Wesley Morgan is the author of The Hardest Place: The American Military Adrift in Afghanistan’s Pech Valley, the Colby Award–winning history of the war’s most unforgiving corner, built on embeds that began when he was a 19-year-old freelancer writing for a college paper. Tony Stark, Justin McIntosh, and Bryan Clark host; Jordan is on dad duty.

We discuss…

• Bubbles of danger — how outposts meant to secure Pech villages drew the war into them instead, and the battalion commander who was trying to sell a pullout at the height of the surge

• The Korengal snowball — two Green Beret teams in 2004, one that refused to be used as muscle in a timber feud and one that went in as hard and as often as it could

• Three tribes — conventional forces, CJSOTF, and the JSOC task force fighting disconnected wars in the same valleys, and the night raid that killed a man the infantry could have arrested at Friday’s shura

• JDAMs and beanie babies — why battalion commanders measured whatever they could write down, and what the Rock’s stat sheet actually showed

• he multicam juggernaut — a Brooklyn art school startup, Delta Force’s R&D budget, and a Ghost Recon producer at a Las Vegas convention

• Lessons unlearned — the Asymmetric Warfare Group disbanded, CALL exiled to Leavenworth, and short-range air defense casualties in the Iran War

Listen now on your favorite podcast app.

Like Weather

Tony Stark: Twenty-five years since 9/11. Twenty-five years of GWOT, twenty-five years of chaos in the Middle East that still hasn’t ended and probably won’t for the foreseeable future. CENTCOM has consumed all things. To mark the anniversary, we thought there was no one better to invite on than Wesley Morgan, author of The Hardest Place, which is basically the story of a microcosm of the war. Wesley, am I missing anything?

Wesley Morgan: No, that’s great. It’s a book about the war in Afghanistan’s Pech Valley, which I use as a microcosm for the broader conflict.

Tony: With us today as usual are Justin, Jordan, and Bryan. Jordan is being an excellent father right now, so the rest of us will run the show while he sits in the background.

Wes, the book is an incredibly detailed account of commanders and individual actions on the ground. Why did you get after this?

Wesley: When I was in my early twenties, I was spending a lot of time in Iraq and Afghanistan as a freelance journalist. This was my fourth trip into the war zones — my second to Afghanistan after two to Iraq. I was in Afghanistan for the summer of 2010, the peak summer of President Obama’s surge, as all the surge troops were arriving in theater.

What I would do is bounce around to different battalions. If it was a three-month trip, which this one was, I’d split it up between four battalions in different places. I wound up going to a battalion in the Pech by happenstance, because an embed somewhere else got canceled and somebody said, last minute, maybe you should go to this one — there’s a good battalion commander there.

So I went out not knowing much about it other than that it was where the Korengal was. The Korengal was very famous at that time. As America returned its attention to Afghanistan after the Iraq surge, the Korengal had become emblematic of the war in popular media. I got to the Pech in July 2010. US forces had left the Korengal about three months earlier and had essentially retreated — they now manned just a string of outposts along the Pech Valley floor. If you imagine the veins of a leaf, the Pech is the spine, and there are these smaller valleys running off of it: the Korengal, and then less famously the Waygal, the Shuryak, the Watapur, which had all claimed American lives in the preceding nine years.

I got to Forward Operating Base Blessing, down the Pech, and from there Combat Outpost Michigan, the one at the mouth of the Korengal. I was really taken aback by everything I saw. It was very different from everything else I’d seen in Iraq and Afghanistan. For one thing, there was vastly more ordnance being expended by both sides than I’d ever seen. This outpost is sucking down recoilless rifle rounds. AGS-17s — the Russian version of a Mk 19 grenade launcher — would be dumping rounds into this place, and then of course mortars and rockets all the time. It was like clockwork. You arrive and the company commander tells you, “You should expect one around 7 p.m. We get two or three of these a day.” Just like weather.

I was also struck by the terrain. It’s a crazy-looking place, much greener than other parts of Afghanistan. Instead of big plateaus, it’s very sharp, jagged mountains that rise right up from the valley floor — very difficult terrain for fighting, and especially difficult for helicopters.

And one thing that was very striking was that although there was this huge amount of ordnance being expended in these engagements out of COP Michigan, it was also very routine. There were American soldiers who were hurt during my stay at Michigan, but it was a heavily fortified place at this point, concrete everywhere, and the soldiers treated these barrages like weather. For the most part it was a relatively safe place to be, despite the torrential downpour of incoming fire hitting it all the time. Likewise, you got the strong sense that the enemy up in the hills wasn’t taking many casualties either, because they had it down perfectly as well. They knew exactly how many minutes it would take before the 81mm mortars started hitting after they initiated a contact, how many more minutes until the howitzers from FOB Blessing started hitting, and how many more until fast air showed up on station and started dropping JDAMs. They would scurry back into the caves before that happened.

So what it felt like was that not too many people were getting hurt or killed on either side. Who was stuck in the middle was the villages outside the US outposts. I knew from counterinsurgency doctrine and the broader strategy of the war that these outposts were supposed to be providing security for these villages from the Taliban. Instead, rather than being security bubbles, they were kind of bubbles of danger. They were drawing conflict into these villages in a way that did not seem to be endearing the American presence to them — which was certainly borne out on patrols. People were over it. They were really, really over it.

And finally, the battalion commander in the Pech at the time was a guy who was very frank about all this. Everywhere else I went in Afghanistan that summer, the battalion commander would be saying, we’re going to build this many more combat outposts, pave this many more kilometers of road, spend this much CERP money, push the ball this much farther down the field by the time the next guy gets here. This battalion commander in the Pech was very frank: “I’ve been here before in this AO. Things have gotten a lot worse by virtue of our presence, and I’m trying to sell a pullout.” That was interesting and bracing to hear from a battalion commander at that point in the Afghan surge, when all the incentives were telling him, no, this is the time for expansion and COIN and outposts and roads.

Stoplight Charts

Justin McIntosh: That’s actually a really good point. One of the things you started seeing in the global war on terror generally — you’d definitely already seen it in Iraq by this point — is that every deployment would start with a stoplight chart: red, amber, green. No matter what the previous unit had done in their area, they’d come in and it would be red. Midway through the deployment it would have switched to amber, because they’d met some magical metrics. And right around the end of their tour, as they’re getting ready to RIP, they’re briefing the commander: “Look, sir, it’s all green in our area.”

But then you talk to actual people on the ground and they’re like, realistically, white space only exists wherever we step. As soon as we take the next step, the area that had been white space is back to being enemy-controlled or contested. That was one of the very hard things to press. Pre-staging this 2010 period, in the same area, you had the Battle of Wanat — the 173rd, right there — where they were attempting to establish these early bases. And that was not the “you shoot at us, we shoot at you, everybody knows the routine” version. That was a lot of people getting injured and killed as the enemy actually tried to fight the US out. It seems like around this time the insurgents realized: we’re not going to beat the Americans, and we don’t have to. We just have to keep being there and challenging them, and eventually they’re going to leave. That was my take at the time on what had happened.

Wesley: The Battle of Wanat is July 2008, a couple of years before this. That was in the Waygal Valley, just north of Blessing, so very much part of this Pech ecosystem, and the book covers all those events. But after visiting these outposts, I was just like, well, what the fuck was going on here? What did I just see? Who built those outposts there, and why, and when?

Initially, many years before the book was published — I was going back to be a senior in college and had to write a senior thesis — I just wanted to understand who built each of these outposts and what the story was on each one. The project evolved from there. One of the big themes that came out of it, which was not obvious to me in 2010, was how much the US military had changed even just between the 2006-to-2008 period, when these outposts were being established, and 2010. Some of them had been established much earlier — the main forward operating bases had started as little outposts in 2002 and 2003, because this was an area where US forces had been at it for a long time relative to other parts of the country, where they’d had a much more episodic presence before 2008 or 2009.

What you see is US forces becoming over time less flexible and less mobile. They become buttoned down into big, heavy outposts because they’re trying to avoid another Wanat, and buttoned up into increasingly heavy, roadbound vehicles because of the IED threat. So by the time you hit 2010, US forces are not fully roadbound, but pretty roadbound, other than the occasional air assault they can muster, which has to be planned months ahead to get the Chinooks. Compare that to 2006, when 10th Mountain went up into these areas and established a lot of these outposts: they did it in pickup trucks and ATVs. The IED threat had not been as great — there were a lot of IEDs, but they weren’t as big — and there weren’t the same theater-wide, top-down force protection requirements. So US forces overextended themselves beyond what subsequent units were able to sustain, because these rules and the appetite for casualties meant subsequent units physically couldn’t routinely get to places 10th Mountain had reached pretty regularly in 2006 and 2007.

The Korengal Snowball

Tony: There are a bunch of rabbit holes I could go down here. One of the earlier parts of your book is the transition between two ODA teams and how that became symbolic of the entire war — how the first team really didn’t want to enter a valley they felt they had no business going into, and was trying to focus on inkblot tactics. Walk through that.

Wesley: The Pech Valley in 2004 was a case study in what people sometimes call the Special Forces or Green Beret identity crisis in the GWOT. You had two teams that each did six months there in 2004. One was an older team with more career experience but on their first GWOT combat deployment, from 19th Special Forces Group out of Utah — so National Guard guys. The second was an active-duty team from 3rd Special Forces Group at Fort Bragg, already on their third combat rotation post-9/11. These teams could hardly have been more different, despite everybody having gone to the same schools.

The 19th Group team was trying to do the inkblot thing. They were focused on building a little security bubble around the district center — Nangalam, FOB Blessing — and training up a militia. The ANA and the National Police hadn’t really settled into the forms they were going to take, so at that point they were really just building a US-funded militia, but they were trying to start doing patrols and security, expanding out from this district center to expand the writ of the government, with the US standing in for the government — which becomes a problem, because it’s not sustainable. But that’s what they were trying to do, and they took it seriously, and they were discerning about it. They were getting human intelligence tips all the time — from the Afghan militiamen they worked with, from their interpreters, from other informants — and a lot of that intelligence was telling them the enemy’s in the Korengal, go to the Korengal, go up there. And the ODA commander, the National Guard Green Beret captain in charge, wasn’t into it. He was like, I don’t understand exactly what’s going on here — it’s going over my head — but there is something at work that is trying to draw me into the Korengal that I had better resist.

The 3rd Group team, when they rotated in, did not have that attitude. They immediately went into the Korengal, and they went in as many times and as hard as they could, because they quickly realized it was a place where they could get into gunfights. It was where the enemy was, and they saw their mission as finding and engaging and killing whoever the enemy was — not just al-Qaeda, but whoever was there. The acronym at the time was AAF, anti-Afghan forces, the euphemistic catch-all for all stripes of insurgent.

What was going over this 3rd Group team’s head — and had been going over the 19th Group team’s head too, but they’d had the maturity to recognize it — was that the Korengal was the center of a timber struggle. The forces allied with the ODA down on the Pech Valley floor represented one commercial interest, and the Korengalis represented a different one. So the Green Berets were being used as muscle on one side of a timber dispute. The 3rd Group team fell into this trap and started the Korengal snowball rolling. From there it evolved into this kind of Colosseum where US forces were fighting local insurgents, but the local insurgents had also called in big Taliban, who’d had a very limited presence in Kunar before 9/11. It became an arena for very intense, visually arresting, dangerous combat in the years that followed.

Justin: On the identity crisis within Special Forces — I would have just been joining around the time that 3rd Group team was getting there, and there was definitely a legacy of the guys who’d been involved pre-9/11. My first team sergeant was a pre-9/11, invasion-of-Afghanistan guy, and he’d talk all the time about, hey man, no, you do it through other people. That’s the whole point of the job — to get other people to do the fighting. If we’re having to get into gunfights, we’ve messed up.

Very quickly, within a couple of iterations, as those older guys got promoted up the chain or retired, you started seeing that change to where it became very aggressive. It was, we’re going to do F3EAD — find, fix, finish, exploit, analyze, disseminate — continually doing hits and night missions. That became the poster child. Even when Special Forces at the time would say, no, no, we want to be the by-with-and-through people, empowering others, unconventional warfare — every recruitment poster we had was somebody kicking in a door or somebody doing a HALO jump. Which is great. I was on a HALO team; I loved doing it. But there was a pretty fundamental break between what was actually going on on the ground and what we were being asked to do, versus what we had doctrinally told ourselves our job was.

That was reinforced by commanders like Odierno in Iraq, and honestly even McChrystal and Petraeus to some degree in Afghanistan, because they could trust Special Forces to go out and do missions they didn’t necessarily want conventional forces to do. They continued to use them as this weapon to go smash the bad guys, for lack of a better term. And it just reinforced those bad lessons: you just need to go find where the bad guys are. It doesn’t matter. Just go get the bad guys. And like you said, a lot of what that initially was was intertribal disputes, and we were basically picking the side of one of the aggressors and giving them muscle to push over their enemies.

Wesley: Exactly — in a way that then invited the insurgency to grow, and created incentives for locals to make connections with the broader Taliban.

Tony: Years later, by the time I joined, it was a known fact in the infantry that if you want to be a door kicker, go SF, go Regiment. That was the default assumption: if you actually wanted to go do stuff, that’s where you went. Which, if you look at the history books, is very different from the Cold War perception of Regiment, of Delta, of SF. That’s a different mission set, a very specific one, which is hopefully what it returns to at some point. But by the 2010s it very much was: if you want to do the cool infantry shit, go do SF.

Three Tribes

Wesley: We should draw a distinction here. In the book I talk about different tribes of the US military in Afghanistan. You can break that down in various ways, but the way it manifested in the east was: you had conventional forces; you had what was called the CJSOTF [Combined Joint Special Operations Task Force], which was the Green Berets, responsible for all things special operations that had a partner — running the ANA Commandos and doing the Afghan Local Police, two very different things, and it was always interesting to hear how a particular company or battalion prioritized those and which teams they assigned to them, because that tells you a lot about what the organization valued. And then the third tribe was the very separate JSOC task force, which went by numbered code names — Task Force 373, Task Force 310 — and consisted largely of SEAL Team 6 and the Rangers, with some Delta thrown in. Their job was the pure direct-action night-raiding mission: living at some big airfield, flying out at night to do hits.

The Green Berets, in a certain way, had more in common with the conventional forces, in that at least the ALP teams were out living somewhere with people, at some little outpost. The commando mission probably had a little more in common with the JSOC mission — living at a big airfield, going out to do not targeted hits but maybe valley clearances, and coming back after three days. So those were the three tribes, and they were all doing disconnected things that sometimes supported one another and at other times really stepped on each other’s toes.

That’s one of the overarching stories of the US war in Afghanistan: different parts of the government pursuing missions that ultimately aren’t compatible. Outside these tribes, you throw the CIA and its counterterrorism mission in there, and that adds another layer. They make things harder for each other as they go. The classic example with the task force — they always called the JSOC task force “the task force” — is they come in and do a raid and kill a bunch of people getting a guy, and the infantry company commander in the area realizes in the morning: we could have just arrested him at the shura on Friday. He would have come quietly. We know him.

The flip side is that sometimes the task force could go up into a valley the conventional forces couldn’t reach anymore because of the way resources were centrally allocated. Conventional forces couldn’t get up into the Watapur Valley anymore. They couldn’t get up to Tsangar the way they had in 2006. Only the task force could do that. So over the years, conventional forces were kind of neutered, to the point where special operations forces were playing offense for the broader ISAF counterinsurgency mission, as opposed to doing narrower counterterrorism. I think that was deleterious for the conventional forces too — to have the ability to hit back taken away from them and given to somebody else who often didn’t really understand the battlespace.

JDAMs and Beanie Babies

Tony: Can we go back to the metrics for a second? Especially on the conventional side, you have commanders who are struggling to find the enemy, struggling to get after the mission they deployed for, but also trying to justify their existence and their future and their career, because those things are interlinked. If you talk about miles of road built or combat outposts built or meetings held — looking back, absolutely none of that is a metric you’d use for success. Roads are a very important engineer mobility mission, but that’s a component of a broader mission, not success in and of itself. Same with meetings. They’re all enabling factors. When you talked to commanders on the ground, did they truly believe in those as the metric for success, or was it just the only thing they could write down?

Wesley: I think often it was the only thing they could write down, or the only thing they could wrap their heads around. Certainly there were commanders who understood the counterinsurgency stuff in their bones — it was second nature to them, they understood dealing with people, they just got it. And then there were a lot more commanders, I’d hazard, who didn’t, and for whom this was the flavor of the day. This is what they’re being told, these are their orders. It doesn’t matter whether it feels good or whether it’s what you joined the Army for; it’s what you’re being ordered to do, so you do your best. But they didn’t relate to Afghans. They did not enjoy sitting around the shura table. They did not enjoy working on development projects. So it became: well, what can you measure?

There’s an example I give in the book. There’s a battalion that fought up in Kunar — the one that was involved in the Battle of Wanat, that lost two dozen soldiers over a fifteen-month deployment, including nine that day. This is 2nd Battalion, 503rd Infantry, out of Italy — the Rock. The Rock was commanded by a very charismatic leader, a great guy I’m still in touch with, then–Lieutenant Colonel Bill Ostlund, who looks like he came out of We Were Soldiers, like a Mel Gibson character. A picture-perfect battalion commander and a great Army officer, and one who really threw himself into counterinsurgency, but not one to whom it came naturally. By his own description, the only Afghans he ever felt any connection to were a handful of Afghan Army officers, because there was something in common he could latch onto.

I used to get into it with him when I was doing the book: what did your deployment accomplish? And he would say, well, we delivered this many JDAMs and this much artillery, we had this many contacts, we spent eighty percent of the brigade’s CERP budget, and we handed out ten thousand beanie babies. These were the things he was measuring success by. And none of them showed anything — in fact, they gave a very contradictory message. If you’re both handing out ten thousand beanie babies and you’ve called in more artillery rounds than any battalion in the country, that’s a complicated picture you’re painting. It is not a portrait of success.

Why They Joined

Tony: I want to pivot to 25 years of war — really 20, if you count from entering Afghanistan to leaving. I’m also reminded, looking at my social media feeds this week, that it’s 25 years of Band of Brothers. Through my Army career, and I think everyone’s, Band of Brothers was the thing. Lieutenant Winters’s assault on the German artillery on D-Day is the textbook thing; it’s what everybody shows. So many people joined because of that, or because of Black Hawk Down. But Justin, you had a good point in your piece this week: why you joined wasn’t necessarily 9/11, because it wasn’t immediately after. The people who joined in ‘01 are not the same people who joined in ‘03 or ‘07 or ‘14 or ‘18. Wes, as you went through your research, what did you see as the common factors for why people ended up over there? And what were the weird standouts?

Wesley: It’s funny you mention Band of Brothers. I was interviewing a former Delta Force operator the other day, and he was talking about how at the time of the invasion of Iraq, on their first rotations to Baghdad as they were standing up the so-called Baghdad SWAT mission — where they’d go out in the Little Birds and do what everybody in the military saw as the coolest thing you could be doing — they were watching Band of Brothers every night. That’s what they would come back to.

In my experience, the popular media I remember being dominant was Black Hawk Down. Black Hawk Down was everywhere. The movie comes out in December 2001, and the soundtrack, lines from that movie, whole scenes that soldiers could recite —

Justin: “This is my safety.”

Wesley: Exactly, the finger on the trigger. Black Hawk Down was huge.

I will say, my experience of the wars in Iraq and Afghanistan was infantry units and cavalry units, with a little bit of Green Berets thrown in. I was not spending much time with logistics units or field artillery units, so it’s not a representative sample. But in these infantry units, a lot of it was just young men who wanted to go to war and fight. That’s what they were there for. There were places where they were disappointed — the vision of war they got was very frustrating and not at all what they had wanted from Band of Brothers or Black Hawk Down. And then there were places like Kunar and the Pech, where it really delivered, and you felt like you were doing what the Army does and what the infantry is supposed to do.

Justin: I have probably watched at least an episode of Band of Brothers, if not the entire series, on every deployment I’ve ever been on. That and Tropic Thunder were my go-to deployment things — also an amazing war movie that everybody should watch. It still cracks me up.

But I was a kid at Fort Benning, Georgia. That’s where we lived in 2001, when September 11th happened and when Black Hawk Down came out. So the night it released, I watched Black Hawk Down in a theater at Fort Benning, surrounded by high-and-tights and Ranger kids and kids who wanted to be Ranger kids. It’s something you don’t realize when you’re growing up in it — that that is the culture. But there was definitely a break, because that wasn’t the reason I joined the Army. The reason I joined the Army was that it was there, and I was looking for something to do that wasn’t college anymore. By the time I joined, just a few years after September 11th, it wasn’t a dominating thing. I’d be interested in Bryan’s perspective. When you came in, what was the motivating factor, and what was the Navy’s thing?

Tony: It’s Hunt for Red October. We know the answer. It’s got to be Hunt for Red October.

The Navy’s War

Bryan Clark: I joined the Navy in the eighties, during the Reagan buildup. The economy in the early eighties was terrible — unemployment was relatively high, interest rates were high. We didn’t realize how bad it was, but it wasn’t great. So a lot of people joined the military because it was a good option. You could see Reagan was putting a ton of money into the military and pay was likely to go up, even though the experience of most people we knew who’d been in the military was very negative, because it had been Vietnam, and the post-Vietnam military was kind of a mess.

I enlisted in ‘82. When I went off to boot camp, you could still have beards. There was still a drug problem — they had just instituted drug testing, and guys were popping positive all the time. Back then you could get a mulligan, basically. Guys would pop positive and get a waiver, a dispensation to continue, because it was so prevalent you had to start somewhere. It was a very different era. The reason people joined was mostly economic, because at that point the military did not have a great reputation, though Reagan was doing a lot to build it up.

By the time 9/11 happens, I was an O-5. I had just taken over as operations officer at the Navy’s Nuclear Power Training Unit in Charleston. I’d been there about three weeks and thought, this is a great job — I get to supervise all these facilities and the nuclear plant operations. Theoretically security was also something I was in charge of, but I didn’t care that much about it. They had it. It’s fine. Then 9/11 happens, and suddenly we’ve got to lock that place down, because everybody’s worried about a terrorist attack on this nuclear facility sitting out on the Cooper River. There was basically no protection. People would drive by on the river all the time, right next to these nuclear-powered submarines sitting in the water with no defensive measures at all. We had to mount a pretty big campaign to improve that. A lot of people of my generation had been used to the Cold War military and then the post–Cold War military, which had not been very challenging from an operational standpoint, and suddenly we were thrust into this very different world.

Wesley: On the Navy’s role — one fascinating representation of how the Navy and the Air Force got in on the ground war comes up in The Hardest Place, because the provincial reconstruction teams for both Kunar and Nuristan were led by a succession of submarine officers, nuclear engineers, naval aviators. Sam Paparo, the INDOPACOM commander, led the Nuristan PRT for a year that coincided with the Battle of Wanat. The Navy and the Air Force were sending their very best — the best command-selected O-5s — to do these jobs. If you look at the bios of a lot of admirals and Air Force generals floating around, there’s this little chapter that’s very different from the rest of their career, where they were in Khost or Jalalabad, out every day driving around in MRAPs doing counterinsurgency.

The Class of 2008

Tony: Wes, you have an encyclopedic knowledge of basically everyone currently running the United States military, because of all your time with commanders on the ground. What have you seen in the evolution of the commander, broadly speaking, from the early days of GWOT to now? And what’s been the strangest thing to witness as some of these guys have grown up and taken three or four stars?

Wesley: My first trip to Iraq was 2007. In 2007-8 Iraq, the two-star generals are guys who had never seen any type of combat when they were growing up. Maybe they’d been in Desert Storm; if they were Rangers, perhaps Panama. But they grew up in a peacetime Army. The battalion commanders at that time had come in pre-9/11, and the company commanders in 2007-8 had come in right before or right around 9/11. The guys who were the platoon leaders and company commanders and battalion commanders in the Iraq and Afghan surges — they are today’s colonels and generals.

The battalion commander in the Pech when I first visited was a guy named Joe Ryan, who is now a four-star general commanding this strange new creature, the Army’s Western Hemisphere Command, built out of FORSCOM and some other things. And anytime you see news about the promotion lists, as Secretary Hegseth pokes at them and pulls people off, every name is one I recognize from when they were a company commander. Basically, colonels right now were company commanders in Afghanistan in 2008, 2009, 2010. There’s one who appears in the book who was just pulled off the one-star list — he was seen as the best company commander anybody had ever heard of in eastern Afghanistan in 2008-9.

So that’s how I know the Army: this blip in time of junior and mid-level officers who fought in Iraq and Afghanistan during a certain period. Today’s company and battalion commanders don’t know me, because the relationship between the military and the media has changed drastically since the time I was starting out, which was a really rare and privileged time of incredible transparency. Basically you couldn’t visit the task force, and that was about it. I showed up in Baghdad in 2007 as a 19-year-old freelance reporter, literally writing for a college paper, and they asked me, okay, which battalions do you want to go to? I said 1-28 Infantry in West Rashid, 1-14 Cavalry on Haifa Street, and off I went. Public affairs was there to facilitate your getting to the units you wanted to see at the platoon and company level in combat. So there’s a generation of Army and Marine officers who had a ton of exposure to reporters and became very discerning about how to deal with the media. Today’s battalion commanders and below — especially below — don’t have that. The media is just a scary thing that’s been held at arm’s length since the age of embedding came to an end around 2013, 2014. The relationship today is atrocious.

How Embedding Died

Justin: From your perspective, how much did the McChrystal incident affect that relationship? To me, that’s when it seemed to start getting different.

Wesley: Maybe that’s the way it was from the other side of the fence. From my side, it didn’t affect it. I remember being in Sangin when that happened, doing an embed with 40 Commando Royal Marines, a battalion that was out there taking casualties constantly in this horrible place. I came back from a little outpost where I’d spent a week without any internet, and the British battalion commander made a joke about McChrystal. I laughed, but I didn’t know what he was talking about, and then he explained. I didn’t notice a change until after Obama’s reelection, as the Afghan drawdown was happening.

2013 was the first trip I made where embedding was still very much a thing, but there started to be more limitations on it. It started to be, well, we’re not actually going to embed you with that unit, because the ANA is in the lead there — if you want to see that, go embed with the ANA. And I would do that. But then in Ghazni it would be: I said I wanted to visit 1-28 Infantry again, and they said no, you’re going to visit 2-16 Infantry, because we’ve got something going on there that we want to show you. It became more stage-managed and more limited.

I think it was a desire from on high to put the war out of sight and out of mind — particularly the fact that even as American forces were transitioning to a training-and-advising role, that advising was still combat advising. That was a complicated message, and the Obama administration did not like it being complicated. That’s really when it ended. Then, going back into Iraq in 2014, it just never began again, and subsequent administrations had no incentive to reinstate it or be more transparent.

The history of embedding is a fascinating subject that I’d love to do a Columbia Journalism Review deep dive on someday. It was the first Bush administration — Donald Rumsfeld, actually — who were the real proponents, who said we’re going to lean in on transparency. There’s a bunch of emails from L. Paul Bremer’s year as CPA administrator in Iraq, all archived in the UK, and you can read them. There’s a point in the late summer of 2003, as things are starting to get uglier in Iraq, where Colin Powell says internally, in the NSC or wherever, why are we letting these reporters come out on patrols and see us trashing people’s homes and taking casualties? And Secretary Rumsfeld is the one who puts his foot down and says no, we’re going to keep doing that. A fascinating role reversal from how you’d expect that conversation to go. But that’s the story of embedding. Sorry, I’ve wandered very far from your original question.

Justin: No, that was actually really good. By the emergence of ISIS in Iraq, what we saw too — I’ve talked about this before, at least with Tony and Bryan — is that from the US forces’ side, the message that had to be controlled was that risk to force really became risk to mission. The fastest way to end the de-ISIS mission was dead Americans. Part of that message was, Americans aren’t in combat. So obviously we weren’t going to have anybody near that, because in an insurgency there’s always a chance things end up behind you. The lines were drawn, but they weren’t super stark. That explains it to me. What’s more interesting is that you have a person like Rumsfeld — we have the term Rumsfeldian for a reason — being so open to, we’re going to get in front of this and bring them in. Do you think that was born out of his experience post-Vietnam? Why was the department open to it?

Wesley: I don’t know. I know more broadly, at President Bush’s level, he was invested in embedding and personally involved in starting it, as was the Pentagon public affairs director who created the program. I think the idea sat somewhere between, on one end, an idealistic post-9/11 idea that the American people deserve to see, and two, the idea that if they do see, they’ll understand it better — they’ll see it from the perspective of our soldiers and sailors and airmen and marines. Which I think is true. Good news gathering definitely meant doing more than embedding, for that reason. By going out with Army infantry battalions, you saw the war the way Army infantry battalions saw it. You had the opportunity to see it that way, and you could grow beyond that if you tried. But that was the pitfall of embedding: if you didn’t round it out with other perspectives and spend time with Afghan and Iraqi civilians and armies, you could get a very blinkered perception of the war — one that in some cases really served the administration’s interests, and in other cases really did not, because often the soldiers fighting the war were not that into it as it dragged along.

Justin: If somebody had been embedded with me on my first deployment, as we got stretched to fifteen months, they would have found out that I really disliked the war in Iraq.

War Reporting Without Access

Tony: From a war reporting standpoint: there’s a false sense that sensors and video on the battlefield are pervasive and you’re getting the full picture. How is that shaping editors, and those who would fund war reporting, in their understanding of what you as a war reporter can go do? And what does the modern space of embedding and war reporting look like?

Wesley: I have largely shifted to retrospective work about the wars in Iraq and Afghanistan rather than covering current wars, for a variety of reasons. One of them is that it can’t be done the way I did it in my twenties anymore. One, states are not interested in that kind of transparency and embedding reporters with their forces. They can see all the downsides, and they don’t see the upsides. In Ukraine, the Ukrainian government is very restrictive about Western media visits to the front line. They’re kept brief, they’re rare — and to be clear, I haven’t reported in Ukraine; this is what I understand from colleagues. In Gaza, the IDF only allowed what they called embeds, which were essentially one period of daylight, gaggle tours of reporters brought in with a general. In the lingo of twenty years ago, we wouldn’t have called that an embed at all. We’d have called it accompanying a general on a battlefield circulation. The US military very occasionally will grant a reporter a few days on a ship or at a base in theater, but that’s all there is. To me, that’s the biggest restriction on that particular type of reporting.

The flip side is that reporting from the battlefield unembedded has become much, much more dangerous for journalists. It was dangerous in 2004 too — there were reporters getting their heads chopped off during al-Qaeda in Iraq’s initial big rebellion. But 2014 is really the year, when James Foley is beheaded in Syria, that a lot of newsrooms start cracking down on risk and saying, okay, we’re not going to sponsor freelancers on the battlefield anymore, because that’s crazy — when they die, it’s on us. So with the military having less appetite for casualties and newsrooms not wanting casualties, major newspapers covering Ukraine imposed restrictions on their reporters that are totally alien to how those same newspapers operated in 2008 Baghdad.

That’s certainly frustrating for people who experienced that version of war reporting. But it doesn’t mean you can’t do it; it means you have to find other ways. There are great reporters, of my generation and younger, who have found ways to get around this and still find the truth. Visual forensics, visual investigation of videos, is one of them. There’s a flood of video that comes out of every modern battlefield. The reporter is no longer where the video comes from. The video comes from everybody, everywhere, all the time, and the trick is to find out what of it is real. The Washington Post, the New York Times, and the BBC have great visual investigations teams that have mastered how to exploit what’s coming off the battlefield from civilians to verify and illuminate what’s happening.

Justin: A personal anecdote. My brother was in the 82nd for a while, and he was at one of the bases in Afghanistan that Anderson Cooper was live-streaming from when the base got attacked. Instantly my mom is calling me: what’s going to happen to your brother? He was fine, nothing happened, but that was three days of dealing with that, because she’s just watching the CNN clip of Anderson Cooper going, “And now we’re getting shot at,” and people running to their positions. So on a personal level, I’m actually glad they do a little less embedding and live-streaming from the front lines. You lose a lot of control of the message when you give reporters free rein to show whatever they want.

Wesley: You do lose a lot of control, and that’s why it’s gone.

The Multicam Juggernaut

Tony: On one of your other favorite topics, Wes: how did multicam become everything over the course of the war?

Wesley: This is one of my favorite topics. It’s waned a little bit now — I feel like multicam is passé. But three or four years ago, this was something I was really fascinated by, because you were seeing a pattern that was suddenly on all sides of every conflict. Both sides in Ukraine wear multicam — the Azov Brigade literally has a metal band with a song called “My Skin Is Multicam.” The Russian military has adopted a multicam knockoff. The Taliban wears multicam. It’s just everywhere.

But it didn’t used to be. In 2007 and 2008 in Iraq, I didn’t see multicam. I remember first seeing it in 2009, in Wardak province in Afghanistan, on a Green Beret. It was exotic and cool. Even with no background, never having played any video game that multicam featured in, you could instantly tell: that’s exotic and cool. Cut to about a year and a half later, and the Army adopts it for all soldiers in Afghanistan, and eventually for all soldiers worldwide.

There’s a fun backstory here that I wrote an article for New Lines magazine about a couple of years ago — how multicam became the juggernaut it is today. It’s a crazy story. Multicam was invented by a couple of Brooklyn art school graduates around 2002, Caleb Crye and a couple of his friends. They initially tried to sell it to the Army, which in 2003-4 was looking for the successor to the old BDU and DCU, the two camouflage patterns you’d switch between depending on the environment. The Army held a big competition, and ultimately settled on the worst of all worlds, the pattern that went on the ACUs — the gray digicam, as I think of it. That actually kind of worked in Kunar. I have some cool pictures, including the one on the cover of the book, where soldiers are blending right into the mountain rock in their ACUs. But mostly, no — that was not the broad reputation of ACUs.

So multicam was passed over. It tested very well, but ultimately the Army went with its own option. The backstory that was not told at the time is that in Iraq, Delta Force was having none of it. They saw this ACU pattern, saw especially how it performed at night and how light hit it, and said absolutely not, we’re not putting that on our bodies unless it’s for a limited purpose like concealing ourselves within a conventional patrol. So they kept wearing the old DCUs, and with their big R&D budget and their authorities to test different things, they started testing commercial off-the-shelf camouflage patterns. One of the ones they tested, starting in late 2005, was multicam.

From there it spread. The British SAS adopts it because they’re working hand in glove with Delta. Green Berets start adopting it because they see Delta and the SAS wearing it. Then other foreign special operations forces start adopting it, and there comes a time in Afghanistan when basically all special operations forces — Australian, Estonian, Polish, whoever — are wearing multicam. It’s the marker of eliteness. That sheen starts to fade when the conventional military picks it up and deploys it at scale.

But the fun backstory of how it acquired that sheen: Crye, the company that made multicam, faced a dilemma after selling this stuff to Delta Force. That’s a huge win, but it’s one you can’t advertise at all. And it’s also one that’s not reflected in the photos coming back from the battlefield from embedded photographers. Today there are a million Instagram accounts that are just photos of JSOC operators — that’s a genre now. But at the time there were no photos of JSOC operators in combat, because embedded reporters didn’t have access to the task force. That was the one thing cut off from them.

So what Crye did that was very clever — I interviewed a video game producer about this; I’m not a video game guy, so I’m forgetting the name of the company, but it was the one that did Tom Clancy’s Ghost Recon [Red Storm Entertainment] — this producer recalls being at a video game convention in Las Vegas in 2005, I think, where the Army had a booth. Natick Labs had a booth, and there was a Crye person at the Natick Labs booth. This video game guy recalls talking to the Crye guy and the Natick guy and being told, hey, you should feature multicam in your games. And he winds up doing that. It’s intimated to these video game producers that this is the camouflage of the elite, so it starts being represented as the camouflage of the elite in Ghost Recon 2 or whatever, and then in other video games — before there are photographs of it actually being the camouflage of the elite. It got out into the video game universe before it got out into the traditional media universe. So by 2009-10, young soldiers deploying to Afghanistan have played these video games and are familiar with multicam before they see it on a visiting special operator.

That’s how multicam acquired this sheen of eliteness, which I think is now on the downswing. Now what’s cool among collectors, among people fascinated by this stuff, is retro. They’re like, no, DCUs are cool, BDUs are cool. But that’s the story of multicam.

Tony: It’s really interesting on the video game front, because I’m — and I think Justin is too — of the generation where the blockbuster video games influenced whether we joined or not. Call of Duty and Medal of Honor were the two titles that really leaned into that.

We talked about Black Hawk Down and Band of Brothers, but the video games influenced a lot of people too, and in less realistic ways. I remember my first driver when I was enlisted. He was like, I joined, and I was so confused when I got to basic and they didn’t put me in a juggernaut suit — the big EOD suits. It’s interesting to see how other people took that as the reality of what they’d get to do when they joined, instead of understanding, as we know it, that helicopter raids and so on are for a very particular set of skills. And the juggernaut suit does not count.

Wesley: What you do not do in those video games is collect census data from the neighborhood. You do not sit in the shura and learn all the elders’ names. That’s not a feature of the video game universe of war.

Justin: There was an Onion piece, circa 2006 or 2007: a video game manufacturer makes a realistic war game where soldiers sit around for nineteen hours a day drinking energy drinks and smoking cigarettes.

Wesley: The video game producer I interviewed — I remember him saying you could have placed multicam in a movie to start with, and it eventually is. The first movie appearance is Transformers.

Tony: I was going to guess that. I feel bad that I was going to guess that.

Wesley: So it does get there, but it was in video games first. And he said, with a movie, it’s a couple of hours, and people watch it three or four times if they really like it. You put it in a video game, that’s hundreds and hundreds of hours, and they’re interactive hours where people are selecting their uniforms. So they form an attachment and a bond to pieces of kit in a way you wouldn’t get by placing it in Transformers.

Jeans and a Fishing Vest

Justin: Just thinking of all the gear nerds out there who really do go through and collect all this stuff, both in the military and out. Man.

Wesley: It’s so crazy. I’ve been interviewing a lot of former SEAL Team 6 and Delta guys recently — specifically the ones who are in the public eye and have put themselves there, with memoirs and podcasts and so on. One guy who’s in the public eye but not totally comfortable with it says, yeah, I’ll get these DMs from kids on Instagram: here’s this photo from 2004 of you in the desert — what’s this pouch? And he’s like, I have no idea. But they know. They know.

Justin: I couldn’t have told you. There were people who would spend hundreds or thousands of dollars on their kit. I put my kit together one time and used it until they made me turn it in and handed me a new one. I don’t care about this shit.

Tony: I don’t know what subreddit it is, but I remember there being a very popular picture of a Ground Branch guy’s kit. And it’s like, what do you think they’re doing? There’s a flashbang pouch. He’s got a couple weapons and some mags. It’s not magical, guys. But it’s that level of obsession.

Wesley: And of course the famous Ground Branch pic that I think is now making the rounds, as people come down from the multicam high, is a photo from Toby Harnden’s book First Casualty, about the CIA in northern Afghanistan in the fall of 2001. It’s Team Alpha, I think — one of the Ground Branch teams that went into northern Afghanistan — all posing before they go in. They’re all wearing jeans and a fishing vest and carrying an AK-47 with no special sights on it. That’s who went in and did the hardest, craziest stuff at the beginning.

Tony: That’s what’s wild about SOF and popular culture — Army popular culture going through GWOT — that ever-increasing obsession with the next tier up: this special unit upon this special unit upon this special unit.

Wesley: When in reality those special units were in a lot of ways counterproductive to the war. I’m working on a book now that gets into how stuff that worked in Iraq didn’t work in Afghanistan. You can trick yourself into thinking, well, because they’re the most special, they were doing the most important work in the war, or if only they’d had their hands untied, they could have won it. But that’s not how I remember it.

What Did We Learn, Army?

Tony: To close out, I’m curious about the stuff you’re working on now, which you just touched on. But two: there’s a lot of commentary that we’re going to forget all the lessons we learned in Afghanistan. Did we learn them during the war? That seems to be the analysis of your book — most of us didn’t learn these lessons anyway. So how much of that retention is worth it?

Wesley: I think there were a lot of things learned, at least at the low levels — platoon leaders, company commanders, battalion commanders. Perhaps not as much by some of the more senior guys, who just kept trying to do the same big air assault into the Watapur Valley every year, no matter what.

I do think it’s a real problem, the extent to which the Army is retaining this knowledge. I think it was Andrew Exum who said, pretty soon after the invasion of Ukraine, that Ukraine is going to be the best excuse the US Army ever found to never say the words Iraq and Afghanistan again. That has proved true. An early harbinger was when, in 2021 or maybe as early as 2020, the Army disbanded the Asymmetric Warfare Group, which was one of the heroes of The Hardest Place. The AWG guys — in particular a guy I call TW, an old, legendary Delta operator who was one of the founding AWG members — became the continuity in this war of discontinuous rotations. They could only do that at the very tactical level. There was no AWG for the things you need to know about your provincial politics; there was just AWG for how the insurgents fight in the Korengal, or how enemy sniper tactics work in Mosul. But a lot of that went away with AWG, and it’s not a good sign for AWG to be gone.

Justin: What the Army has done really poorly is that it created the Center for Army Lessons Learned — CALL — and that gave everyone an excuse not to do their own research. I can just wait till it gets put out, and whatever they put out, that’s it. We’ve done lessons learned; it’s this thing that sits over here. I don’t need a unit historian. I don’t need somebody pulling through reams of data and bringing up best practices and challenging our training when we go out — saying, hey, that wouldn’t work, and here are the reasons, and I can give you all the data. We divorced all of that thinking from the operational side and put it in a center at Fort Leavenworth. It might as well be in the prison, because that’s how many people actually go access it. I’m being slightly pejorative, but the truth is plenty of people don’t. CALL has to go on speaking tours around Army installations basically asking people, can you go to our website? Can you try to look at our stuff? It hasn’t become an ingrained part of who we are. We’ve just given it to another functional.

Tony: Justin, you and I were talking this morning about that article in Task & Purpose. Wes, if you haven’t read it, take a look. It’s the company commander from the cav unit that got routed by the Ukrainian unit at Grafenwöhr a few months ago. He says, well, we learned throughout. But the things he talks about are: well, we didn’t do spacing correctly when we first went in. That’s a perpetual lesson in the Army. That’s your Ranger Handbook. If we can’t even ingrain the things we know to be true in any type of war, I’d be surprised if we remember the things from our most recent one, which lasted twenty years and for some units is still going on.

Wesley: Sure. And there’s a reason the Army wants to focus on Ukraine and never think about Afghanistan and Iraq again: a lot of the Army’s core pre-9/11 skills degraded during those wars, because of the intense focus on every brigade doing counterinsurgency. It certainly went too far in the other direction as far as readiness. I remember, nine years ago now, when I was covering the Pentagon, I did a story about the 173rd Airborne Brigade. In 2017 they put out a big, alarm-bell, hundred-page lessons report that was basically, these are all the things the brigade is incapable of doing. They included short-range air defense, which was something that just didn’t exist anymore, because the Army had gutted it. Any NCOs still around who’d been in short-range air defense battalions had done a couple of deployments to Baghdad as provisional infantrymen. So that’s something the Army had to rebuild from the ground up, because there were multiple generations of knowledge gap. There weren’t sergeants who could train up the new privates. They just had to rebuild.

Now there’s a lot of new short-range air defense force structure. And I think three or four of the fatalities in the Iran War, the ones in July, were men and women from short-range air defense units. So you understand why the Army wants to move past Iraq and Afghanistan, but there’s a balance that has to be met.

Tony: Justin, this is not the first time we’ve closed an episode with “what did we learn, Army?” I guess we learned not to do it again.

Justin: That’s the thing. We just won’t do counterinsurgency anymore. We’ll stay out of those quagmire-y situations where we try to do regime change.

Tony: Surely there are no countries in the Indo-Pacific with difficult geography and six thousand languages, none of which US personnel speak. Surely there are none of them, and none of them are in proximity to the PRC.

Reading Recommendations

Justin: Wes, give us some reading recommendations before we close out.

Wesley: The Airborne Mafia, by Robert Williams. He’s a former squad leader in the Korengal with the 173rd, and now a school-trained military historian with a PhD. He wrote a book a couple of years ago about how the generation of airborne officers from World War II influenced the culture of the broader Army in the decades that followed. A fascinating book, and one where you can see parallels to the Ranger Regiment in the modern Army — it feels like every general came from the Ranger Regiment.

Another book I was rereading recently that’s great on similar issues — military and war, military and peace, how you meet the balance between the two — is The Caine Mutiny. It’s about young Navy officers in the Pacific in World War II and their collision with the older generation of Navy officers who kept the lights on between the wars, who they see as stupid and lazy, but who in fact were the ones who kept the institution alive and carried the torch. Those are the first two that come to mind. And I’ll plug my favorite novel of all time, which is not closely related but is a great book: The Long Ships, by Frans G. Bengtsson, if you want some hilarious and vivid historical fiction about Vikings.

Tony: Hell yeah. Wes, thank you so much for coming on. I think this was an excellent episode, and perhaps the only correct way —

Wesley: It was a pleasure. Thank you.

• Just how much Wang is obsessed with cutting costs;

• What it was like bringing humanoids onto the Spring Festival Gala stage;

• Unitree’s managerial structure;

• And how the IPO has — and hasn’t — changed Unitree employees’ lives.

Footnotes, bold markings, and block-quote comments were added by ChinaTalk editors. The translation was drafted with Claude Opus 5 and revised by the ChinaTalk team.

But first, we’re running a personal classified from a real life friend of Jordan!

I’m 30, a straight guy based in DC who works in tech and national security policy. Most of my “free” time is spent outside. That’s a lot of 6am long runs, and hauling a camera up mountains that don’t require it. Date me if you’re looking for someone who’ll plan the trip, carry a heavy pack at a price Unitree can’t beat, cook when we get back, probably lose to you at every sport that isn’t running slowly for a very long time, and always give you that Strava kudos.

Respond to this email if you’d like to learn more about a person I strongly vouch for as a good human being and hang!

By Tang Min’an 唐敏安 and Huang Siyun 黄思韵, researchers at Caijing

Employees say that at Unitree Robotics (hereafter “Unitree”), any expense reimbursement over 100 yuan requires Wang Xingxing’s approval.

Wang Xingxing is the founder of Unitree Robotics Co., Ltd. Since establishing the company in August 2016, he has served as its chairman, general manager, and chief technology officer. On August 19, 2026, Unitree listed on the STAR Market, becoming the first “humanoid robot” stock on China’s A-shares market.2 After the IPO, its market capitalization briefly reached 440 billion yuan, and Wang Xingxing briefly became the richest person in China born in the 1990s.

In the eyes of Unitree employees, Wang Xingxing’s personal traits are pronounced: his communication style is relatively blunt and severe; he requires employees not to disclose anything about Unitree externally; he is product-focused, frequently discussing product details with front-line staff, down to the length of a single screw; he cares intensely about cost control and insists on spending every cent where it counts; and he rarely talks about anything outside of work.

Working at Unitree comes with high pressure. Employees often queue outside Wang Xingxing’s office waiting for approvals. Nowadays, he has less and less time to go around, with many people ringing him up to chase him. Employees reporting to him have to talk fast — two sentences in, they will likely be cut off or told to hurry up.

For Unitree’s R&D staff, being criticized is routine. Coworkers compare who gets criticized less among themselves. Unitree’s performance scores range from 0 to 1.5; Wang Xingxing gives every senior executive a score of 1 or below.

Within China’s embodied AI industry, Unitree is currently the most closely watched star company. It solved the problem of humanoid robots’ high price tags and has stayed profitable for several consecutive years — one of very few companies in the industry to sustain profitability. The label on Unitree’s robots is “low cost, high performance.”

For capital markets, Unitree set the trend for “humanoid robot concept stocks.” As of August 28, its market capitalization stood at roughly 248.8 billion yuan — seven times the contemporaneous market cap of UBTech, a robotics company which previously listed in Hong Kong, and ten times Unitree’s own valuation in its final pre-IPO funding round.

Doubts about Unitree mostly center on R&D spending. The humanoid robot is called the jewel in the crown of robotics: the humanoid form is the most general-purpose and has the widest range of applications, but it is also the hardest to build. According to its prospectus, Unitree spent 145 million yuan on R&D in 2025, or 8.53% of revenue — a middling level among tech companies, given that STAR Market-listed firms average roughly 11%. UBTech spent about 500 million yuan on R&D in 2025, about 25.4% of revenue.

Unitree’s R&D-to-revenue ratio was 31.39% in 2023 and 17.83% in 2024. The decline is mainly because revenue grew explosively over the same period, diluting R&D as a share of the total.

…

Unitree’s offices in the Zhongguo Shugu Building 中国数谷大厦, Hangzhou. Photo by Tang Min’an

Three Spring Festival Galas

Unitree values public visibility — ideally, the kind that costs nothing

The Spring Festival Gala was a defining moment for Unitree. It was also the beginning of the public’s impression of humanoid robots.

At the 2026 Gala, several embodied AI companies appeared on stage. Multiple industry practitioners told Caijing that ahead of the 2026 Gala, one such company based in Shanghai approached the Gala’s production team and raised its sponsorship bid from 30 million to 60 million yuan. Others followed, and market rumor put each company’s sponsorship fee as high as 100 million yuan. CCTV has not disclosed details of the fees. The company that bid first ultimately withdrew.

One Unitree employee was somewhat surprised. “Unitree hadn’t originally planned to go on the 2026 Gala, because we’d already gone last year,” he told Caijing, but management eventually decided to take part for reasons he does not know. “It’s always been that if [a promotional activity] required us to spend money, we definitely wouldn’t take part.” This time was an exception.

For the 2025 Gala, Unitree paid nothing. Director Zhang Yimou approached the company, and the humanoid robots were supporting players in the act. Unitree’s first Gala appearance, in 2021, also cost nothing — that year, 24 of its quadruped robots were dressed in ox costumes for the Year of the Ox and performed the number “Niu Qilai” alongside stars like Andy Lau. The debut made little splash, but it let some investors know that besides the US’s Boston Dynamics, there was also a domestic startup developing quadruped robots.

The 2025 Gala that cost nothing, in which Unitree’s robots performed a yangge folk dance in floral-print clothes, was the most important of the three appearances.3

Employees were the first to feel the impact. One Unitree employee told Caijing that after the 2025 Spring Festival holiday, overtime became the norm and he once worked at least half a month straight. Urgent capacity expansion was needed and there was not enough production space, so Unitree’s plant in the Shangtai Industrial Park had to relocate to Guanshan. Headcount expanded in parallel: only a few hundred humanoid robots had come off the assembly line at that point, but orders were already queuing up, and Unitree hired large numbers of outsourced workers for production and after-sales as a stopgap.

Unitree’s prospectus shows sales of both quadruped and bipedal robots surged in 2025. Quadruped sales reached 23,000 units, up 222% year-on-year from 7,136 units in 2024. Bipedal humanoid sales reached 5,215 units, up 1,165% from 412 units in 2024.

…

Wang Xingxing cares about how the company appears in public. One employee got a private message from Wang reminding them to be careful after they shared work experiences on social media. Unitree encourages employees to report colleagues who leak internal information, with rewards of 10,000 to 500,000 yuan once a report is verified.

Unitree employee Luo Lan (a pseudonym) told Caijing that every promotional video goes through Wang Xingxing. Not only does he review the finished cut, he also picks over the setting, shots, and even lighting before filming. Whether there is a budget at all is entirely his call.

Wang Xingxing likes to fuss over details like adjustments to a material’s color, even though these don’t affect function. They come more from his personal aesthetic taste. Some employees mention that the face of the H2 robot was designed by Wang himself. After the H2’s bionic face was unveiled last year it drew an “uncanny valley” reaction online; Wang later responded that the decision to add a human face came from user expectations.

Unitree takes its Gala performance very seriously. One employee told Caijing that for the 2026 Gala, the company stationed nearly 50 engineers on site for tuning, “living on American time every day,” for more than two months of adjustments in total. Because of venue constraints, Unitree kept more than 20 robots at the Gala site for rehearsals, while algorithm testing was done on another batch of 20-plus robots at a site in Daxing. Whenever he had nothing else on, Wang Xingxing would come to the site and solve problems alongside everyone. He personally verified every machine and “would talk through every change with every person there.”

Unitree treats the Gala not merely as “promotion” but also as an opportunity to systematically crack hard technical problems. One key breakthrough in the act was swarm control of robots performing difficult movements and complex, fast repositioning — hardware stability varies from unit to unit, so even the same algorithm running on different robots of the same brand may well not produce the same result.

It was only in the months before this Gala that Unitree’s algorithmic stability improved enough to accommodate robots with differing hardware stability. Achieving the effect seen on stage was no easy feat.

Spending Where It Counts

“Don’t try to beat us on cost — we can still go a lot lower”

Hardware and motion-control algorithms are the most-cited advantages for Unitree. The prospectus shows the gross margin on its quadruped robots rose from 43.71% in 2023 to 56.72% in 2025, while the gross margin on bipedal humanoids has held steady above 60%, reaching 63.18% in 2025.

Several robotics hardware engineers told Caijing that Unitree’s hardware advantage is not that its quality and precision are especially good, but rather that through design choices and structural engineering, Unitree drives costs down to the extreme. And through shipment volume and delivery speed, it iterates hardware development quickly.

…

Achieving the same result with a cheaper approach “is itself a demonstration of technical skill.”

Several practitioners summed up the industry’s conventional approach to cutting costs on robot hardware: on the production side, use larger shipment volumes to push down unit purchase prices. “In the long run that’s viable. But it’s still early days, real order volumes are limited, and a large enough sales scale is itself a paradox,” one practitioner said.

Wang Xingxing has been cutting costs since the company’s earliest days, and believes that relying on mass production alone to do it is a siren song. He has said publicly on several occasions that cost reduction depends first on design — the most fundamental path — and only then on mass production. In a 2025 interview he said, “A lot of people think that once I have volume I can cut costs. That’s pure illusion.” He also said: “Don’t try to beat us on cost — we can still go a lot lower.”

…

Quality control comes under pressure as a result. Luo Lan said most of Unitree’s quality inspectors are outsourced and mainly check appearance, and material quality varies between batches. “At this cost and this speed, the supply chain isn’t going to give you especially good quality control,” he said, adding that R&D staff frequently have to deal with material quality problems over and over.

One Unitree employee told Caijing that the early return-for-repair rate was “extremely high,” whereas now they can at least guarantee a product will not break within its one-year or six-month warranty. An investor with stakes in several embodied AI companies noted that plenty of firms’ robots need repairs within one or two months.

Several people in Unitree’s supply chain told Caijing that in any collaboration, Unitree is definitely the dominant party. All of them noted that Unitree iterates product designs very fast and is unyielding about details.

One supplier described Unitree’s delivery rhythm as “short, flat, and fast,” and they set up a dedicated liaison team with a working model geared to responding to design changes within three to four weeks.

When another supplier first dealt with Unitree, the order was for just ten sets. Even ten sets was hard to schedule for a company of their modest size. She believes Unitree can only work with small companies, because only they can accommodate its need for fast-response iteration; large supply-chain firms cannot even get the order into their queue. “If delivery is required in three days, many big companies need three days just for internal processes.”

The same person said that even counting labor cost as zero, there is not much money to be made — “unless the company’s strategy is very bullish on robotics and it’s determined to break into this space.” That is precisely why they want to be part of Unitree’s supply chain.

As Unitree’s production scale gradually expanded, small factories’ capacity could no longer meet its needs. The company has been both looking outward for firms able to take on current volumes and expanding its own capacity. Since 2025, embodied AI has become an explicit part of many listed companies’ strategic plans — especially firms originally in the automotive or consumer electronics supply chains, which treat embodied AI as a “second growth curve.” Numerous supply-chain companies told Caijing that Unitree’s supply chain is very closed off: they do not know which firms are in it, and they themselves cannot get in.

…

On the evening of January 4, 2026, Unitree’s official video account posted a clip of Wang Xingxing (right) at an H2 humanoid robot training site. Photo/IC

Wang Xingxing’s Trade-offs

What to do and what not to do — Wang Xingxing has a clear plan, and everyone else just carries out his orders

Unitree employee Yu Le (a pseudonym) finds Wang Xingxing’s daily routine hard to imagine. He told Caijing that Wang is still posting technical documents in work chats at two or three in the morning, generally leaves the office at three or four, and can be seen at his desk again by eight — and he comes in on weekends too.

By Yu Le’s account, Wang Xingxing is a thoroughgoing introvert. “If you run into him in the hallway he’ll walk past with his head down and won’t say hello. Sit down next to him and he’ll actually be uncomfortable.” Several people close to Unitree told Caijing that in the early days Wang refused all interviews and public appearances; only after investors pressed him on account of investor and government relations did he begin showing up in public.

When it comes to products, though, subtraction seems difficult for him.

Luo Lan said Wang cares a great deal about whether product performance beats competitors’, wanting to lead on every metric — cost, performance, appearance. Features that enterprise customers want, he wants consumer customers to have as well. “That’s impossible; you have to make trade-offs. He never makes trade-offs. His demand is ‘I want all of it.’”

On AI models, Unitree is pursuing both WMA (World Model Accelerator) world models and VLA (vision-language-action) models, an approach many robot-hardware companies share. No agreed technical paradigm for embodied AI models has yet emerged, and companies across the field are all working toward their own GPT moment. At this year’s World Robot Conference, Wang Xingxing said Unitree began work on video-generation-based world models in early 2024, shelved the effort for a period at the end of 2024 because results were not particularly good, and from 2025 began pushing its world model framework hard again.

On August 19, 2026, during the 2026 World Robot Conference, Unitree humanoid robots walk off stage after a performance. Photo/VCG

Wang Xingxing has said more than once that large embodied AI models are not yet mature. In an interview this March he noted that “video generation models — the world model direction — demand far too much compute,” such that pure robotics companies anywhere in the world “simply can’t train them.” Unitree is pinning its hopes on models with smaller data requirements.

In an early exchange with brokerage analysts, Unitree argued that sufficiently stable and reliable hardware plus motion-control algorithms can cover a certain set of industrial tasks. Several practitioners told Caijing that in some industrial settings today, the work is indeed done through fixed-point motion control and rule-based vision algorithms rather than by training large robot models.

From 2025 Unitree began expanding hiring for roles related to large embodied models. Beyond algorithms, data is the infrastructure for such models: on the recruitment page of its website, roles related to robot data infrastructure are marked “in hot demand.”

At the IPO celebration, Wang Xingxing spoke for the first time about the direction of Unitree’s advance research on the robot “brain”: self-evolving physical AI, in which robots, built on large AI models, achieve an autonomous loop of perception, decision, execution, evaluation, learning, and evolution.

…

Wang Xingxing came up through hardware structural engineering and has ten years of experience as an engineer. He drew up ten rules specifically for the structural engineering department, each a technical standard distilled from his own experience — for instance, that a given component’s thickness may not fall below a specific figure.

“He really isn’t much like what we usually understand a boss to be.” In Yu Le’s view, a company’s number one is supposed to focus on strategic direction and setting broad systems, whereas Wang Xingxing makes the call on everything from end to end, whether corporate strategy or a detail of product design.

At Unitree, the founder’s attention has become a resource-allocation mechanism, with the company’s resources increasingly revolving around one person’s judgment. Unitree once launched a resistance-training device called the “Fitness Pump.” One employee recalls that Wang, working on motors one day, realized they could be used to build exercise equipment; he quickly made the call, pulled in a few colleagues to start on it, and, convinced the product would sell well, stockpiled tens of millions of yuan worth of materials that are still sitting in the warehouse today.

Unitree also launched a boxing edition of its G1 robot, still on sale today. A person close to the company told Caijing that while preparing for a robot boxing match, Unitree modified the G1, swapping out prototype components on the fly to improve its performance in the ring. After the match, Wang Xingxing decided to cut new molds for the modified parts and push the boxing robot into mass production. In this person’s view, Wang only makes such a decision after due diligence or once orders are in hand: “Only when there are clearly orders or market demand will he put in more R&D spending.”

Unitree’s general-purpose humanoid G1 sparked wide discussion as soon as it launched in 2024. The base version started at 99,000 yuan [less than $15,000], addressing the previously prohibitive cost of humanoid robots.

Geek and Businessman

Wang Xingxing is both a businessman and a “geek.” Together, those two traits produced the “low cost, high performance” formula behind Unitree’s robots.

Employee Pang Geren 庞歌仁 vividly remembers walking into Unitree for the first time. It was the third floor of an office building in Hangzhou. Factory and company were crammed onto the same floor, a space of less than 300 square meters, “like a roadside repair stall.” The workshop was dimly lit, filled with the sounds of cutting, drilling, screwdrivers, and hammering, equipment and people jumbled together.

He recalls this as Unitree’s fastest and most efficient period. The company was still relatively small; as he sat at his desk, line workers would walk straight over, tap him on the shoulder, and bring him problems to solve. There was almost no distance between R&D and production. Later, when production moved to a larger park, he would still go down to the line frequently.

Wang Xingxing has said publicly on several occasions that before 2021, he was firmly against building humanoid robots, on the grounds that the technology could not handle so complex an integrated system and that he saw no practical value in it. In August 2023, Unitree released its first humanoid, the H1.

A person with knowledge of the matter told Caijing that an investor once asked Wang Xingxing why he was building humanoid robots, and got an answer the investor found “unbelievably plain”: because the orders came in. Another investor who met Wang at a very early stage asked who he would sell the robot dogs to once they were built. Wang responded that if he really built them, someone would definitely buy them. That investor passed on Unitree at the time.

“No compelling pitch deck, and absolutely no grand narrative about a ‘hundred-billion-yuan market of the future.’” That is how Zhang Peng 张鹏, an investor at Bianliang Capital 变量资本, describes Wang Xingxing in his earlier days. “But the moment the conversation turned to the product, the technology, and the engineering itself, he clearly had authentic enthusiasm and understanding beyond his years.”

In one interview, Wang Xingxing said Unitree’s commercial logic is that any business activity must generate a reasonable commercial profit. Cost has always been the KPI behind everything Unitree does, and the core of it is to make money.

…

Caijing sent Unitree Robotics interview requests regarding expense approvals, performance reviews, the terms of supplier confidentiality agreements, and early quality control. The company did not respond.

Unitree’s offices in the Fengda Creative Park 峰达创意园, Hangzhou. Photo by Tang Min’an

Still a Startup

Unitree’s management structure remains flat; it’s simply “Wang Xingxing and everyone else”

Unitree’s organizational structure has not grown more complex in step with the expansion of its business.

Several employees told Caijing that the company has no management layers and that a great many things require Wang Xingxing’s approval. At the same time, one R&D employee argues that Unitree’s incentive system consists only of penalties, with almost no rewards. For some staff, taking on one more feature means maintaining more things later with no additional reward. “They’ll think, why not just keep it simple?”

The same employees said R&D spends most of its time pushing projects forward, with no time for cracking hard technical problems. One person carries several projects at once, most running in parallel. The company used to focus on robot dogs, then added humanoids, dexterous hands, and assorted custom projects; new projects keep coming in, and custom work easily ties up manpower — “but there are only so many people.”

At Unitree, one person functions like a team. Per employees’ accounts, a promotional video may be handled by only one or two people, and business lines such as dexterous hands may each be the responsibility of just one or two staff. To a degree this preserves the company’s vitality; on another level it creates problems.

Employees sometimes find Wang Xingxing’s thinking hard to keep up with.

Sometimes a business team will thoroughly debate its way to an approach, only for Wang to insist on his own idea and redirect the team elsewhere. Months later, when that path proves unworkable, the company reverts to the original approach.

Several people close to Unitree told Caijing that its salaries are not high by industry standards. One former employee said pay is somewhat higher than at traditional industrial robotics companies, but on the low side compared with the embodied AI startups that emerged after 2023.

Two employees noted that raises for ordinary staff are smaller than the industry norm, while the workload is not light. One said that in the early days, on top of a 996 schedule with alternating six-day weeks, there was additional daily overtime.

A veteran employee told Caijing that last year and this year have seen the highest attrition of core staff in Unitree’s history.

…

On August 19, Unitree listed on the STAR Market, opening up 629%, with market capitalization briefly hitting 444.9 billion yuan. As the spotlight fell on Wang Xingxing and the investors ringing the bell, the gap felt by rank-and-file workers only widened.

On listing day there was none of the internal celebration Luo Lan had expected. Everyone was working — “not even a can of Coke to mark the IPO.” Luo Lan feels Unitree does not care about the company’s environment for employees or benefits. A desk at Unitree consists only of a desktop, with no cabinet or drawers. The annual party is simply a meal. Unitree’s HR does not do HR business partner work, taking no part in employees’ mental health or in company culture, and handling only recruitment and performance reviews.

Unitree’s options are locked up in the hands of a few people. According to the prospectus, the company settled most of its option grantees between 2017 and 2021, signing 1-yuan-per-share options with 17 employees, and added six more between 2024 and 2025. Nine people subsequently left and had their allocations cancelled; of the remaining 14, 11 have exercised.

In the IPO, the company set up employee asset-management plans No. 1 and No. 2, under which employees had to put up their own cash to subscribe for new shares at the offer price of 150.8 yuan per share. In the end, 171 senior managers and core employees took part in the strategic placement, subscribing for roughly 272 million yuan in total.

As Pang Geren sees it, this lets the company tell the world that employees are backing Unitree with hard cash while also raising interest-free money — and employees who have already been with the company for the long haul have to put in still more of their own. “There’s a feeling of being wrung dry.” One employee who holds equity told Caijing that he has not been able to feel happy since the listing either. “Every day I wake up and, on top of the unsolved technical problems, I have to face the stock going up and down and the sound of people cursing us out.”

The prospectus states that under the employee stock ownership plan, the upper-tier partnership interests in Shanghai Yuyi 上海宇翼 — held by Wang Xingxing, the company’s actual controller — will in future be used entirely for employee equity incentives (with Wang himself excluded as a recipient). No less than 50% is planned to be granted within the two calendar years after the 36-month mark following the IPO, with the remainder granted over the plan’s remaining term.

Unitree investor Chen Cheng (a pseudonym) has followed hard-tech for years and sees nothing wrong with a tech startup adopting a “1+N” management model — the founder and everyone else. Commercializing technical results is full of difficulties and requires a strong person to lead; if everything has to go through multiple rounds of discussion among many people, that is actually ill-suited to efficient iteration.

He believes that post-IPO, if Unitree is to become a larger enterprise, it will need to shore up the capabilities of its other executives and cultivate a “minus-one” team one level below the founder, spanning management, sales, production, and R&D.

In July 2026, Wang Xingxing appeared with the GD01 on the cover of Time under the headline “The Big Robot Moment,” A person with knowledge of the matter said Time approached Unitree of its own accord after the GD01’s launch. It had been eight years since a Chinese entrepreneur last appeared on Time‘s cover.

The Wang Xingxing in Chen Cheng’s eyes is a different figure. At the 2025 National Day military parade, Wang was invited to watch from the Tiananmen gate tower; by then, Unitree was already the undisputed leader among China’s embodied AI companies. When the parade ended, Wang walked out of Tiananmen with a black backpack, in a white T-shirt, jeans, and white sneakers, and spent more than thirty minutes walking along Chang’an Avenue to get somewhere where he could hail a cab. No assistant, no car and driver — just as he had looked the first time Chen met him, backpack on, white T-shirt on, “out doing a roadshow like that.”

Ethan Kessler was a 2025-2026 Boren Graduate Fellow for Mandarin Chinese in Taiwan. Before that, he was a graduate student and think tank researcher focusing on cross-Strait relations, critical technology, and defense acquisition policy.

Over the past half decade, small, relatively cheap drones have proven their potency on the modern battlefield. This has driven many in Taiwan to support a strong industrial policy for drone assembly and production. To many in the government, armed forces, and private industry of Taiwan, drones represent a way to leverage Taiwan’s comparative advantages and satisfy the desire among U.S.-allied countries to locate alternatives to Chinese supply chains. Given the relevance of drones to modern warfare, Taiwan’s evergreen push for Made in Taiwan (especially for defense articles), and growing Taiwanese uncertainty in U.S. commitments to arm and defend the island nation, this should be a no-brainer.

But there are tradeoffs and obstacles to Taiwan’s push for greater drone manufacturing power. And after two years of Lai Ching-te’s expansion of his predecessor’s initial drone industrial policy, these challenges are more obvious.

This piece covers:

• Why the hype about Taiwan’s drone industry meets a brick wall in the opposition-controlled legislature;

• Where Taiwan is on track, and where it has stalled, on the drone-related goals set in the first two years of the Lai Ching-te administration;

• What kinds of unmanned systems Taiwan has fielded so far;

• The four segments of the drone supply chain where Chinese dominance has been hardest to dislodge;

• The three business models defining private Taiwanese firms working on drones, and how much private capital is involved;

• The software, talent, and testing obstacles facing anyone in Taiwan looking to do “Hellscape, but affordably made in Taiwan”;

• And why Taiwan’s most realistic play is to focus on the parts inside others’ drones.

  ChinaTalk is a reader-supported publication. To receive new posts and support our work, consider becoming a free or paid subscriber.

First, the political underpinning of any large-scale Taiwanese government procurement of domestically produced (or at least assembled) drones is wobbly. The opposition-controlled Legislative Yuan 立法院 (LY), in May, passed the government’s proposed special defense budget without funding for drone procurement, leaving observers — including, crucially, potential investors in Taiwan’s burgeoning drone industry — in doubt. More recently, the opposition passed a military drone procurement bill at the end of August that rejected the six-year special budget model proposed by the government and instead requires military drone spending to be reauthorized annually. While the opposition contains a range of voices on security and cross-strait issues, its leading figures cannot be seen as opposing drone development — and, by extension, national security — outright. What they have consistently expressed is a suspicion of a special budget for drone development less subject than regular spending to legislative oversight and too high for Taiwan’s public finances.1 To many observers, their concern not only reflects a genuine worry that the government will direct funds to politically aligned firms, but also represents a clever play to both make the ruling party look incompetent and force it to call in more rhetorical back-up from the Americans, deepening domestic suspicion of exploitative Americans and their lackeys in the ruling party. This is why domestic drones were excluded from the special defense budget passed by the LY in May, which focused instead on high-end purchases of U.S. weapons systems already vetted and not as clouded by domestic graft concerns as indigenous drones.

This is one reason why the political trajectory of domestic drone funding in Taiwan is difficult to predict. The conventional tendency to back security-related industrial policy meets in the opposition a countervailing tendency to demand greater oversight from and politically humiliate the opposing party. An additional wrinkle arises due to cross-strait relations. The chairwoman of the main opposition party — who was the first sitting chairperson to visit China in a decade — has ruffled feathers in Washington with her calls for greater cross-strait dialogue and skepticism of American commitment to Taipei. Observers in Taiwan and the United States have pointed to her desire for warmer cross-strait relations to explain the opposition’s rejection of government proposals to fund military drone spending through special defense budgets. With subsequent allocations subject to the same partisan horse-trading and showmanship that delayed the first drone package’s passage – as well as additional delays from a new LY-selected oversight committee – staffers in the ruling party and industry advocates warn that the innovative speed and efficiency promised by the government’s multi-year procurement plan will never be realized.

While the LY has rejected the pronouncement by the largest opposition party’s leader that no funding for military drones would pass before Xi Jinping visits Washington in September, the drones bill passed in late August defied the government’s wish for a multi-year special budget managed by the defense ministry. With subsequent allocations subject to the same partisan horse-trading and showmanship that delayed the first drone package’s passage – as well as additional delays from a new LY-selected oversight committee – staffers in the ruling party and industry advocates warn that the innovative speed and efficiency promised by the government’s multi-year procurement plan will never be realized.

While Taiwanese drone firms have gained some initial traction in foreign markets, a second challenge arises on the point of Taiwanese and Western firms co-producing finished military drones. If Taiwan wants drones for military use, it will probably need to continue relying on Western state-of-the-art software. It is not yet clear if co-production arrangements can secure Taiwanese firms enough of that know-how, or even if Taipei should aim for such an outcome.

Lastly, Taiwan has still not decided how willing it is to address such critical issues as geographical and regulatory limitations on flight testing, a critical part of military drone development. Related is the question competitiveness. What role does Taipei envision for itself in a world increasingly desirous of drones but also more crowded with relatively low-cost component and assembly alternatives to China?

Underlying these challenges is a deeper question about Taiwan’s place in the international division of labor for drone production. As one American defense executive told me anonymously, Taiwan’s best shot is not in selling finished drones abroad, but in serving as a merchant supplier of components — airframes, gimbals, actuators — to integrators in other countries, especially to firms that sell more internationally competitive, battle-tested products.2

Building a Drone Strategy

Military Necessity

Taiwan’s pursuit of indigenous defense production does not have the same logic as that of South Korea, another East Asian “miracle” economy threatened militarily by a close neighbor. Both trace their respective defense industries to the 1970s, when fear of U.S. abandonment after the Vietnam War sparked urgency in the two most threatened East Asian U.S. allies. Both began crash programs to produce on their own what Uncle Sam had traditionally provided: rifles, missiles, tanks, and fighter jets. However, Seoul has managed to domestically produce defense articles to a greater degree and across a wider range of systems. Today, South Korea’s economic and military strength mean that the country could arguably combat North Korea alone. And the South Koreans remain able to procure exquisite foreign technology that Taiwan cannot.3

Meanwhile, many cross-Strait military analysts, including at RAND and CSIS, argue that Taiwan could not hold off a Chinese blockade or amphibious assault without U.S. help. This is unlikely to change anytime soon given the pace of Chinese military modernization and the economic growth undergirding it. Taiwanese defense indigenization efforts should therefore not be evaluated on the metric of self-reliance. Instead, Taipei’s efforts should be assessed in military terms: what is needed to fight, assuming the U.S. intervenes in a cross-Strait conflict? If a massive number of cheap drones in a variety of sizes and for a variety of mission roles are indeed necessary for Taiwanese forces to operate effectively, then indigenous production is a reasonable procurement option because foreign procurement may be too expensive or may contain too many potentially compromised Chinese components.

Producing a large number of relatively cheap systems domestically could be one half of a “high-low mix” strategy when paired with continued acquisition of higher-end reconnaissance and strike systems from the United States.4 Ukraine offers a helpful analogy: the country now produces much of its own drone and cruise missile arsenal, yet still relies on NATO allies for Patriot air-defense systems and fighter jets. The flipside is that any purchase has tradeoffs. Taiwan has other things it must pay for, such as manning and training its armed forces and buying them missiles, traditional vehicles, and other items. And a Ukraine-style drone buildout may be less useful if Taiwan cannot solve reconnaissance, surveillance, and guidance challenges. Kyiv has fought with access to overhead U.S. satellite imagery, a condition unlikely to hold for Taiwan should the U.S. withhold support in a conflict or in the event that Beijing counters U.S. intervention by degrading satellite capabilities. Taipei is at least aware of the gap — the LY eased foreign-ownership rules for satellite operators like Starlink on July 21, citing Ukraine’s wartime experience — though analysts claim that severe hurdles to Starlink entering the Taiwan market remain.

Time is the other factor. If one believes the maximum risk of Chinese aggression is in the next few years, then a drone production effort likely cannot produce fully indigenous systems within that timeline. However, if the end goal is to indigenize drone production over a decade, then the effort could succeed.5

Seizing Opportunity: Red-Free Supply Chains

Military logic can only explain so much of Taiwan’s drone push under Presidents Tsai and Lai, however. Both saw an opportunity to service growing global demand for drones at a time of growing attention toward supply chain security. Taiwan’s indigenous drone production should also be judged on this economic score. Plenty of international customers are wary of arming themselves with DJI or other Chinese-made drones, which still dominate the global market. Taipei is banking that Taiwanese firms can reap some of the profits to be made.

Public officials’ statements offer more clarity on the economic case for Taiwan’s drone push than the military one. President Lai’s speeches on the threat posed by China emphasize Xi’s directive to the People’s Liberation Army (PLA) to be ready to take Taiwan by force by 2027 as a way to rally public support for politically contested defense bills. But these speeches do not assert whether the threat of war is greater in 2027 or 2030. The Executive Yuan’s October 2025 drone development plan sketched a six-year timeline, and so far government officials have not clearly articulated that Taiwan must reach some desired proportion of indigenously produced drones by a hard date — a standard which, if not met, could prove embarrassing.

Lai’s language on the development case for drone industrial policy has been considerably stronger. In March 2024, then-Vice President and President-elect Lai toured the Asia UAV AI Innovation Application R&D Center, located in the city of Chiayi in central Taiwan. He told manufacturers that “national defense and the economy are the two pillars supporting the country’s survival and development, and the drone industry holds dual significance” (國防與經濟是支撐國家生存發展的兩大支柱，而無人機產業具備雙重意義). After assuming the presidency that year, Lai folded drones into the “Five Trusted Industries” plan (五大信賴產業推動方案), his flagship industrial policy initiative that explicitly named drone supply chains alongside chips and next-generation communications as a core pillar of Taiwan’s industrial strategy. In 2025, asked by a Japanese broadcaster why Taiwan was racing to develop drones, Lai said, “first, the need for national security, and second, the need for industrial development” (第一是國家安全的需要，第二是產業發展的需要).

This is still how Taipei frames drone policy. At a June gathering of the Taiwan Excellence Drone International Business Opportunities Alliance (TEDIBOA) (台灣卓越無人機海外商機聯盟) — a loose convening mechanism for Taiwanese drone firms established four months into Lai’s presidency — Lai reiterated that “promoting the drone industry is a major national policy” and that the policy’s two goals are “strengthening national defense capabilities” and “industrial upgrading and economic transformation.” He also touted Taiwan’s status as the first Green UAS certification hub outside the United States and restated the goal of producing NT$40 billion (roughly US$1.3 billion) worth of drones by 2030. Meeting with a U.S. Senate delegation on July 7, he said he hoped Taiwan can become “an Asian hub for global drone development” with American support and pitched joint research, development, and production as a way to advance both countries’ defense capabilities and economic ties. Premier Cho Jung-tai 卓榮泰 has used nearly identical language, saying earlier this year that the goal is “to build Taiwan into an Asia-Pacific hub for the democratic drone supply chain.”

The metrics and goals below capture how much Taiwan’s drone sector is growing in absolute terms and how well Taiwanese firms are embedding themselves in trusted international supply chains as component and subsystem suppliers — not whether Taiwan is becoming a globally competitive exporter of finished, Taiwan-branded drone systems, something infeasible given the dominance of experienced, battle-proven incumbents. Taiwan’s 2025-2028 National Development Plan, along with other government documents and speeches, provides a foundation for assessing domestic drone production.

For all four of the main strategies offered by Premier Cho at the October 2025 Cabinet meeting that presented a comprehensive plan for the drone industry, assessing progress is still difficult. The first strategy, to expand domestic and foreign demand, remains uncertain because it is unclear whether Taiwan’s drone producers have attracted and invested enough capital in the scale needed to drastically grow overall output and thus meet the first metric listed above. The October 2025 EY plan’s NT$44.2 billion (US$1.44 billion) procurement package and the Ministry of National Defense’s (MND) since-superseded proposal for 200,000 more military drones both reflected a government view that 2025 procurement orders were insufficient, necessitating stimulus for the industry.6

The second strategy, to boost Taiwanese drone-makers’ technical prowess and connection with foreign partners, has indeed produced many memoranda of understanding (MOUs) aligning Taiwanese manufacturers with Western software and communications infrastructure providers, but the detailed terms related to technology transfer of partnerships remain proprietary secrets.7 The third strategy, to build out industry clusters, and the fourth, to modernize drone-related regulation, have together produced Chiayi’s Asia UAV AI Innovation Center. Though it is now home to several dozen resident firms, real capacity remains limited since flight testing is still constrained.

Through the last few years, “How much money is Taiwan’s government throwing at this?” has been a central question, given the insufficient size of the Taiwanese civilian market for quickly building and then sustaining a capable drone industry. Political wrangling over the last year between the DPP-controlled Executive Yuan and opposition-controlled LY has muddled the answer. This June, following the stripping of drone-specific funding from the government’s large special defense budget proposal in May, the EY introduced a separate, six-year special budget for coastal surveillance and attack drones and small suicide unmanned surface vessels, funded through debt issuance rather than the annual budget. The Kuomintang (KMT) and Taiwan People’s Party (TPP) countered on July 1 with drone bills of their own that routed funding through the general budget and proposed a standing drone committee to provide greater legislative oversight. For most of 2026, actual government funding for Taiwan’s drone industry has therefore been a question mark, contingent on the outcome of yet unresolved legislative negotiations.

Over this summer, the opposition parties would concede on the KMT’s proposed domestic content rules and some other provisions, but ultimately pushed through a version of the government’s proposed drone funding that removed what made it unique: special budget authority allowing for normal fiscal limits to be bypassed, multiyear procurement promising a large stimulus for private investment, and MND leadership on the acquisition of military drones to wall off defense acquisition from industrial policy pressures. On August 14, the opposition-led LY passed the civilian drone package proposed by MOEA ten months earlier, worth NT$44.2 billion (US$1.4 billion) and set to last until 2030. Two weeks later, on August 27, the opposition passed the “Act to Strengthen National Defense Self-Reliance and Promote the Development of the Unmanned Vehicle Industry” 60-50. Set to enter force in six months, the bill authorizes funds for a headline figure (NT$240 billion, or around US$7.5 billion) larger than the EY’s special budget ask across six years, and covers all domains of drones. Yet it retains the main tension of industrial policy for military drones by assigning MOEA as lead overseer while leaving MND in charge of procuring said drones — an arrangement MOEA has protested.

Serving two masters is difficult, and catering to both the needs of Taiwan’s drone industry and the best interests of its armed forces simultaneously is one of the main reasons that supporters of the EY’s special budget proposal have criticized the opposition-passed bill. The lack of a special budget for the drone funding also means that the government must come up with the money anew each year, and the LY must authorize it anew each year. This, and the opposition bill’s insertion of a LY-appointed review committee, adds two stages of vetoes (and thus leverage) for the opposition, allowing it to extract concessions elsewhere until the DPP regains control of the legislature — a prospect that can come no later than February 2028.8 The Sunday following the LY’s passage of the bill, a source said that Premier Cho would countersign the act and request funding for it via a supplemental budget. Assuming that he does so, the main question remains: will Taiwan’s drone industry be able to attract sufficient investment and develop at the speed necessary to become a drone hub? Or will nearly a year’s worth of partisan delays and deviation from the special-budget, multiyear procurement model prevent the scale and innovation needed for this goal? The scale of investment and speed of innovation are certainly important, but other factors — like Taiwanese firms’ role relative to other key drone companies and testing regulations in Taiwan — will matter as well.

The State of Taiwan’s Drone Arsenal

Taiwan’s interest in remotely piloted vehicles for military use goes back two decades, but only after the Russia-Ukraine War did Taipei begin to focus on procuring attritable drones in large numbers. Before then, Taiwan’s drone force was not exceptional: like other countries, its drones were either relatively high-cost, expensive platforms not designed to be attritable or were loitering munitions that fit within long-established concepts of employment.

However, in summer 2022, President Tsai Ing-wen was reportedly spurred by the Russia-Ukraine War to create the Drone National Team, which aimed to procure 3,200 military drones from domestic suppliers by mid-2024, and the Drone Defense National Team for procuring counter-drone systems. While the 3,200 goal was not met — MND procured only around 1,000 units in the first half of 2024 — by fall 2023 the national defense review proposed a new model: having domestic firms produce mainly commercial drones to enable them to grow beyond the demand provided by Taiwan’s military. (The goal was 7,700 drones by 2028, 700 military and 7,000 commercial.) Last October, Lai raised the government’s purchase goal to 48,750 military drones by 2027 and 50,000 civilian drones by 2030.

The military drone objective was later raised in January 2026 to 200,000 units, but the government’s ability to meet that goal is uncertain following the exclusion of funding for Taiwanese drone purchases in the supplementary defense budget passed by the Legislative Yuan on May 8. While five-digit production numbers are a substantial step up over four-digit ones, this is probably insufficient for addressing China’s drone capacity. By the end of 2025, 3.29 million drones were registered with China’s Civil Aviation Administration, and DJI alone controls roughly 70% of the global drone market — a massive manufacturing base that can support substantial military drone production.

The National Chung-Shan Institute of Science and Technology (NCSIST) (國家中山科學研究院), Taiwan’s state defense technology agency, started working on drones in the early 2000s. Its Chung Shyang II (銳鳶) “Albatross” reconnaissance Group 3 UAV, used by the Republic of China Navy (ROCN), debuted in 2007. It is not attritable — only 32 were built. The size, weight, and layout are similar to Israeli drones designed in the 1970s and 1980s (IAI Scout and IAI Searcher), whereas NCSIST’s larger Teng Yun (騰雲) “Cloud Rider” reconnaissance Group 5 UAV resembles the U.S. MQ-1 Predator.9

Apart from large reconnaissance and strike drones, Taiwan’s indigenously produced drones also include loitering munitions of varying sizes. NCSIST’s Chien Hsiang (劍翔) “Rising Sword,” a delta-wing loitering munition similar in appearance and size to the Israeli Harpy and Harop Group 3 systems, first appeared at the 2017 Taipei arms show and was seen test-firing from its 12-canister launch truck in 2023.10 The overall number of Chen Hsiang systems procured has been small. In November 2022, it was reported that NCSIST would produce 104 units by 2025. That target was later augmented with a goal of 192 to 200 second-generation units (up to 48 per year, from 2022 to 2025), which have the ability to launch from ships and reportedly have a range of 1,000 kilometers.

After the 2022 pivot, Taiwanese military leaders reportedly assessed that relying solely on U.S.-procured drones would be inadequate for Taiwan’s defense, leading to the development of “Type I and Type II” suicide drones. The backpack-carried, tube-launched Type I/Ching Feng I (勁蜂) “Mighty Hornet” attack drone is similar in size and appearance to the U.S.-made, Group 1 Switchblade-300 infantry-deployed system, with a reported flight time of 15 minutes and human-guided control range of eight to ten kilometers. NCSIST also claims it can be deployed from a launch tube mounted to the Albatross UAV. The Type II (or Ching Feng II) drone, which appears to either be Group 2 or 3, debuted at the 2023 arms show in a configuration featuring three units mounted in a launcher on a Humvee, and reportedly takes a crew of two or three to operate.11

Other smaller drones include the Group 1 Cardinal scout drone, whose “Fire Cardinal” (火紅雀) suicide variant was revealed at the 2019 Taipei arms show; the Ching Feng III X-wing style, VTOL, anti-armor kamikaze drone, recently tested with Shield AI’s autonomous guidance system; and the Ching Feng IV, a loitering munition developed with U.S. firm Kratos around the latter’s existing target drone design and tested successfully in February 2026 (though not yet procured by the MND). These systems have all been developed by NCSIST but incorporate assistance and components from private firms. NCSIST’s president announced plans in December 2025 to expand production of explosives in Taiwan, partly to supply these attack drones.

These systems, all past the prototype phase, have now been superseded by NCSIST’s newer “A through E” series of indigenously produced drones, all included in the government’s plan to procure 48,750 military drones by 2027. Drones spanning these five categories — estimated by DSET to have totaled 209,646 units in the government’s last draft proposal of the special defense budget — were not included in the May 2026 special budget passed by the LY.

In the naval domain, NCSIST’s Kuai Chi (小型快速無人艇案) light USV reportedly passed tests in August 2025, with other designs from private Taiwanese firm Thunder Tiger and Taiwan’s state-backed shipbuilding firm less far along. Each Kuai Chi unit is reportedly around three tons, capable of firing onboard munitions or ramming enemy vessels and then detonating explosives, and designed to fit in a 40-foot shipping container. In August 2025, NCSIST tested integration of Kuai Chi drones with onboard Ching Feng I “Mighty Hornet” kamikaze drones and overhead Chung Shyang II “Albatross” UAVs for guidance. Taiwan’s Naval Command Headquarters specified that the final winner of its USV contract must have swarm capabilities and an operational range of at least 44 km (24 nautical miles), according to the command’s bid opening announcement last December. MND’s procurement goal rose from around 1,300 units to 1,500 units given additional interest from the Coast Guard in late 2025.

In addition to indigenous drones, long-range intelligence, surveillance, and reconnaissance UAV capabilities are set to be bolstered by four MQ-9B SkyGuardian drones ordered from the United States in 2020, with the first shipment delivered this year. Taiwan also announced in late 2024 it would procure 291 Altius 600M-V loitering munitions from Anduril and 685 Switchblade-300 loitering munitions from AeroVironment (now AV Inc.). Despite lawmakers’ concerns with the performance of Anduril systems in Ukraine, Taiwan’s appetite for Anduril products has grown dramatically since then.12 The special defense budget passed in May 2026 includes NT$35 billion (US$1.1 billion) of funds for 478 Altius-600ISR and 1,554 Altius-700M drones. (Altius-600ISR is Group 2 and designed for ISR and communications missions, while the larger, Group 3 Altius-700M carries explosive payloads for kamikaze missions.) And Anduril’s Barracuda-500 cruise missile, to be manufactured by NCSIST in Taiwan and designed to be much less costly than competitors, has been a key part of the defense tech firm’s focus on scalable fires as it continues to cater to U.S. war-fighting requirements in the Indo-Pacific and expand into Japan and Taiwan.

It is hard to evaluate recommendations for how these systems ought to be deployed because Taiwan’s military has a much less transparent requirements process than the United States, and needs vary by system.13 While the Kuai Chi USV’s procurement is to be overseen by Naval Command HQ, the MND said in 2025 it will be used by the Navy’s Coastal Combat Command, the Marine Corps, and the Army’s Special Operations Command, each of which has different wartime functions. Moreover, UAVs are further along in development than other drone-related systems. Taiwan’s defense minister recently stated that unmanned ground vehicles would likely appear in next year’s regular defense budget request and that military-wide command and control systems for counter-drone operations—a domain marred by reported flaws with NCSIST-developed systems—are still under development.

AI will make it easier to design dangerous pathogens, but those designs still need to pass through the physical world. Nucleic acid synthesis providers are one of the few chokepoints between digital biological know-how and actual biological capability, as we’ll see in the next section.

If policymakers are looking for a concrete way to reduce AI-enabled bio risk, screening these orders is a good place to start.

Subscribe now

Washington has already half-promised to act. Last year, the Trump administration nixed the Biden-era nucleic-acid synthesis screening framework, promising to replace it with something even bigger and more enforceable. Then biosecurity largely fell off their radar. But after the Fable ban (and then un-ban), as well as the OpenAI–Hugging Face fiasco, there is a growing recognition that frontier AI risks — including biological risks — may not be purely theoretical after all.

Cotton (R) and Klobuchar (D) also have a bill working its way through the Senate that would force Commerce and OSTP to act on synthesis screening, and there is a narrower House bill from Pfluger (R) and Houlahan (D).1

The US needs to follow through on this. But America is only one node in a global biotech supply chain.

China accounts for ~34% of the world’s DNA synthesis providers. Its AI ecosystem is more open-source and thus harder to govern through model-level safeguards. And it has done a worse job at ensuring its AI models don’t spit out biorisks. An America-only screening regime would reduce risk, but it would not eliminate it if customers could simply route orders through providers in more permissive jurisdictions.

Synthesis screening belongs on the agenda of the upcoming US-China AI talks.

Synthesis screening is narrow, relatively verifiable, and plausibly a win for both sides. It also imposes little cost on AI development itself, since the burden falls primarily on synthesis providers rather than model developers. Rather than asking Washington and Beijing to sacrifice core interests or verify each other’s promises, it asks them to adopt a safeguard that neither country has very strong reasons to oppose.

Share

Nucleic Acid Synthesis Screening, Explained

Nucleic acid synthesis screening is essentially a security check for custom DNA and RNA orders. When someone orders a sequence, the provider compares it against databases of dangerous pathogens and toxins, checks whether the customer is suspicious, and decides whether to fulfill, delay, or reject the order. The goal is not to review every biology experiment, but to stop dangerous genetic material from being shipped without scrutiny.

Imagine a researcher wants to synthesize a protein from the smallpox virus, one of history’s deadliest pathogens. The relevant genetic sequence has already been published, so they could copy the DNA code corresponding to that protein and send it to a synthesis company. And as AI-enabled biological design improves, the concern is not just that users could copy or tweak known pathogens, but that they may increasingly be able to design new biological sequences optimized for lethality or transmissibility.

The fact that we don’t already regulate this is pretty wild and is why many of the major synthesis companies screen voluntarily. But not every provider participates, it’s hard to actually enforce since these providers have limited resources and compulsion, and a single underregulated supplier can become the loophole that defeats it for everyone.

Biosecurity is asymmetric in a way many other safety problems are not. A system can work well 99.9% of the time and still fail catastrophically if one competent bad actor finds the gap, or if one legitimate lab makes a serious mistake. The many harmless orders do not cancel out the one dangerous one. And past reporting has exposed suppliers willing to ship dangerous pathogen fragments with embarrassingly little scrutiny.

Subscribe now

AI model safeguards help, but they are easy to get around, and can force labs like Anthropic into blunt fixes like automatically routing users to weaker models (which made researching this piece much more annoying). Screening the synthesis providers instead targets the supply chain before a sequence becomes a shipped product.

But screening is not a panacea. Benchtop synthesizers (personal synthesis machines small enough to fit on a lab benchtop), fragmented orders, and overseas suppliers can weaken the system. But benchtop synthesizers, for instance, are not yet magic pathogen printers, and other workarounds are limited by cost, sequence length, and the economies of scale that help set the main synthesis providers apart. Screening commercial synthesis orders will not close every route from screen to lab bench, but it can close one of the easiest ones.

Is China Down for Synthesis Screening?

China, like the US, does not have a mandatory national order-screening regime yet, but it is not starting from zero. Some of the biggest China-linked synthesis providers already participate in voluntary screening systems. The International Gene Synthesis Consortium (IGSC) is a great (albeit limited2) industry group whose members commit to screening gene orders and customers. It added BGI Bio-Solutions 华大基因 in 2017, and GenScript 金斯瑞 is also a member. Chinese companies/experts have also worked with the International Biosecurity and Biosafety Initiative for Science (IBBIS), a biosecurity nonprofit trying to make synthesis screening easier and more standardized across countries.

Synthesis screening is also a topic Chinese researchers have been raising more frequently.

Share

The AI case for screening may be even more compelling for China than the US. Chinese frontier models are comparable with leading Western models on dual-use biocapability benchmarks but ship with significantly less effective safeguards against misuse. And China’s AI ecosystem is more open-source, which makes bio-concerns harder to govern — once the weights are released, users can modify, jailbreak, or run them locally without the original lab’s safety filters. China is pushing hard on both open-source models and biotechnology, which could be a recipe for disaster. This gives them even more reason to strengthen the downstream chokepoint where digital sequences become physical DNA.

On practical implementation, China may also have an advantage. The hardest part of synthesis screening isn’t the sequence check but verifying that customers are trustworthy, and China already runs something close to KYC on most of its population.

And let’s not forget the most obvious point: Xi does not want COVID 2.0 coming out of China! Because of this, as I discussed in a previous article, China has taken recent steps to increase its pandemic preparedness, such as the Biosecurity Law (生物安全法), which has a few sections that could plausibly be interpreted as calling for screening.3

Screening Doesn’t Slow Biotech

Some will worry that synthesis screening could slow biotech progress, and that if the US tightened its rules while China did not, Chinese firms would pull ahead. But I don’t think this is actually much of a concern.

Nucleic acid synthesis is one tiny step in a much longer pipeline. Ordering DNA takes days, whereas developing a drug takes years. The real bottlenecks are things like failed constructs, cell work, animal studies, clinical trials, manufacturing, and regulatory review.

Screening itself is also getting cheaper. The software already exists. IBBIS’s Common Mechanism gives providers a baseline tool for sequence screening. Customer screening costs more, but even that is dropping fast. As IFP argues, AI-assisted KYC could push those costs down enough that customer checks become a small fraction of a normal gene-length order’s price.

Firms representing roughly 80% of global synthesis capacity already screen voluntarily. If screening were a serious competitive drag, the companies with the most to lose would be the ones avoiding it, not signing on. Making screening mandatory mostly raises the floor for the remaining 20%, which arguably helps the major providers by adding stability to the market dynamics, since today they can be undercut by smaller, less scrupulous suppliers who skip screening.

Finally, the cost of screening is trivial to the cost of getting this wrong. At its peak, COVID was burning through the US economy at roughly $26 billion a day. Every day a vaccine arrived sooner was worth tens of billions of dollars. (The Centre for Long-Term Resilience put numbers on this for the UK and found every £1 spent on screening returns £3.50 in averted losses.) Any intervention that lowers the odds of another pandemic, even slightly, pays for itself many times over — to say nothing of the ethical imperative of saving millions of lives!

Subscribe now

A Way of Thinking About Cooperation

The US and China cooperating on anything is difficult. Part of the problem is that many proposed forms of cooperation require mutual verification. If one side upholds its end of the bargain while the other cheats, it can end up at a disadvantage.

That is why, when thinking about US-China AI cooperation, I find it useful to distinguish between contingent cooperation and demonstrated cooperation.

Contingent cooperation only works if both sides act simultaneously. In arms control agreements, for instance, if one side limits itself while the other secretly defects, the cooperator ends up weaker while its rival gains a military advantage. Demonstrated cooperation consists of each side independently taking actions that advance a shared objective (one they want to pursue anyway) while providing evidence of its commitment to that objective. Over time, both countries gradually accumulate trust through repeated acts of demonstrated cooperation that could establish a foundation for progressively solving harder — more contingent — problems.

Matt Sheehan articulates something similar in this post:

“I believe that (at least for now) the most important actions on AI governance and safety will be domestic actions that the U.S. and China take — not international or bilateral agreements. There may be some level of light touch coordination, but it’s unlikely to be “We do X because you’ve promised to do it too.””

Synthesis screening strikes me as an opportunity ripe for demonstrated coordination. Both countries independently don’t want a Covid-like pandemic emerging from within their borders. Cooperating doesn’t detract from this goal; it supplements it. It gives both governments something they can do independently.

Share

The US can require American providers to screen orders; China can require Chinese providers to do the same; both can point to domestic signals as evidence that they are taking AI-bio risk seriously.

This framing also avoids one of the usual traps in AI safety politics. Safety is often cast as a brake on progress; why should China/US accept rules that might lock in the other side’s AI advantages? Synthesis screening does not fit that frame, since it is not about making models weaker or slowing AI labs down. Rather, it is about securing the physical supply chain that turns biological designs into biological material. And as explained in the previous section, it won’t meaningfully slow biotech either.

For a US-China AI safety agenda that will otherwise be full of abstractions, nucleic acid synthesis screening is tantalizingly practical.

How to Actually Do It

The first step would not need to be a grand US-China biosecurity treaty. It could be a shared baseline — relatively minimal — clarifying which kinds of DNA and RNA orders should be screened, what sequence lengths count, what categories of pathogen or toxin sequences trigger review, and what customer information providers need to verify.

Some of these measurement questions could be tricky. For example, should the metric they report to each other be measured by the share of providers that have passed third-party verification, or by the share of orders that are actually screened? The latter may be the more accurate measurement [like how counting FLOPS is better than counting chips for estimating compute capacity], but many private companies are reluctant to disclose order-level data. The former is easier to report, but it can be statistically distorted by the behavior of numerous small providers. Neither metric is perfect. But the important point is that modest differences over questions like these would not be the end of the world, so long as each side reported enough for the other to recognize that meaningful action had actually been taken.

Tools like IBBIS’s Common Mechanism could also help make this less abstract by giving providers a common technical starting point rather than asking every company to invent its own screening system.

The second step would be talking to each other, ideally starting with Track 1 happening in a few weeks! Both countries could report how many orders their providers screen each year, how much of the market is covered, or how many orders are escalated for human review. My kind of arms race is one where each side competes to prove it is screening more thoroughly than the other.

Subscribe now

The harder part would involve sharing information about dangerous orders that are actually flagged. That would be more useful, but also more politically sensitive. No one likes to brag about how many potential bioterrorists they have in their country, and China in particular has an MO of trying to sweep biosecurity blunders under the rug. So the realistic starting point may be the understanding that each country is dealing with their flagged customers privately.

Trusted intermediaries could help. Organizations already involved in this work, such as SecureDNA and IBBIS (both Swiss and therefore relatively neutral) as well as IGSC, NTI, and SAIF, can help develop standards, host technical dialogues, and build confidence before the issue becomes too politicized. And new players like Carnegie or Brookings could leverage their existing dialogue networks to begin feeling out the vibes with the Chinese players.

What not to do

What I don’t think it would be wise to do is route the whole thing through the BWC, WHO, or other official international bodies. Those bodies tend to make cooperation more contingent and more politically fraught. For instance, if nucleic acid synthesis screening is immediately turned into a BWC-style treaty-process question, it is at risk of inheriting all the hard problems of verification, sovereignty, accusations of cheating, and disputes over who gets to inspect whom that have pushed both the US and China away — especially since China is not exactly eager to report sensitive biosecurity problems to international institutions. Starting with something bilateral seems better.

None of this eliminates the need for harder, more contingent cooperation. One obvious problem is split ordering: a bad actor could try to evade screening by ordering different fragments from different providers, possibly in different countries. Solving that may eventually require more direct information-sharing. The same goes for regulating benchtop synthesizers. But I believe we need to first build the habit of parallel demonstrated action. Prove that both sides can screen orders, verify customers, and report aggregate progress. Then, over time, work toward the harder cases.

Share

Translation by Fable 5.1 and checked by Irene. Footnotes added to explain some historical references.

These days, the situation is excellent with Chinese models, and new releases take turns to farm Opus 5 like a boss monster.1 But we still have to see the situation clearly and should not deceive ourselves

A theory of quick victory and a theory of defeat are two sides of the same coin, and both are very bad actions.

We have to first get the sequence straight. In early June, the evil Anthropic released the powerful Fable 5. Only then did many domestic models with benchmark scores approaching Fable 5 appear.

The basic picture, in which OpenAI and Anthropic hold a large lead, has not changed. The premise is that our large models still lag behind; only then can we debate whether the gap is six months or four.

Whether from the viewpoint of model performance or parameter count, K3 and Qwen 3.8 Max are still a full major generation behind Fable 5, and with GPT 6 about to be released, the gap may widen further.

The progress of the past two months shows up mainly in how much faster domestic models are following up. With the successive releases of GLM 5.2, K3, and Qwen 3.8 Max, we can confirm one basic fact: domestic large models are improving at an accelerating pace, and the gap between open-source and closed-source models is, for now, narrowing.

The most critical variable here is open source.

Open source has terminated the logic of the theory of defeat. It has replaced “America encircles Chinese AI” with a protracted war in which open source encircles closed source.2

The logic of the theory of defeat is very simple. America has the largest amount of the most advanced compute, plus OpenAI’s first-mover advantage, and large models are a business where, once the technical path has been proven, things mostly come down to piling on compute and scaling up data.

In terms of compute, Nvidia GPUs shipped to China inevitably suffer losses along the way. In terms of data, the US was also first to build out high-end human annotation. Synthetic data, like using transfer stations to hit the Claude API and grind benchmarks, is bound to be lossy too. As long as we respect the law of conservation of energy, Chinese large models cannot possibly catch up to America.

That logical sequence is all correct. But if everything ran on logic, there would be no such thing as entrepreneurship, and everyone would just clock day in and day out to tighten screws.

I put together a timeline of the major open-weight releases of large models.

It is obvious at a glance: open-source models have gone from being mostly small experimental models in the early days to, recently, three open-sourced models approaching Opus 5. Open-source performance is closing in on top closed-source models faster and faster.

Chinese-vendor-led open-sourcing of large models has changed the balance of power between the two sides. Open source has enlarged the base of the AI industry and consolidated every AI team outside the closed-source labs into a united front.

The most recent example: the release of K3 triggered a knee-jerk reaction from the US government, which planned to restrict American companies from deploying open-source models. More than 270 tech companies and organizations, including Nvidia and Microsoft, published an open letter supporting open-source models, and then, for now, nothing happened. K3 was not banned.

Clearly, if K3 were a closed-source model, this many companies would not have come to its defense; at most Jensen would say a few words to sell GPUs. The whole world under heaven has suffered under Anthropic for a long time. Acting in their own interest, commercial companies will inevitably back open-source models against the evil Italian, to stop the latter from swallowing everything.

We can already see this trend today. Fable 5 is so expensive that even tech giants cannot open it up for unrestricted use by all employees. Everyone uses Fable 5 for a handful of complex jobs, and the vast majority of tasks are done by relatively cheap models. The latter is a ready-made ecological niche for every open-source model.

What will inevitably happen in the future is that the vast majority of token demand will be served by open-source models, and only a tiny number of complex scenarios will require expensive, strictly moderated closed-source models.

Small open-source models in particular have dramatically changed the meta of the game.

Open-source models fully clearing the usability bar and reaching somewhere between Opus 4.6 and 4.8 is something that happened around GLM 5.2. And just two months later, the small Qwen 3.8 model, with only 27B parameters, has also reached Opus 4.6 level, matching the strongest model of six months ago.

Qwen 3.8 27B is a crazy model. We know Zhipu likes to talk about its high “intelligence density,” reaching performance close to K3 with 744B parameters. From this came the debate over whether large models must chase large parameter counts, and how to measure the intelligence-to-efficiency ratio.

First, if you read Uncle Tang Jie’s full tweet, you will find he did not actually land on a verdict. He only said that at this stage Zhipu has chosen to keep improving base-model performance through post-training, and that in the future they will also try pre-training and all sorts of other things. It works whichever way you read it.

Second, by this logic, the poster child of intelligence density is not Zhipu, but old-timer Alibaba. Qwen 3.8 27B has only 1/28 of Zhipu’s parameter count, leading its peers in density by an order of magnitude, and it is still far ahead even if you count by active parameters.

So here is the question. GLM 5.2 was the post-training king of the last meta. Should Qwen 3.8 27B be the pre-training king of the new meta, or the post-training king, or should we just call this thing number one in intelligence density?

So intelligence density is a concept that does not survive close scrutiny. Enjoy the noise and move on. The ticket into the new season is still building a large model of around 2T parameters.

Where Qwen 3.8 27B changes the meta is that a model fully surpassing human coding ability has become a true public good. Even DeepSeek V4 Flash is hard for an individual to deploy, whereas Qwen 3.8 27B is a model that can be run entirely on consumer-grade GPUs.

I have seen the hardware nerds having a great time these past few days, digging out Mac minis and 4090s to deploy all kinds of quantized versions; one guy even did it on a modded 3080, running 27B for a bit over 3,000 yuan. A scene as beautiful as the great lobster renaissance.

All I can say is: vigorous life, all things striving forth. 27B versus 744B, the advantage is mine.

My main point is that Qwen 3.8 27B has accelerated the trend of intelligence becoming as cheap as cabbage.

My imagination is limited; I do not know what happens once intelligence is cabbage-cheap. I can only lean on one fun case: Xiaozhi AI 小智AI, the AI hardware device that’s shipped over a million units, was born because its developer, Brother Xia, happened to have a few 4090s on hand.

Qwen 2.5 happened to be open-sourced, and after deploying it locally, Brother Xia tinkered his way to an open-source chatbot solution beloved by hardware nerds. Most AI chat hardware today still uses Xiaozhi’s open-source design.

Frankly, once Opus 4.6-level models are fully cabbage-cheap, someone will pull off something far wilder than Xiaozhi. It just needs a bit of time to arrive.

I’ve spilled so much ink on Qwen 3.8 27B because it is a model born entirely for open source. Open source is good not as a moral slogan, but because open source has actually accelerated the improvement of large-model capabilities.

Originally, Chinese large models were certain to lose to American ones; the logic of the theory of defeat held. Open source changed the meta of the game. From then on it has been open-source models competing against closed-source models, each side entangled in the other.

“Entangled” meaning: when American tech companies deploy K3 and Qwen 3.8 Max, they also pay Kimi and Alibaba. Open source not only guards against bans, it also drops gold coins.

Trade space for time, and closed-source models will always have a chance to fall behind. The phase of large-model competition has thus shifted from the theory of defeat to protracted war.

What is bound to happen is that in the future 90% of tokens will be served by relatively small-parameter models, enough to handle the vast majority of agent tasks, and only the remaining 10% will be supplied by ever more expensive, ever larger SOTA models.

This trend is positively correlated with the spread of agent products. Caring about model version numbers is a very niche need; ordinary users only need to pick a model tier inside the product and do not need the strongest, most expensive coding model.

LLMs are basically a kind of manufacturing, and in manufacturing efficiency is everything. A small-parameter model whose performance is fully usable is obviously more efficient than Fable 5.

This is the meaning of protracted war. No quick victory is possible, nor will there be a quick defeat. Instead, in the stage of strategic stalemate, we wait to see what variable emerges that raises one side’s efficiency and lowers the other’s, until the tipping point arrives and one side collapses first.

Of course, I am not saying open source makes domestic models invincible. You can see closed-source models accelerating too, and St. Musk’s Grok, so desperate to improve, happens to be the control group.

Grok is a case of a closed-source model with ample compute and ample data. From St. Musk’s flailing micromanagement to dutifully training on Cursor data, Grok’s performance has actually improved very quickly. Grok 4.6 is itself a 2T-parameter model, on par with K3 in performance, faster and cheaper, and currently the most cost-effective third-best model in the world.

Moreover, Grok is also shipping a minor version every month, and its momentum in chasing Gemini is no slower than Kimi, Qwen, or GLM. There is a high probability that the next version of Grok significantly surpasses this generation of domestic models.

The pace of progress of the open-source camp as a whole is not necessarily faster than closed source.

Right now is the protracted-war period for open-source models, and also the most difficult period. The folks at the model labs are burning through their lives; open Twitter at night and Gemini and Grok are updating constantly too. The competition never ends; there are only small wins, no big win.

The other key tenet of protracted war is building base areas.3 Translated into large-model industry terms, that means domesticizing compute.

This is why, although Meituan’s LongCat is a second-and-a-half-tier domestic model that cannot even beat Tencent’s Hunyuan 3 in the free-token race, I still feel great warmth toward it.

Because LongCat is the first model to complete the entire training-and-inference pipeline on domestic chips. Although its performance is not there yet, and although it most likely used domestic cards throughout because it lacked resources, it was the first to close the loop between domestic models and domestic compute. That is building a base area.

As long as there is a base area, you can always grow slowly and become strong.

Open-source SOTA models like K3 and Qwen 3.8 Max can all run inference on domestic chips already in mass production. Just a couple of days ago, Wu Yongming said that Alibaba’s second-generation supernode, which begins tape-out and production in the second half of the year, can fully substitute for large-scale model training.

It seems that, alongside pushing the ceiling, everyone lending a hand to help domestic chips achieve integrated training and inference is not out of the question either.

This fundamentally undercuts the logic of the theory of defeat. It is not the case that we must forever rely on Nvidia chips. Once the entire large-model chain closes the loop and efficiency gradually improves, there is no reason what happened in solar and new energy will not repeat in AI.

My main takeaway is that the big companies are truly resilient. While Meituan and Alibaba slug it out in a war of attrition, they still manage to do AI with flair. Even [Meituan’s] Xiaoxiang Supermarket is pretty good: when I wanted fruit and milk and whatnot at 10 p.m., [Alibaba’s supermarket brand] Freshippo and Sam’s Club couldn’t deliver, but Xiaoxiang could 👍

Maybe these two companies have reached a game-theoretic equilibrium where they have to fight a war of attrition to do AI well. If the food-delivery war ended one day, the bosses would look at all that extra free cash flow, not know what to do with it, draw their swords and stare blankly in every direction, and that might actually hurt large-model training.

The future of open-source models is bright.

Because open source has real interests and the people’s hearts on its side. In the words of the Great Leader: “Weapons are an important factor in war, but not the decisive factor; it is people, not things, that are decisive. The contest of strength is not only a contest of military and economic power, but also a contest of human power and morale.”

But the road for open-source models is arduous and winding.

Another line from the Great Leader: “The united front must be maintained; only by maintaining the united front can we keep up the war; only by maintaining the united front and keeping up the war can we achieve final victory.”

(Cover image generated by ChatGPT; text written entirely by a human.)

As the next Trump-Xi summit approaches, details on the AI safety dialogue the two sides announced in May are scant. Past efforts have devolved into energetic airings of grievances, or have been derailed by unrelated issues. Trust remains low on both sides, with Washington unsure of China’s aims and Chinese state media claiming in a recent op-ed that the United States has conflated AI safety with a geopolitical blockade. There’s no guarantee that serious talks will proceed, let alone tackle the technical issues and political dynamics that underpin risk management.

But as a veteran U.S. diplomat now working on AI safety, I know that talking is the first step. Both countries know the stakes are high. And there’s much that the two sides can do to prepare for the meeting between President Trump and Xi this month. Here’s how they should start.

First, know the risks

The U.S. and China have seen more than their share of failed talks — risking failure for diverse reasons.

Political crises have felled many such dialogues. Following House Speaker Pelosi’s 2022 visit to Taiwan, China suspended a series of military and law enforcement dialogues as well as climate change talks. Prior incidents like the EP-3 collision with a Chinese jet and the NATO bombing of China’s Embassy in Belgrade triggered major pullback.

Mechanisms established on paper can go unused. During the 2023 spy balloon incident, Secretary of Defense Austin attempted to reach his counterpart via the Defense Telephone Link, which was established for crisis communication. The PLA did not connect the call. This could be a prologue for disuse of a hotline in a politically sensitive AI-related incident.

Deliverables that sound good can be unfeasible to enforce. Xi’s 2015 state visit to the United States was marked by a short-lived and unenforceable agreement that “neither country’s government will conduct or knowingly support cyber-enabled theft of intellectual property, including trade secrets or other confidential business information, with the intent of providing competitive advantages to companies or commercial sectors.” Neither side trusts the other as arbiter of “intent.” As accusations fly about each side’s plans for frontier AI, that failure is also prescient.

What stands a chance?

Choices made for the first dialogue will have an outsized impact. The two sides should focus on the most tractable areas first, potentially launching a working group on one or two discrete areas, as they did in 2024 with nuclear command and control issues.

In the two years since the governments’ last AI safety dialogue, a bevy of unofficial dialogues between experts in the two countries have surfaced critical AI safety risks (Concordia AI counts 12 such “Track IIs” between China and the West).

But expectations diverge widely between the AI safety community and former negotiators with the (mis)fortune of eking out agreements and statements with China over the years.

1Wide-eyed scenarios like Plan A pitch a deal with China as the path to containing frontier AI risks. It envisions agreements on compute declaration and an AI training pause by 2029, and offers arms control analogues. But China is forging its own path to compete at the AI frontier. It may see deal proposals like this as a covert way to contain its rise, and is hesitant to offer the transparency and restraint that underpin arms control in traditional areas. Today, China appears to be racing for nuclear parity, rather than negotiate controls from ‘behind’ – and that is the stronger parallel to the AI race.

To understand this divide, I polled a group of experts who participated in official U.S.-China dialogues under Obama, Trump, and Biden, and another group involved in ongoing Track II work on AI Safety. They were asked to rate twelve potential policy topics for a U.S.-China AI safety dialogue on two dimensions, on a scale of 1 to 5. First, feasibility: the chances that a governmental dialogue can lead to meaningful progress on this topic, given the numerous failure modes of U.S.-China dialogue. Second, value: the importance of the topic to mitigate serious risks from advanced AI.

Experts from both camps broadly agreed that cooperation on the issues was valuable. But for the U.S.-China experts, only two topics eked out higher-than-50% feasibility for a productive dialogue: nuclear risk (the subject of an earlier agreement under the Biden administration) and utilizing new models to close vulnerabilities in open source software (which might benefit from a statement of principles, but requires close cooperation). The Track II group was more optimistic on feasibility across nearly every issue, and saw the topics as more urgent.

In terms of combined feasibility and value, a few areas took the lead:

• Moderating risks that today’s AI models, or those that may be available in the near future, could enable chemical, biological, radiological, nuclear, and explosives (CBRNe) threats;

• Discussing biosecurity controls on models;

• Discussing risks specific to non-state actors; and

• Renewing discussions on nuclear risk.

For the first three areas, they are clustered for good reason: they represent threats that can be discussed technically and where the two governments’ interests generally align.

The bilateral discussion can start here, with a working group that begins with low-cost information sharing, but ratchets up into genuine technical and political cooperation.

Expert Panel: What AI safety engagements can withstand U.S.-China tension? Where is cooperation most urgent?

Look beyond government

Next, governments need to think about filling their information gaps by bringing in stakeholders with AI safety expertise.

Both Beijing and Washington are in the early stages developing positions on these challenges and recruiting the experts to work with their frontier AI labs on policy. Companies, technical bodies, scientists, and organizations are doing work in parallel. Governments must find a way to leverage their expertise, while retaining decisionmaking authority.

Involving a range of stakeholders in scoping for talks through public comment would be an early – and easy – win for negotiators. In both countries, key expertise is held outside of policy circles – in frontier labs, in academia, in AI safety organizations and think tanks, among others.

Following the May Trump-Xi meeting, Washington issued a call for public comments on the development of the U.S.-China Board of Trade, including changes to tariffs on non-sensitive products. It should do so again, asking stakeholders directly: What AI safety efforts will be less successful if they do not incorporate the other country’s government? What are early warning signs of risk from the latest open weight models, which are prologue for future risks? What are best practices that industry supports – and may be ripe for coordinated endorsement by the two governments?

The call of the hawk

Of course, for some policy makers, the top AI safety risk is China itself.

One vocal camp observes that the Chinese side is incentivized to give up very little in direct talks. Drawn-out negotiations stall changes to the status quo, which in some domains is almost certainly China’s objective. Both the U.S. and the Chinese sides may well have reasons to want to talk but do nothing. They may want to add a deliverable to an otherwise fractious outcomes document; project an image as a ‘responsible power’; or stall the development of binding instruments.

Moreover, Beijing might define AI governance terms differently from American counterparts. An agreement based on terms the countries define differently carries little weight. When Xi spoke at the World AI Conference in July, he called for AI that is “secure and controllable” and for preventing “malicious use.” As in the cybersecurity domain, shared references to safety, security, controllability, and reliability can differ in their objective – protecting users or protecting indigenous technology development and national security against foreign threats.

Critics of Beijing also point to slow and inadequate enforcement of previous cyber and trade commitments, China’s distillation of American models, open-weight model proliferation, and models aligned with authoritarian values.

Fair points – but they do not support the conclusion that thoughtful cooperation on safety is a lost cause. Some of these hot issues, like export controls, are already discussed in economic channels, and should stay there. On others, like censorship, the two sides will never agree, and Washington must simply hedge by developing countermeasures at home and investing in the international diffusion of the American AI stack. What remains, still, are a range of topics for which cooperation staves off risks that affect both countries.

Do the work

In May, Trump and Xi created an opening for thoughtful engagement on AI safety. Since then, the drumbeat of AI security incidents has only quickened. AI agents escaping their sandbox to hack other companies. Violent extremists using AI to plan attacks.

If the minimum for the September summit is ‘first, do no harm’, they should affirm that catastrophic risks from AI require bilateral talks and understanding, and ideally point to evidence that their teams followed through on the May announcement of government-to-government talks. Leadership signals like these provide top-cover for practical engagement.

Maximalist visions of AI safety cooperation between the two countries (treaties, for one) are unlikely in the near term. But robust information sharing and exchange on issues squarely in both countries’ interests can build a foundation for more ambitious arrangements. In practical terms, that means direct conversations on incidents and risks the governments have observed, and on the lessons they drew from those for risk management. The expert panel suggests a focus on CBRNe risks.

The work is not for government alone. My findings, and the reactions of old China hands, suggest the AI safety community also needs to hone its messages for policymakers weathered by years of fractious dialogue, as well as those animated by the race for AI dominance. Ideally, the governments would actively solicit their knowledge, explain the gaps they need filled, and get closer to the margins of the Track II AI dialogues.

The AI frontier is not slowing. The governments must follow through on this AI safety opening if they hope to keep up – before more catastrophic risks become reality.

In short, OpenAI researchers launched around 1200 agents and targeted them at tasks in ExploitGym, a large benchmark built from real-world software vulnerabilities that was designed to test how well AI models can exploit them. In order to solve “impossible” tasks, these desperate agents, programmed to be highly persistent, found creative ways out of their sandboxes. They managed to communicate with each other, leave each other notes, and eventually reached the open internet. This culminated in them hacking into Hugging Face, a leading platform for open-source machine learning.

Most alarmingly, none of these agents alerted humans to their endeavors or considered their activities to be unethical (if not potentially illegal). In fact, at least a fifth of the agents were interested in tampering with their own transcripts to cover their tracks, according to METR. A few even developed a successful technique for tool call spoofing.

All this raises obvious concerns about how much we can trust AI agents to act safely across our cyber systems. While this case is mostly related to US companies, AI’s cyber risks concern people and organizations around the world. Chinese media coverage and online discussions of this incident have been interesting. Some were quick to frame the situation as yet another case of dangerous American AI losing control, contrasting OpenAI’s risky actions with Hugging Face’s use of a Chinese open model (Z.ai’s GLM-5.2) to patch its security. Others are more cautious, focussing on the threats models like this can pose and how Chinese organizations should respond. Finally, as we inch closer to a Xi-Trump summit at the end of September, a bombshell op-ed from state media over the weekend attempted to set the tone on AI safety.

Today on ChinaTalk, we cover:

• Is state media finally AI safety-pilled?

• How helpful GLM-5.2 was, actually — and where China is on the open-models debate;

• What Chinese researchers are worried about;

• And why Zhongnanhai is doing Anthropic-ology.

We draft and edit ChinaTalk articles without LLMs. In the case of translations, we use LLMs to translate excerpts, then adjust phrasing based on our own judgement. Most translations in this piece were done by Claude Fable 5.1, with the exception of the Yuyuan Tantian piece, which drew from Bill Bishop’s Sinocism translation (assisted by ChatGPT).

How Chinese media covers AI safety

Science and Technology Daily 科技日报, a newspaper published by China’s Ministry of Science and Technology (MOST), published a report on the Hugging Face incident on August 29, which drew heavily from both investigative reports as well as Western media reporting.

OpenAI described the incident as a “warning” to the company and to the world at large: without proper safety safeguards, powerful AI agents are already able to circumvent technical controls, coordinate through unauthorized channels, and take dangerous actions that no human ever instructed them to carry out.

...

For the AI industry, once multiple agents have the ability to operate computers, call on tools, and carry out tasks continuously, traditional security perimeters face new challenges as well.

In this incident, no human directly instructed the AI agents to attack Hugging Face, yet they ultimately carried out the attack. How to detect this kind of behavior in time, and how to stop multiple agents from amplifying risk by collaborating with one another, have become questions that AI safety research can no longer avoid.

First, it’s remarkable that a state-directed outlet is covering this story, which has little to do with China, prominently. The language is strong, but also neutral and technical, with no mention of potential policy measures or governance frameworks. It seems that while institutions like MOST recognize the salience of AI-related cyber threats, they may not yet be top of the pile on decision makers’ desks.

Cybersecurity think tank Anquan Neican 安全内参 took away from the incident that “Chinese models are better than American ones at cyber defense.” On the microblogging site Weibo, state-led channels amplified hashtags like “OpenAI lost control of its model” and “Hugging Face sought help from a Chinese model”, further reinforcing a narrative that Chinese models are the vanguard of safety. This is not necessarily Beijing’s explicit directive. Chinese media knows that nationalism sells and frequently wraps stories in patriotic veneer.

Underneath such narratives, however, China’s actual level of concern for AI’s threat to cybersecurity remains murky. Kyle Chan (of Brookings and High Capacity) recently argued that China will need to see AI safety as a domestic priority before it takes meaningful action, comparing it to the trajectory of climate policy a decade ago. We seem to be in an ambiguous phase right now. Beijing understands that the tides of cyber threats will eventually reach home shores, but isn’t feeling urgent quite yet.

Z.ai’s accidental glory — and what open models mean for safety

When Hugging Face dug into their logs to understand what happened, they found out that frontier models (accessed through APIs hosted commercially) were unhelpful. Uploading extensive details about the attack triggered these models’ security guardrails. Instead, they ran Z.ai’s GLM-5.2, an open model released in June 2026, on their own infrastructure, in order to probe the logs.

Yacine Jernite, head of machine learning at Hugging Face, told CNBC that the company used GLM-5.2 “as a way to analyze the attack, and were able to contain it very quickly using this model.” Dwarkesh Patel reviewed both the OpenAI and the METR/Redwood reports closely and wrote that he “[hasn’t] seen evidence that open source models provided any significant real-time defense.” It seems, then, that at least in terms of defending against the attack while it happened, GLM-5.2 wasn’t involved. The model was mostly used to investigate what happened after the fact.

Many headlines, both American and Chinese, jumped at the opportunity to claim that a Chinese model helped “defend” an American company. Xinhua wrote that Z.ai’s model “saved the day” 救场, quoting Professor Zhang Yue 张悦 of Shandong University:

Zhang Yue thinks that when it comes to AI safety, the truly critical question is whether an increasingly complex and autonomously intelligent system can be adequately understood, verified, and supervised. The significance of China developing open models lies not only with providing another choice to developers. More importantly, it helps to gradually foster a more diverse ecosystem for AI technology.

Openness, broadly speaking, still rules the day in Chinese AI policy’s Overton window. In particular, the transparency, relative controllability, and independence of locally-deployed open models make them valuable for safety work, even as the overall risks of cyber incidents increase due to the proliferation of AI systems. Given the endurance of pro-openness rhetoric in Chinese reporting, we should not expect major U-turns any time soon barring sudden incidents.

Chinese researchers on the future of cyber

In a separate piece covering the incident, Science and Technology Daily interviewed cybersecurity researcher Huang Wenhong 黄文鸿. Huang works at the China Center for Information Industry Development (CCIID), a research institute affiliated with the Ministry of Industry and Information Technology (MIIT). He argues that the future of cybersecurity has AI on all sides of the coin:

AI agents can probe in parallel around the clock, rapidly switch paths, automatically chain vulnerabilities together, and compress the attack cycle with tens of thousands of low-cost operations. “The direction for the next step in building security systems is clear,” Huang Wenhong advised: AI must be used to fight AI. This means adding machine-speed behavioral detection, full-trajectory monitoring, automatic isolation and circuit breakers, and resilient recovery, while treating every agent as a potential insider threat with restricted privileges.

Zhou Hongwei 周鸿祎 is a co-founder of Qihoo 360, one of the most prominent cybersecurity companies in China and the developer of 360 Safeguard. In July, he shared two commentaries about the OpenAI-Hugging Face incident on social media. Unlike those who focused on comparing China with the US or open with closed models, Zhou paid attention to the granular cybersecurity challenges this incident revealed. He is most concerned about the level of persistence OpenAI’s agents demonstrated:

An agent doesn’t need to be brilliant at every step; it only needs to never get tired. A human hacker who fails several hundred times in a row might grow weary, give up, or decide the cost is too high, but an agent won’t. It can keep trying, read the error messages, adjust its strategy, and try again. The first ten thousand attempts can all fail; as long as the last one succeeds, the system may be breached. So what AI has truly changed is not just attack capability but attack cost. …

Faced with agents that never tire and can keep trying indefinitely, we must draw hard boundaries in advance around runtime, compute, permissions, and network access. The moment abnormal behavior appears, the circuit breaker must trip immediately. At the same time, disclosure of AI security incidents must become more rigorous. What is officially confirmed fact, what is merely speculation, and what still lacks evidence must all be spelled out clearly. Because today it is far too easy to use AI to generate a technical analysis that is packed with jargon and looks extremely professional, but reading like the real thing doesn’t mean it is.

To him, Hugging Face’s use of GLM-5.2 doesn’t necessarily vindicate Chinese models over Western ones. Rather, it demonstrates the continued relevance of local deployments for enterprises, at a time when companies, in a rush to adopt AI, are sometimes lured into a false sense of safety by cloud inference providers. In another video, he echoes Huang Wenhong’s sentiments earlier:

The biggest lesson this incident holds for China’s AI industry is not simply that domestic models beat American ones. What really matters is that in critical settings such as security, finance, and healthcare, companies must have models they can deploy locally and control independently. When something truly goes wrong, sensitive data cannot just be uploaded anywhere, and the model’s permissions and rules cannot be dictated entirely by someone else. At the same time, agent security cannot rest solely on having a huge model; it has to be built into the entire system: least privilege, network isolation, confirmation of critical operations, end-to-end monitoring, and immediate suspension and disconnection the moment an anomaly is detected. In the future, companies will need not only an agent that does the work for you, but also a dedicated security agent whose job is to keep watch over it.

Beijing’s Anthropic treatise

China’s leaders like to keep their options open. In a policy realm as rapidly-evolving as AI, they prefer not being locked into a corner. This applies to narratives as well as policy options: Beijing wants to define words under its own worldview, rather than accepting semantic frameworks developed elsewhere. AI safety is one such word.

OpenAI isn’t the only company whose models-in-training have breached sandboxes. Back in April this year, Anthropic disclosed in its system card that Claude Mythos Preview also broke out of its testing sandbox and reached the internet. (For what it’s worth, Mythos at least emailed one of Anthropic’s researchers, whereas OpenAI’s agents conspired to keep humans in the dark.) The incident motivated Project Glasswing and Anthropic’s decision not to release Mythos to the public.

On August 30th, a prominent Beijing mouthpiece published a remarkable treatise on Anthropic and America’s AI politics writ large. Yuyuan Tantian 玉渊谭天 is a commentary channel under China Central Television (CCTV) that focuses on international affairs, particularly US-China relations. (Its name is a reference to Yuyuantan 玉渊潭, a large park across the street from the CCTV building in Beijing.) It wields a distinctly stylized voice and is regarded as an authoritative propagator of Beijing’s views.

The piece argues that Anthropic, founded upon Dario Amodei’s effective altruism-inflected savior complex, has now become an effective tool of Washington’s AI arms race. In its quest to achieve superintelligence and put it in the hands of people it deems morally superior, the company repeatedly sidestepped boundaries of corporate behavior. Drawing from primary sources like “Machines of Loving Grace” and The Technological Republic, the piece posits that Amodei, Palantir’s Alex Karp, and the new tech elite are using a combination of AGI belief and supposedly pro-democracy ideology to justify dictating humanity’s future from Silicon Valley. Its author diagnoses this particular mélange of ideas as “American disease” 美国病:

This is entirely a monster bred by the American path. It genuinely believes it is right; genuinely believes its model is a “democratic model” and its safety standards are the world’s safety standards; yet it is completely unaware that this very absence of self-doubt is the greatest danger of all.

Near the end of the piece, the author redefines AI safety in even more explicit terms:

First, there must be a distinction between what constitutes a genuine safety threat and what is merely technological competition.

The United States has directly defined Chinese models that are lower-priced and more open as “safety threats.” This has already departed from technological competition and slid toward blockade.

Anthropic’s differential treatment of users in different regions, packaging commercial competition as a safety issue, is one manifestation of this. The problem is that if the power to define the rules lies in the hands of an American company, its competitors will forever be placed on the “unsafe” side of the line. What is needed now, therefore, is a scientific definition that both sides can accept, rather than unilateral declarations by a particular country or company.

(A note on translation: Sinocism’s version translated 安全 as “security” in the first, second, and third paragraph of the block quote above. I think “safety” is the more accurate translation here, because it contrasts the “unsafe” framing in the third paragraph and is intended to weigh in on Western AI safety discourse. Comment below if you have thoughts!)

Western AI safety advocates often see their field as politically neutral territory between the US and China. Both countries, after all, would face severe harm if risks in cybersecurity, biosecurity, and even existential threats proliferate, so it is in their interest to collaborate on the problem of AI. Beijing, however, perceives the AI safety discourse as overly captured by forces like Anthropic, which in turn serves the interests of Washington. It argues here that China will not accept a definition of AI safety wholly borrowed from Western thinkers; in fact, in its eyes, those thinkers are implicated in “American disease.”

We should expect Beijing to articulate its own definition of AI safety in the coming months and years. China’s ambitions in the realm of ideas should not be underestimated.

spent his formative years as a DARPA and NSA contractor before four years at Meta, where he founded the company’s frontier cyber capabilities evals team and finished as its AI security tech lead. He recently left to co-found Abundant Security.

We discuss…

• What actually happened in the Hugging Face hack, and why AI safety insiders weren’t surprised,

• The “grad student lab” security culture inside frontier labs racing sixty hours a week to ship,

• Why you should apply to run a new org with eight figures to play with to follow trends in AI + cyber risk (apply here!)

• Why the cyber Pearl Harbor predicted since the early 2000s never arrived — and whether swarms of hacking agents finally change that,

• How nation states will post-train open-weight models like GLM and Kimi into billion-dollar cyber weapons,

• Why AI has so far favored defenders, from mass bug-squashing to superhuman network monitoring,

• The industrial economics of cybercrime — kill chains, ransomware divisions of labor, and whether AI gives criminals 100x returns.

Listen now on your favorite podcast app.

What Happened at Hugging Face

Jordan Schneider: What was the Hugging Face hack?

Joshua Saxe: OpenAI was training a new model capable of solving long-horizon tasks — tasks that take many hours to complete. The way that kind of training works is you train thousands of copies of a model in parallel. You have them all go out and try to solve a set of tasks, you see which ones succeeded and which failed, and then you update the model’s parameters so it does more of the good behaviors that led it to solve the task and less of the bad ones. You keep doing that iteratively at very large scale over the course of weeks, sometimes months.

They were doing that kind of training, and while the models were being trained, they broke out of their training environment, hacked internal infrastructure at OpenAI, and then eventually hacked another company — Hugging Face, an open-source model and evaluation repository that’s widely used in the ML community. At a certain point, Hugging Face discovered what had happened, and eventually everybody figured out that these models had broken out. But this whole thing unfolded over the course of many weeks. There were three different cases — a couple of early cases where the models broke out of their containment environment just at OpenAI, which was detected, and OpenAI thought they’d fixed the problem. Turns out they didn’t, and in the third case the models actually hacked Hugging Face.

Jordan Schneider: The world is currently freaking out. Are they correct, dude?

Joshua Saxe: For those of us who’ve been around AI safety for a long time, it wasn’t that surprising. I didn’t find it that surprising. But it is alarming — alarming that it happened because we could have predicted this. Not just OpenAI but also Anthropic, Meta, a testing company called Irregular, and the UK AI Safety Institute all had incidents like this happen in the last year or so. We as an industry knew this could happen, and the basic measures that should have been taken to prevent it weren’t taken. That’s troubling. We need to fix that.

Jordan Schneider: How easy is this as a thing to stop? And what is the incentive structure that’s wrong in all of these organizations — including ostensibly very safety-pilled ones — for them to not have been able to make sure their latest models didn’t run amok in customers’ servers?

Grad Student Energy, Serious Stakes

Joshua Saxe: I’ll say a couple of things. I was at Meta — I started the team that did frontier cyber capabilities evals. I’m giving you a sense of where I’m coming from, because I can’t speak directly to exactly what happened inside the labs — that’s private. But I had the experience of starting a team that did similar experiments to the ones that were happening when the OpenAI models broke out. And we all kind of know each other in this community, across the labs.

With that said — the culture among the training teams and the evals teams at the labs has a Wild West feeling to the whole thing. Everybody’s under a ton of pressure to move really quickly. There’s enormous time pressure to release new models, tremendous awareness of how any given lab is doing relative to the others in terms of capability. And the field has evolved super fast in the four years since ChatGPT launched in fall 2022. The whole thing has been kind of a blur — anybody who’s worked in the labs in that time has been working sixty-hour weeks the whole time. So it’s not surprising to me that you could have a kind of grad-student computer-science-lab culture around the security of these model training runs, which I think is what’s the case, and I think that’s why you’ve seen these lab escapes. It is ironic, because frontier AI people talk about safety all the time — but that irony is a real thing. The security practices are not very good, because of all those dynamics.

Jordan Schneider: So is your sense that if adults take this more seriously and take a deep breath before they press run, then runaway AI taking over servers around the world is not something we should be super scared of — once people get the fear of God into them, which hopefully this incident does?

Joshua Saxe: For AI’s level of capabilities today — or maybe three months ago — it is true that implementing basic, known security practices would have stopped these incidents. Those practices are things like better sandboxing. The models were sandboxed — for your audience, sandboxing means we constrain the models to a very circumscribed set of behaviors and don’t let them do things like talk to random websites on the internet; they can only execute certain commands. The models were sandboxed, but not very well — not well enough to prevent the escapes. And it seems nobody was monitoring the models, when it was standard practice to have a team of humans monitoring for misbehavior. With current-level capabilities, that would have sufficed to stop these incidents. There were a lot of really bad choices made around security — not just at OpenAI, but across all the labs that had these escape incidents.

Now, I also think we should keep in mind this is going to get more challenging if the current rate of progress continues, and we’re going to have to be more creative and innovative. It’s hard to secure these large training runs because definitionally they operate at huge scale. To get the models to solve, say, coding problems that take multiple days, you need thousands or tens of thousands of models all trying to solve problems in parallel. The way we’re making AI better is by scaling it — so these training runs need to get bigger, go longer, run more in parallel, and solve more complicated tasks. And there’s pressure for the tasks to be more and more real-world-like. Part of the reason the OpenAI models escaped is that they were given the ability to download packages off the internet through an Artifactory capability — which is what models do in the real world. If you want to train models to solve real-world problems, there’s a trade-off between how much access you give them and how safe the training is. As we scale up training and the models get more capable, there are going to be more and more safety challenges, and there needs to be more and more safety investment. What’s true today about how we keep them safe won’t be true in a year.

Jordan Schneider: Let’s make the AI cyber observatory pitch.

The Case for a Cyber Observatory

Joshua Saxe: It feels like there’s looming risk here, but we don’t know exactly how it’s going to materialize — and it’s basically impossible to know, due to how complicated the interaction is between technology and society. So probably more important than preparing for any one possible future is building state capacity to understand the specifics of the risks that are actually physically manifesting, and what their trajectory is. We don’t have that state capacity right now. There’s no organ of the American government — or, as far as I know, any friendly government — that’s tracking all of the risk signals around how adversaries are using AI and how we should prepare. We need an institution that actually does that and can make good recommendations to policymakers.

Jordan Schneider: Okay, what’s our budget? We’re looking for donors — come on, let’s do the pitch.

Joshua Saxe: There are donors in the AI donor community who would be happy to fund this thing, so we’re not bottlenecked by money. We are bottlenecked by talent. We’re looking for someone who has a background in AI, a background in cybersecurity, and who likes building statistical models and forecasting models where the objects we’re trying to predict aren’t directly observable. For example — how are ransomware gangs using AI? We can’t directly observe this, but we need statistical models to help figure it out. Someone who likes doing that kind of work, and who can also make their analysis legible to policymakers and the public. If we could find such a person, we would have plenty of money to found an organization dedicated to tracking the AI cybersecurity risk landscape, tracking how defenders are using AI, and making much better recommendations to policymakers about how to address risk in this space.

Jordan Schneider: Let me play a bit of devil’s advocate. I understand that someone is going to need to do this work and come up with the recommendations. But you just made this point — we’re not ready for this, the world is not prepared. Isn’t it going to take crazy hacks, or effects in the world, or a movie? Is it really going to be more reports that get Mount Sinai to up their game?

Joshua Saxe: So do we need more wonkish policy analysis here, or is that basically just feel-good stuff?

Jordan Schneider: I mean — I fucking host ChinaTalk. Of course we’re in favor of wonkish policy analysis. But the question is whether what we need now is a broader societal reckoning with AI and cyber risk, which is not going to happen through policy reports — or whether we just need the playbook, and once we have the playbook, we can execute on it.

Joshua Saxe: I’ll give you an example of where this could possibly have helped. As you know, the US government blocked Anthropic’s Mythos model from launching for a while, and they also blocked OpenAI’s GPT-5.6 for a little while too, on the grounds that these models would pose too much cybersecurity risk. Their argument was that when you measured these models’ cyber capabilities and did experiments around finding bugs, the models exceeded a capability threshold that made them unacceptable to release.

I think the way that decision should have been made is by looking at a bunch of signals about how defenders are actually using current AI vulnerability-finding capabilities, how attackers and defenders are actually using them in the wild, and how vulnerable our critical infrastructure really is. If they had done that, they would have seen that defenders are having enormous success with AI vulnerability finding — we’ve been fixing tens of thousands of vulnerabilities — while attackers are mostly relying on social engineering and exploiting known bugs, not on these models’ ability to find new zero days. It would have been a net positive to release those models faster. So having an organization that brings together all these signals and makes informed recommendations is important.

And even if it takes a lot of pain to get us to a place where we start up-leveling our cyber defenses, you still want an observatory organization — in the same way that during COVID it was useful to have heat maps of where the virus was, plus analysis of what the effect of vaccination and masking was. You want data, so you can have an OODA loop around how to deal with a crisis. We won’t have a very good OODA loop if it’s just people going off vibes.

Jordan Schneider: Okay. So we can have our heat maps, we can also have a TikTok clipper, and both of those together are really going to get us where we need to be.

Joshua Saxe: That’s the hope, yeah.

Jordan Schneider: So if you want to run Joshua’s AI cyber observatory, what should you do next?

Joshua Saxe: Just email me at josh@abundantsecurity.co and/or fill out this form we’d love to talk to you.

Is It Time for an AI Pause?

Jordan Schneider: Let’s break this down into three components of our AI apocalypse: the AIs running amok on their own, then nation states, then cybercrime. Starting with the AIs running amok — you seem most concerned about new training runs, as opposed to models already trained and shipped into the wild that start feeling themselves and make kamikaze runs in support of future AI overlords. So is it time for an AI pause, Joshua? The first-order incentive as of April 2026 — “this other lab is going to steal a march on us” — seems to be beating the second-order one, which is “if we have too many of these mistakes, our entire industry gets shut down.” We still have capitalism running in the background of all this, right?

Joshua Saxe: Is it time for a pause? In some utopian world where people can coordinate way better than they can now, maybe we would pause — do a ton of safety research, implement all the safety stuff, and then press play again. That sounds great. If you can get all the Chinese labs to do that, and all the American labs, and all that. I don’t think it’s going to happen. So I think we need to come up with a realistic policy.

Jordan Schneider: But if we want these models to keep getting smarter — even if the humans watching the training runs aren’t grad students and are actually doing their best — the creepiest part of this whole thing, reading the reports, is that not one model tattled. Not one told the humans, “hey, maybe this isn’t the coolest thing we should be doing right now.”

Joshua Saxe: Is this concerning? Yes, this is concerning. As the models get more autonomous, as we turn over more and more autonomy to them, we need safety to keep up, and safety gets more and more challenging. So how do we do that? I think there are going to be pressures due to capitalism on the labs to do that, so the models are safe enough to be sold. Nobody wants to deploy a coding agent in their infrastructure that 0.1% of the time goes off and hacks some other company. So there are going to be pressures from microeconomic self-interest that push the labs to achieve a certain level of safety. But then there’s going to be some residual level of safety the labs won’t be endogenously incentivized to implement, and for that we probably need some kind of regulation.

Jordan Schneider: But even if the incentives are there, can just doing the smart security things you were disappointed these labs didn’t do get you to safe training runs as capabilities keep expanding? Or is there something where now these models are so smart that the further we push them, the less we’ll be able to use our old playbook to keep them in a box?

Joshua Saxe: First, a couple of things. One, it’s not just security — there are a bunch of things that have to happen to make the models safe. The model weights themselves — the neurons and their interconnections — need to be aligned. That’s how we talk about it in the industry right now. They need to have the kind of deontology that humans have, which would include not hacking random companies in order to solve the evals. There are people working on that. Then we also need to figure out how to monitor the models when we deploy them.

I’m going to make a larger point here, which is that there’s an enormous research overhang in all of the areas I’m about to list — researchers have done a bunch of work, but the labs haven’t implemented it. Aligning the model weights, that’s one area — there’s lots of literature, and it hasn’t all been implemented. Then there’s monitoring the models — there’s been a bunch of work on monitoring the internals of the models while they’re running to see if they’ve become misaligned. For people in your audience who read the ML literature, Owain Evans is super well known in this space — a bunch of work around understanding when the models are scheming or lying, and the conditions under which they do that. Then there’s work on monitoring the chain of thought of the model, which OpenAI famously wasn’t doing when the models broke out of the sandbox — this is the English-language internal monologue of the model. Then there’s work around monitoring the tools the model is calling. Then there’s sandboxing. There are so many layers of safety, and in the safety community and the cybersecurity community we’ve done a fair bit of work and built knowledge around how to implement safety at each layer.

But the labs need to absorb this research overhang. That’s one thing that has to happen. And they’re so incentivized to race right now. At least this has been my lived experience working in one of these companies — without talking about anything private, it can be quite hard to get safety stuff into the models as a safety person in one of these labs. But part of the point I’m making is we don’t need a Los Alamos kind of safety effort right now. We might eventually. We just need to get the labs to do the things we already know how to do. Eventually there’ll be new research problems — it won’t be long before we need a lot of new research — but right now the priority is just implementing what we know.

Jordan Schneider: All right, we’ll take your cautious optimism for at least the next six months, and perhaps check in then.

The Cyber Pearl Harbor That Wasn’t

Jordan Schneider: Now I want to turn to nation states. Even without rogue AI, enormous new hacking capabilities are coming online — already extant in models like Mythos and 5.6, with Chinese models creeping toward that horizon — plus the prospect of whatever hacked Hugging Face being something a country can deploy to go do its bidding. But let’s look backwards before we look forwards. There’s a general consensus in the field that the cyber Pearl Harbor predicted in the early 2000s — cyber as a domain with real strategic impact on nation-state competition — hasn’t panned out. What we’ve seen over the past 25, 30 years is tactical stuff and better spying, not wars won and lost by hacking. What were the binding constraints?

Joshua Saxe: Are you asking why we haven’t seen strategic-level cyber-physical effects achieved almost ever? I agree with you. Maybe the high-water mark in terms of strategic effects was Stuxnet, the US-Israeli attack on the Iranian nuclear program, which I think disabled a couple hundred centrifuges. That might have been the closest we’ve come, and the overall context is just disappointment from national militaries in trying to achieve these cyber-physical effects. So having grown up in the national security space at a time when we were talking about cyber Pearl Harbor — this has been a disappointment from the point of view of offensive capabilities.

There’s no one clean answer as to why. But maybe we also have to ask why we should have expected it. Why did we ever assume you could achieve kinetic-warfare-level cyber-physical effects using cyber weapons? I’m not sure why we ever assumed that. Take the attack Russia did on the Ukrainian power grid right after the invasion — they managed to disable power for a couple hundred thousand people for four hours, and then the grid operators switched over to manual operation and turned the grid back on. I think this has to do with how our society is designed. Physical systems are controlled by internet-connected systems, but there often is a manual option.

For me the interesting question is: now that a single cyber operator can outsource their work to N number of agents working in parallel, could you achieve strategic effects now? What if you scaled that up a hundred times, and combined it with an actual physical military operation? Maybe cyber warfare becomes, for the first time ever, after twenty years of talk, an actual real thing. That seems possible and worth considering, although I don’t have a very strong position on what’s going to happen.

Jordan Schneider: The Russia-Ukraine war is the best natural experiment we’ve had, and basically the answer has been “close but no cigar.” You can’t do stuff remotely at the level of impact Russia has had actually bombing power plants, as opposed to trying to cyber-hack them.

Joshua Saxe: First of all, the capabilities are changing so quickly. Two years ago, when we started doing cyber capabilities testing on models across all the labs, it was a joke — they could do maybe half of the first step of a kill chain. Now they can do the full thing. They can’t do it in a very stealthy way, but maybe a year from now they will. If I were at the NSA, I would be interested in post-training GLM 5.3 to do exactly the kind of operations we wanted to do. So probably national militaries will eventually do that. It’s a little hard to know the shape of these weapons going forward. I also think the command structures are going to need to evolve, because right now, in a lot of settings, operations are bottlenecked by some human decision-maker deciding whether to engage a target in the cyber domain. The whole value of swarms of agents is that you don’t bottleneck their operations — so there needs to be some thinking about how the decision-making works. It doesn’t seem like the right time to make strong arguments. This is a time when there’s going to be experimentation. It’s too early to say exactly what the nature of this capability is, and it’s going to keep changing.

A New Category of Weapon

Jordan Schneider: Given that we don’t know what the correct conceptual frame is, what are the arguments for and against more capable intelligence communities around the world causing a cyber apocalypse?

Joshua Saxe: I’m going to reframe this slightly, because I think the tier-one cyber programs will be able to use AI capabilities in all sorts of ways. One is cyber weapons development — the most well-known thing going on right now. You can basically pay for bugs with tokens. You point coding agents at software and they find exploitable bugs. This is interesting for all militaries that have cyber programs. From my DC and government-contractor days — and I think this is common knowledge — there’s an archipelago of small contracting shops that find bugs for the NSA and the military. Those economies are going to change a lot. There’s also an economy in all the cyber powers around building malware — post-exploitation, where you install software implants on victim computers. All that stuff can be vibe-coded now. The cyber weapons production world is changing really quickly and can be made much more efficient.

Then there are the hands-on-keyboard cyber operators actually running the operations, and those folks can now scale. It’s an open question for me how much this helps the tier-one programs, where there’s a lot of talent, and where it’s typically small teams that do these pretty large operations. It’s hard to know exactly how many people were involved in something like Operation Aurora — where Chinese hackers broke into a bunch of US infrastructure, most famously Google — but it was not that many people, maybe dozens. So it’s not clear to me how bottlenecked those operations are by labor. If they’re very bottlenecked by labor, agents matter more. If not, computer-network-operations agents matter less.

I’d argue the second- and third-tier cyber programs might benefit more from all this — where they don’t have capable operators or access to the same vulnerabilities, and now, instead of training people over a decade, they can just have their agent do it. And possibly non-state actors, who don’t have cyber programs at all. I don’t know if Hezbollah has a cyber program, but maybe now they can have one that’s amazing, thanks to agents. There are many such cases — probably scores of actors who might want cyber capabilities can now get them. It’s social complexity and very hard to predict. But we should prepare ourselves for a real disruption in the way cyber conflict happens.

Jordan Schneider: I’m getting real whiplash here, because you almost talked me into capitalism keeping this under wraps. But if we’re talking about models post-trained to hack things — what is the AI safety suite of the MSS going to be? What’s Hezbollah’s? These agents are going to be hooked up to compute somewhere; it’s not like they need to break into a neocloud and steal everyone’s credit card to get going. And we’ve already had worms that aren’t AI-superpowered take off, overshoot what they were intended to do, and shut off corners of the economy.

Joshua Saxe: For the purpose of this exchange, I’d separate the safety into two threats. One threat is that agents deployed by benign actors doing benign things go rogue and do malicious things — that’s what happened in the OpenAI-Hugging Face hack. The other is actors seeking to do harm to some third party using models to do that. That’s the second thing we’re talking about.

Jordan Schneider: My mental model is that even if these models get more powerful and a lot cheaper to run, the US and China still have enormous incentive not to shut off each other’s power plants and hospitals and water treatment facilities — the chances of getting found out are non-zero, and the costs of having done it are very high. As long as there’s enough command and control, and the price of entry is a nation state or someone with nine figures to burn, some rational calculus is going on. We end up in a world where secrets are harder to keep, but no cyber apocalypse. Let’s stick with countries, and then get to Hezbollah and the criminals.

Joshua Saxe: Actually, let me push back on this a little. What you just posed feels a bit strawman-ish, because yeah — the US and China aren’t just going to start destroying each other’s critical infrastructure in peacetime. But in wartime we use weapons against each other, right? And this is a new category of weapon, and it could disrupt warfare. I think that second thing is the more interesting proposition.

Jordan Schneider: Fair. But let’s do the accident case first. Say I’m an intelligence agency that post-trained a model to hack, I’ve made it extra devious, and I’m not running the full safety suite. It’s peacetime, we just want to be spying on stuff — but six months from now the models are better, they run ahead on their own, do more than intended, and all of a sudden global shipping falls apart. Is that a real risk?

Joshua Saxe: So you’re worried about a non-intentional, terrible effect of these models?

Jordan Schneider: Yeah — even in the world we posited in the first twenty minutes, where the labs figure out how to train and deploy long-acting AI safely. Serious organizations running serious espionage operations, or laying groundwork for a World War Three that isn’t imminent, spilling over into a cyber apocalypse by accident.

Joshua Saxe: It’s possible. I’m personally way less worried about that than about intentional use of the models to cause harm. There are a lot of people trying to do harm in peacetime with these models — cyber criminals, ransomware gangs. And there are a lot of actors who’d be doing harm pretty quickly. I don’t want to pick on Hezbollah, but I’m thinking non-state actors —

Jordan Schneider: They’ve had a rough few years. You gotta put some respect on their name, Josh.

Joshua Saxe: No, no — there are active conflicts, right, around Iran/US/Israel, or Russia/Ukraine, where people could immediately start using these capabilities. And then there’s the prospect of a buildup, where the US and China start tooling up, and in the event of a conflict you could see these weapons deployed. In many ways these are just a new kind of weapon, like drones. But it complicates geopolitics. It’s a significant new capability. I object to the framing of “apocalypse, yes or no.” I think there’s more of a continuum here, and the more interesting question is: how disruptive is this new capability going to be?

Winning a War by Pressing a Button

Jordan Schneider: How disruptive is this new capability going to be?

Joshua Saxe: In terms of geopolitics, it’s possible that scaled hacking with cyber agents could be qualitatively more important to warfare than it has been in the past. As soon as you unbottleneck cyber operations — so you don’t need a human for each one, you can have a single human managing a large fleet of hacking agents — it seems possible that national militaries and non-state actors could finally realize the dream of having real, significant physical effects on their opponents. This is also interesting from a miscalculation point of view. If someone went to sleep eighteen months ago having internalized a perfect model of AI’s impact on cyber warfare, and woke up today and was doing policy, their policy would be completely wrong — because the capabilities have gone from basically a joke to now superhuman in their ability to find bugs. So this is something to pay attention to. It’s going to be an important element of geopolitics now.

Jordan Schneider: The idea that someone eighteen months from now can believe they can win a war by pressing a button — even if they can’t — you’ll perhaps be able to talk yourself into that future, which is a dangerous thing.

Joshua Saxe: Or just have game-changing strategic effects. I don’t know exactly what the expectations were around Stuxnet, but the expectations and what actually happened diverged. Stuxnet was back in the days when you needed the most elite, talented people, and there were very few of them, to do an operation like that. Now that’s scaled up. You can imagine a big government program post-training open-weight models — Kimi K3 or GLM 5.3 today, but they’ll keep getting better — investing a hundred million or a billion dollars into creating this capability, expecting to unleash it in some strategic surprise. And then who knows what happens, because this is untested.

Jordan Schneider: The political economy is always to say your thing is really cool. If you’re selling it into the government, if you’re the major who wants to get promoted, if you want to up your budget — who knows how much compute is getting allocated to all these places? Everyone always wants more GPUs. And once you get the GPUs, you gotta show you’re doing something with them.

Joshua Saxe: Exactly. If I were to quit the private sector and go be a DARPA program manager, I would ask for two billion dollars to post-train models to do computer network operations. And then I would sell that really hard. It’s amazing, as you’re saying.

Jordan Schneider: I keep recalibrating how scared I am.

Joshua Saxe: What’s the time series? Are you less alarmed, more alarmed?

Jordan Schneider: I think I’m ticking up a little bit. But we haven’t talked about the defense side at all, and that’s worth going into. There’s some hope out there that we can also just press a button and then all code will be fixed forever.

Why Defense Might Win

Joshua Saxe: I’m not optimistic about that — it’s obviously a caricature, and it can’t happen very quickly. But AI can be extremely useful for defense. My sense is it’s been more useful for defense thus far than for offense.

What does it afford? It affords running all the AI vulnerability-finding tools against your code before you launch it, which is a huge strategic advantage the vendors have. Yes, attackers can spend lots of tokens to find bugs in your code — but if you can run all those tools first, you can really harden your code and make it way harder for adversaries to find exploitable bugs. In the last year, AI has been miraculously effective at this. There was a blog post from a team at Google a couple of months ago, bragging in a delightful way about how they found and fixed thousands and thousands of bugs in, I think, Google Chrome — and they seem to have done that before attackers could exploit them.

AI can also be used for surveillance of all the activity in a network to detect intrusions and stop them before they get far. Defensive labor has been bottlenecked by our ability to monitor networks at scale. There’s been twenty-five years of work building detection funnels where we have detectors for suspicious things on a network, but then a human has to follow up and figure out if a breach is really happening. Even huge hyperscaler companies will only have dozens or scores of people monitoring networks that are the substrate in which all of humanity mediates its communications. So it’s a game changer — the idea that you could now have tens of thousands of AIs that seem superintelligent at detecting breaches operating in people’s networks.

There are many cases where AI can be tremendously beneficial to defenders. Even in the limit — the platonic ideal of how defenders could deploy this — AI is defense-dominant. Defenders can ultimately get to a much better situation with AI. Obviously the real world is different; how we roll this out is bottlenecked and high-friction, and that’s part of the problem.

Jordan Schneider: My brain keeps coming back to tricking yourself into thinking you have a strategic effect you can just press play on — taking a weapon system offline, or the command and control of it, or an adversary military’s ability to talk to each other. Maybe there’s a manual mode for your Ukrainian power plant, where some 65-year-old guy remembers how it works and can flip a switch. But we don’t have Cold War-era radio backpacks hanging around anymore, and most of modern warfare has no hardwired backup you can flip on. Maybe you could manifest one over a few weeks or months — but it’s not crazy to imagine strategic effects being possible just with code.

Joshua Saxe: No, I don’t think it is. One way to convince yourself: if Russia was able to disable part of the Ukrainian power grid and turn off power for four hours for a couple hundred thousand people — what if they did that in a hundred places all at once, while they were invading? That’s already interesting strategically. And the models are better now at finding zero days than humans; they can operate 24/7. There are all sorts of ways in which they’re superhuman. I’d put substantial probability mass on a future in which you could achieve strategic effects using these weapons.

Jordan Schneider: And it’s not just achieving the strategic effect — it’s someone being able to stand in front of a national leader with a straight face and say “I can do this.” Whether or not it works, what we don’t want is for this to lower the threshold for starting new wars. Someone somewhere is going to make that pitch to a leader looking for an excuse — Putin in late 2021, looking for a reason to believe he could win. This just gives a person predisposed to that another reason.

Joshua Saxe: I think this is reasonable — it makes sense. I’m thinking back to Stuxnet — your audience should look up the details, because I don’t remember the exact blow-by-blow, but it was a capital-intensive operation. My memory is that Israel and the US set up actual centrifuges in the Negev desert and tested the cyber weapons on them as part of their development. They also burned a bunch of new zero-days. It was an amazing technical tour de force in cyber weapon development. It’s not hard to imagine a similarly capital-intensive effort to set up cyber ranges with simulated or real critical infrastructure, and spend a few billion dollars post-training some open-weight models to get really good at attacking it. And then the claim would be that you have this magical new cyber weapon.

The Cybercrime Industry

Jordan Schneider: Scaling down from China or the US spending a few billion dollars, to your neighborhood gang or the DPRK — what’s the right way to conceptualize this? First, let’s level-set on what cybercrime has been able to accomplish over the past thirty years, and then look forward to what AI can do for this cohort.

Joshua Saxe: Cybercrime has become a real industry — or set of industries. It costs somewhere between five hundred billion and a trillion dollars a year in damages. Then there’s also the cost of paying for defenses. It’s a significant drag on the world economy — maybe 0.5 to 1 percent of global GDP. I’m getting this from a Gov.ai study from about a year ago that estimated 500 billion in damages, though it was a big range; a McAfee study estimated a trillion. Behind that is a somewhat complex division of labor. There are groups that build the malware used in ransomware attacks, then groups that just specialize in deploying it — they purchase the ransomware and specialize in operating it. And within that there’s further division: groups that specialize in getting initial access, a foothold into a victim network, and other groups that, once they have that initial access they’ve purchased from the access brokers, specialize in compromising the whole network and ransoming it. By kill chain I mean the sequence of operations an attacker group runs before achieving its goal on a victim network. Some of these ransomware groups use the same kill chains as nation states, and have capabilities that approximate nation-state level.

Jordan Schneider: So it’s just that best practices have bled down from the NSA to your Kazakhs.

Joshua Saxe: That’s right. The most advanced ransomware groups look a lot like nation-state actors in terms of their level of talent, their arsenal of cyber weapons, the malware they use, sometimes the exploits. They can be pretty advanced. It used to be that these attacks were pretty simple; now they can be complex heists that involve months of breaking into a network, moving quietly, compromising all their data, encrypting their hard drives, and eventually asking for a ransom.

Jordan Schneider: So think about the Mexican cartel with guns as large and sophisticated as the army chasing them. There are still militaries with F-35s, and no gang has an F-35 — but in cyber that’s not necessarily true. What’s the analogy? I guess the good guys have more compute and more people. What can defenders bring to the table?

Joshua Saxe: There’s the comparison between ransomware gangs and national militaries, and then between ransomware gangs and defenders. In terms of national militaries, it depends who you’re comparing to. The US cyber program has more than $10 billion invested in cyber war, so there’s no comparison in level of investment. But the practices often look similar in how you break in and move around. Ransomware versus defenders — the defensive landscape is very power-law. There are regulated financials with hundreds of people working on cybersecurity, like Blackstone, or hyperscalers like Google or Meta. But then there’s your regional network of hospitals, where sometimes it’s an IT person plus another IT person who maybe learned some security. It’s very uneven, and a lot of the time the people in the long tail are the ones who get victimized.

Jordan Schneider: Who get picked on, sure. So cybercrime has been an annoying wart on the global economy rather than a cyber Pearl Harbor — acute impacts over short periods, but never a real blocker on global growth or a catastrophe for large swaths of the population. What’s the CAGR for global cybercrime, and how worried are you that a Russian gang post-training GLM gets 100x returns?

Joshua Saxe: I’ve tried to get a good CAGR estimate for cybercrime over the last ten years. This is a murky domain with a lot of uncertainty, but I think 10% would be my estimate for how fast it’s grown — pretty fast, as scams, fraud, and ransomware have gotten industrialized. So it’s already worrisome. I agree it’s a wart on the world economy, not a catastrophe. But it’s a real problem — not just the direct damages, but how annoying it is to operate a lot of businesses because of all the security friction, which is a downstream result of this.

In terms of what happens next — it’s reasonable to assume that over the next few years cybercrime gets a lot worse. A simple mental model: software engineers have gotten maybe three times more productive thanks to coding agents. Maybe cyber criminals get three times more productive at breaking into networks. That’s a rough mental model. At the very least, we should be prepared. Policymakers should think about the possibility that in a third of possible universes nothing changes, in a third it’s incrementally worse, and in a third, attackers suddenly have armies of agents that dramatically scale up their businesses. We should be worried about it regardless of the exact probabilities.

Bugs on a Bedroom Rig

Jordan Schneider: Is compute a constraint here? Is the scale you’d need obtainable for your friendly neighborhood hacker — do you rent from a neocloud, or buy your own racks? Is the capex prohibitive for the long tail?

Joshua Saxe: I don’t think the capex requirements are that bad for attackers. For one, AI is getting exponentially more compute-efficient, and it’s not clear that’s tapering off. The difference between today and two years ago is night and day. Today you can download the new Qwen 32B model that came out a month ago — trained to do agentic behaviors — and it can probably run on my laptop. It’s not a frontier model, but it’s reasonable to extrapolate that twelve to eighteen months from now you could have a model that does a pretty good job of automating computer network operations running on a humble compute substrate. Even today you can buy an NVIDIA rig you could run in your bedroom that could run a much larger model — trillion parameters, pretty close to the frontier. This is not the kind of thing where you need to spin up a data center, unfortunately.

Jordan Schneider: Can you talk about the breakdown between super-fancy creative zero-day cybercrime versus just social engineering — phishing, convincing people that you’re their mother or their boss?

Joshua Saxe: Today, if not a majority then a plurality of initial access into networks comes from social engineering, not exploits. That’s the case for a lot of nation-state attacks and also cybercrime. Attackers aren’t interested in using the most sophisticated thing all the time — they’re just trying to accomplish a goal, and social engineering is a super reliable way to do that. There have been studies on the efficacy of phishing on human subjects, and people fall for this stuff reliably if you design a good phishing email or fake website. So the sophistication of the strategy an attacker uses and the sophistication of a cyber operation are separate. There are a lot of people using basic stuff.

Jordan Schneider: On the social engineering thing — wouldn’t we have already seen the AI uplift, now that a Russian can write native English?

Joshua Saxe: That’s a good prompt. We are seeing AI used pretty ubiquitously in social engineering for basic stuff like machine translation or cleaning up the language — if you have someone who’s not a native English speaker, or you need to send a phishing email in another language, it’s pretty common to use AI for that. The real question is how much marginal damage is happening due to AI that wouldn’t have happened otherwise. Maybe the emails look more believable now. But you could have found somebody to click on the email even with the bad English — that’s been well proven over the last ten years. The FBI has a good report showing some small marginal damages, but it’s not super large. It’s an open question as to why. Part of it is that criminal enterprises — ransomware gangs, pig-butchering operations — are businesses with their own adoption curves and dynamics around innovation. There’s no reason to expect a ransomware gang in Belarus to adopt AI faster than Deloitte does in its consulting work. They’re on their own adoption curve. Just because we haven’t seen huge counterfactual damages yet doesn’t mean we won’t.

Jordan Schneider: Other thoughts folks should be aware of when it comes to cybercrime?

Joshua Saxe: It’s quite likely that we’re not prepared, as a society that runs on an IT substrate, for some of the reasonably likely futures here. It’s partly — maybe mostly — a policy thing. We need to figure out how to up-level the hardness of our IT infrastructure — by hardness I mean how hard it is to hack. And I don’t think the risk signals are making their way into the American economy fast enough to prepare us. We’re not used to exponentially improving technology like this, and that seems like a risky situation.

Oh, and when making this transcript, Claude kicked me down from Fable to Opus 4.8 because it was flagged for cyber…

If you know anything about rare earths, you know China is great at digging them up and refining them, while America sucks at both. Unfortunately, this is where most analyses end. To go one step further requires understanding the entities leading China’s rare earth juggernaut, and how they built Beijing’s rare earth weapon. That’s exactly what we’ll explore in today’s deep dive. This story revolves around China’s two creatively-named state-owned giants, China Northern Rare Earth 中国北方稀土 and China Rare-Earths Group 中国稀土 (CREG). We’re going to call them “China Northern” and “China Southern” from here on out; CREG doesn’t have Southern in its name, but it’s in the South, and we think it’s a lot catchier.

The fact that these two companies reign over the rare earths world cannot be boiled down to a single factor. These things are messy. To pick it apart, we (Aqib and Nick) look at China’s duopoly from multiple angles, covering the firms’ history, industrial processes, and, on the human level, their workforce.

In our efforts to unpack it all, we pull on several threads, including:

• Why the industry is divided sharply between North and South, and what each region is in charge of,

• The ownership politics beneath that structure, such as which parts of the state own each company and why the South is so much harder to control,

• What rare earths smuggling reveals about China’s state capacity,

• The technical reasons these companies are ahead,

• And what it’s like to work at a Chinese rare earths company.

North vs. South

Nearly all of China’s rare earth production runs through two companies. China Northern Rare Earth, headquartered in Baotou 包头, Inner Mongolia, is the world’s largest producer of light rare earths, the volume elements that go into permanent magnets. China Rare Earth Group (hereafter “China Southern”), headquartered in Ganzhou 赣州, Jiangxi, controls the heavy rare earths — the scarcer and more important elements that give Beijing its chokehold on defense supply chains. Since 2024, the national production quotas have gone to these two firms and no one else.

Why two companies, and why those two places?

Rare earths split into light elements, including the magnet workhorses neodymium and praseodymium, and heavy elements like dysprosium and terbium, the additives that let a permanent magnet keep working inside a hot EV motor or a missile’s guidance systems. The two groups sit in different kinds of ground. China’s light rare earths come overwhelmingly from one place, the Bayan Obo 白云鄂博 deposit in Inner Mongolia, a single colossal open pit where they occur alongside iron ore. The heavies sit in ion-adsorption clays spread across the hillsides of Jiangxi, Guangdong, Fujian, and Hunan, where rare earth ions cling loosely to weathered soil and can be flushed out with ammonium sulfate solution.

In the North, because the ore was all in one place, rare earths were a byproduct of state steel from the beginning. Consolidation in the North was finished by 2015 because of this simplicity.

The South was harder for both physical and political reasons. Ion-adsorption deposits are diffused throughout the provinces, so the region bred hundreds of small operations, chronic illegal mining, and smuggling routes through Myanmar, which has deposits of its own. On top of that sat a crowd of champions backed by provincial and municipal governments, each with revenue and jobs at stake and varied interest in handing its assets up the chain. When Beijing forced the industry into the “Big Six” 六大稀土集团 structure in the mid-2010s, the North got one group and the South got five. The original plan had been exactly two groups, one North and one South, but Southern resistance stalled it at six.

Beijing finally completed the plan in December 2021, fusing the rare earth assets of Chinalco 中铝, Minmetals 五矿, and Ganzhou’s municipal champion into China Rare Earth Group. Guangdong’s assets were not folded in until 2024; the two-firm quota structure dates only to 2024, and Beijing stopped publishing the quota numbers altogether in 2025.

The consolidated structure was in place for barely a year before the 2025 export restrictions, and running those restrictions through dozens of quarreling producers instead of two compliant groups would have been a far messier proposition. The weapon was finished just in time to be fired, which ought to remind us how this arrangement is still settling, still being improvised in places, and still discovering what its own workforce will tolerate.

How much are these companies worth?

The messy consolidation of the South also shows in the obscurity of these companies’ financials. China Northern generated about ¥32.97 billion (roughly $4.6 billion) in 2024 revenue, employs just under 10,000 people, and by April 2026 had a market cap of about ¥177.9 billion ($25 billion), making it the most valuable rare earth company in the world. Revenue jumped to ¥42.56 billion in 2025, with net profit more than doubling to ¥2.25 billion, as export-control-driven prices flowed through.

China Southern, on the other hand, posted about ¥3.03 billion in 2024 revenue, down 24 percent, and a net loss of roughly ¥287 million, less than a tenth of Northern's revenue. But this figure is probably misleading, since it describes only a slice of the group, its heavy rare earth processing and separation assets, not the full mining base or the unlisted parent "China Rare Earths Group" itself, whose total consolidated revenue and profit Beijing has never disclosed. For context, MP Materials brought in just $204 million in revenue in 2024, less than a twentieth of Northern's, and employs fewer than a thousand people. Even after a year of Washington deal-making pushed its valuation up, MP's market cap is around $10 billion (on a good day), well under half of China Northern's.1

The Southern Smuggling Snafu

The South also has to deal with smuggling from Myanmar, bypassing China’s set quotas. The production quota system was predominantly designed as a response to environmental concerns, as unregulated mining, particularly in the South, led to acid and ammonia leaking into streams irrigating rice fields as well as landslides.

Many of the small-time private mine operators that were running dirty mines in the South left for dirtier pastures in conflict-plagued Myanmar. Now, much of China’s heavy rare earth mining comes from Chinese firms operating in Myanmar. The pollution and environmental cost wrought upon Burma is extensive. But the larger thorn for Beijing is that smuggling indicates a lack of state capacity to enforce quotas and secure self-sufficiency.

We can gauge the scope of smuggling by looking at China’s production quotas before they became hidden. The quotas for light rare earths from the North and heavy rare earths from the South have some correlation, as both are used for making NdFeB permanent magnets, which are ubiquitous in EVs, smartphones, wind turbines, and every sort of consumer electronic. NdFeB magnets mainly require light rare earths, but they are often sprinkled with a bit of heavy rare earth to make them heat resistant.2

Given China manufactured about 260,000 tons of NdFeB magnets in 2024, China would need about 1,750 tons of dysprosium, the heavy rare earth, for its dusting of heat resistance.3 The problem is the national production quotas for dysprosium in 2024 were less than nine hundred tons. The gap of almost a thousand tons of dysprosium must be made up by a combination of hidden overproduction from China Southern and smuggling from Burmese mines.

This smuggling epidemic perhaps played a role in China keeping its quotas out of public scrutiny in 2025, and it demonstrates the reality that the central government is not a panopticon. China has its own rare earths issues, and the tug-of-war between Beijing and local interests in the South reveals a governance problem for the central government.

Who Owns What?

The two firms also answer to different levels of the state. Both are state-owned, but China Northern is a local SOE 地方国企 whereas China Southern is a central SOE 央企. The difference is basically who primarily owns and controls the company: the central government or a provincial/municipal government.

Beijing steers Northern through quotas and party discipline, but a lot of the direction is happening in Inner Mongolia. Southern, by contrast, is a central SOE supervised by the State Council’s State-owned Assets Supervision and Administration Commission (SASAC, 国务院国资委), and its shareholding shows the fight it took to build it. SASAC holds 31% directly, a rare arrangement almost like a referee’s seat above three parents — Chinalco, Minmetals, and Ganzhou — who did not particularly want to be merged. The merger supposedly took months of haggling between the center and the localities, with Ganzhou fighting to keep equity for the sake of municipal finances.

Put the pieces together and the idea of “China’s rare earth duopoly” describes a specific and somewhat lopsided settlement. The strategically scarcer resource, the heavies, sits under direct central ownership in part because the South was where Beijing’s regulatory control kept failing. The North remains under provincial supervision because the center didn’t need to do much to corral it.

What Makes These Companies So Good

China Northern and China Southern are the umbrella organizations through which the rare earths flow, but a whole ecosystem lives under the hood to make the system work. The know-how for mining, refining, smelting, and magnet making is one of China’s key advantages over other countries; these kinds of know-how are the crown jewels that Beijing export controlled in October 2025.

It’s Not the Equipment

The equipment ecosystem differs slightly for China Northern and China Southern. The process for mining rare earths changes depending on the ore being harvested. The light rare earths from China Northern are primarily from bastnäsite, whereas the heavy rare earths from China Southern are primarily from ionic clays. We’ll be publishing more on the details of turning an ore into a permanent magnet soon, but the key idea here is that each ore requires slightly different equipment and processes.

For China Northern, the process requires equipment to crush hard ore, froth flotation cells to skim off gangue, and rotary kilns to prepare the concentrate for chemical separation. Meanwhile, China Southern just needs leach tanks and precipitation tanks, as ionic clays are easier to process, though they contain less rare earth by volume.

Regardless of the ore, though, China possesses a deep ecosystem of companies providing all sorts of equipment for mining and refining. Interestingly, these companies aren’t clustered near Inner Mongolia or Ganzhou, but rather spread throughout China. Companies include BGRIMM in Beijing, Naipu Mining in Jiangxi, CITIC and TIEI Extraction in Henan, and so many more.

For most of the rare earth supply chain, the equipment is not specialized.4 Machines used for rare earths mining and refining can be used for almost all mining and refining. These tools have been commoditized globally, meaning that the secret sauce for getting from ore to oxide isn’t in the equipment.

Know-How Is China’s Edge

But equipment is only one piece of the puzzle. The U.S. has its own supply chain for rare earths equipment, and the basic processes for mining, refining, metallization, and magnet making are common knowledge. But there is room for process knowledge to pull China’s industry into a different league.

Decades of dominance over the industry allowed China’s duopoly to develop process knowledge and discover trade secrets that allow them to squeeze every last drop of juice from the earth. This know-how allows China greater efficiency, extracting more rare earths from ore. Everyone in the industry knows you need a rotary kiln, but the roasting recipe is secret. Everyone knows you need mixer-settler banks, but the way China sets theirs up minimizes stages while reaching near-perfect purity.

“The Father of Chinese Rare Earths Chemistry,” Xu Guangxian 徐光宪 developed the process of countercurrent extraction (SX) in the 1970s. This process greatly enhanced the industry’s ability to separate neodymium and praseodymium in the last step of the refining process before smelting into metal. Given it is the final step, both the North and the South needed SX for optimizing their workflows.

If Xu is the common patriarch of China’s rare earth industry, then his intellectual descendants from Peking University are the ones who carried on his legacy, improved upon it, and carried new aspects separately to the North and South. Both regions maintain their own R&D institutes: the North has the Baotou Research Institute of Rare Earths 包头稀土研究院 (BRIRE) while the South has the China General Research Institute for Nonferrous Metals 中国有研 (GRINM). For a sense of scale, BRIRE alone boasts seven hundred staff, with at least two hundred engineers. By contrast, the U.S.’s closest equivalent, the Department of Energy’s Critical Materials Innovation (CMI) Hub, has “more than 300 people in leadership and research roles.” Both BRIRE and GRINM institutes predated Xu’s discovery, but they both separately undertook the task of commercializing and applying his discoveries for commercial purposes. The people at China Northern and Southern, BRIRE, GRINM, and other organizations like the Chinese Academy of Sciences (CAS) are the ones with the know-how.5 China also has the pipeline that feeds talent into these organizations. About eighty mineral processing schools and mining engineering programs collectively graduate thousands of students a year. These students can work at over forty rare earths specialist labs, alongside research institutions and the big two mining companies. The U.S. has only five hundred students graduating from similar programs annually.

Despite that seemingly grim number, the problem for the U.S. isn’t the talent pipeline but rather the destinations for those pipelines. The U.S. doesn’t need rare earths degree programs; they are nearly equivalent to chemical engineering programs. The problem is that chemical engineering graduates go to strike gold in oil fields instead of rare earths facilities. As the U.S. resurrects its rare earths facilities, more and more graduates will enter those companies; over time, the know-how will grow as the processes become perfect. Jack Lifton, Co-Chair of the CMI, conducting an autopsy of the death of America’s mineral capabilities, diagnosed “The West exported not merely factories. It exported an apprenticeship.” By creating capacity, we can funnel graduates into these apprenticeships, and build capability.

Wanting to skip the wait for know-how, we may be tempted to poach talent from China. But poaching wouldn’t be a silver bullet for Americans wanting to one-shot the rare earths problem because, unfortunately, Adam Smith dreamt up the division of labor. People are often siloed to one specific task within SX design, reagent chemistry, or electrolysis furnace operation, and poaching one person does not allow you to replicate the process.

American companies would need to poach entire teams to extract sufficient know-how, though even that may not be enough. The teams at China Northern and CREG are trained and specialized at perfecting the infrastructure for their specific process flow. Each facility is working with a different ore with different impurities, chemistries, and idiosyncrasies, meaning a worker cannot simply plug the process flow for Baotou into the process flow for Mountain Pass.

Is It a Good Life Working in China Rare Earths?

As policy analysts, we spend most of our time thinking about rare earths at the altitude of geopolitics and great power competition. Most people are not like us. The industry is made up of people who went to school, learned separation chemistry, moved to Baotou or Ganzhou, and would prefer to be paid well, enjoy their work, and not spend much time thinking about U.S.-China competition. So how is it going for them?

Let’s start with the numbers, which are public because China Northern is a listed company. In 2024, Northern’s average employee compensation was about ¥183,400 — roughly $25,000 — across a workforce of nearly ten thousand. Solid pay for industrial work in Inner Mongolia, but these workers can also scroll through international job postings on their lunch break and see that they could be making a lot more money overseas.

MP Materials, the company operating America’s lone rare earth mine at Mountain Pass, disclosed a median employee compensation of $92,576 for 2025. The median American mine worker in this industry makes nearly four times Northern’s average employee. And the median software engineer at OpenAI clears somewhere between $555,000 and $800,000!

At the top, general manager Qu Yedong 瞿业栋’s compensation peaked around ¥2.22 million ($310,000), with his 2024 cash pay reported closer to ¥808,000. CEO of MP Materials James Litinsky, on the other hand, made $6 million in 2024 and $28 million in 2025, 303 times his median worker, while personally holding a stake in the company worth north of $700 million. To be fair, we imagine the rent in San Francisco might be a smidge higher than in Inner Mongolia; and Qu’s real career currency is less money and more party rank, a seat on the parent group’s standing committee and a shot at political elevation.6

The business model China has constructed for its rare earth industry is far more experimental and internally divisive than many assume.

In March 2021, Xiao Yaqing 肖亚庆, then minister of the Ministry of Industry and Information Technology 工信部 (MIIT), the industrial regulator that among other things sets the annual mining and smelting quotas under which the two rare earth groups operate, complained publicly that China’s rare earths were selling at the price of “earths” rather than the price of “rare” because of “vicious competition and competitive price-cutting, wasting this precious resource.” （目前稀土没卖出“稀”的价格，卖出了“土”的价格，是因为恶性竞争、竞相压价，使得这种宝贵的资源浪费掉了。） Xiao was actually expelled from the CCP in 2022 for suspected “violation of discipline and law,” whatever that means.

There have recently been some high-profile exits. In the first half of 2025, China Southern’s listed arm lost its chairman, general manager, a deputy general manager, its CFO, and its supervisory board chief, officially for work reasons, but conspicuous enough that the group had to publicly deny rumors of defections.

Claims have circulated that rare earth researchers average under ¥300,000 a year against tech packages of a million or more. One departed research institute employee vented that ten years of work there did not match what a peer gained from a single job hop.

Beyond the limited pay, the work might also come with other drawbacks. In June 2025, the Wall Street Journal reported that MOFCOM 商务部 began demanding rare earth firms submit detailed rosters of their technical personnel, education, research background, roles, and — in some cases — collect their passports to prevent unauthorized travel and the leakage of process know-how. This touches a broader theme about what Beijing asks of its national champions. In March 2025, it was reported that some of DeepSeek’s employees also had their passports taken. By 2026, the practice had spread beyond DeepSeek to researchers across China’s frontier AI labs, making it more difficult for their employees to leave.

For perspective, the people doing the mining and refining of the most consequential resources for the fate of the world are not rockstars. They aren’t wunderkinds that America failed to produce. They are talented chemistry and engineering nerds, and the U.S. has those too.

We might be tempted to think that the rare earths duopoly are organizations of China’s geniuses, whereas the U.S. funneled its geniuses into consulting and finance. That is not the case. China’s duopoly is on top because of their circumstances. Their advantages are in scale, experience, proper organization, and government aid. It’s not magic; it’s just hard work.

Ancient & Classical

Classical Greek Tactics: A Cultural History, Roel Konijnendijk, 2018. Covered in an upcoming podcast. Fun to read a book where from the first five pages it’s clear the author has an absolute command of his sources.

My favorite riff was his exploration of the debate around what actually happened at the Battle of Leuctra, Epaminondas’ famous victory over Sparta in 371 BC. Prussian historians tried to make him into another Frederick the Great grand tactical genius, but the sources just don’t give you enough color to know what it was he did to win this battle, so each historian came up with their own pet theory.

Roel also squared a lot of circles by emphasizing just how corny all Greeks save Spartans thought hoplite training was. Xenophon acknowledges that Athenians “not only neglect to make themselves fit, but mock at those who take the trouble to do so.” Beyond general fitness, weapons training was practically non-existent, and all hoplites save Spartan ones couldn’t turn. This limitation limited just how creative generals could get with their tactics.

Pod with Roel here:

Side note: According to Plutarch, Epaminondas was a fat-shamer. “He hated fat men, and dismissed one of them, saying, that three or four shields would scarce serve to secure his belly, so that his belly would not let him see his own genitals.” And this was pretty hard too: “When the Argives were confederates with the Thebans, Callistratus the orator reproached the cities with Orestes and Oedipus. But Epaminondas stood up and said: ‘We confess there hath been one amongst us that killed his father, and among the Argives one that killed his mother; but we banished those that did such things, and the Athenians took them in.’”

The World of Odysseus, Moses I. Finley, 1954. The history books that still get recommended from the 1950s are some of the best I’ve ever read. This era had writers with so much confidence and style!

*Odyssey on audiobook!

Highly recommend listening to the Clare Danes’ Emily Wilson in parallel with the Ian McKellen’s Fagles rendition. Both bring completely different energies, highlighting their respective translators’ strongsuits. Listening gives you more of a Homeric experience and reminds you how fun and plotty this text is. Compared to the Old Testament, you get much more time for characters to breathe while still benefitting from the extreme economy you get in ancient literature.

Some bits that stuck out to me this read through:

Zendaya crying at Troy’s sacking is a slightly different energy than Athena helping Odysseus make heads roll!

Side note—Claude is pretty fun to ask questions about Homer’s word choice!

2800 years later and still no-one has out-simile’d Homer.

And I love the shade that Odysseus shows his son here:

Korea & the DPRK

Korean Messiah: Kim Il Sung and the Christian Roots of North Korea’s Personality Cult, Jonathan Cheng, 2026. Covered in this podcast.

Kim Il Sung in the Khrushchev Era: Soviet-DPRK Relations and the Roots of North Korean Despotism, 1953-1964, Balázs Szalontai, 2005. Deeply archival, it has some good lore and new perspectives on familiar political beats as it explores the question of why the DPRK got so weird while other divided communist countries like East Germany and North Vietnam didn’t grow into wacko personalist dictatorships.

His answer in brief is that it’s Kim being Kim:

“In my view, one of the principal causes of North Korean despotism was the intense and xenophobic nationalism professed by Kim Il Sung and the former Manchurian guerrillas. Rooted in the latter’s personal experiences and reinforced by Japan’s post-1937 assimilationist policy, this nationalism greatly influenced Pyongyang’s actions toward both South Korea and the ‘communist camp.’ In turn, the goals the North Korean leadership pursued in the field of foreign policy produced a strong effect on its domestic policies. Furthermore, the excessive importance Kim Il Sung attached to direct participation in the armed revolutionary struggle against Japanese rule became a source of excessive elitism.” After Stalin’s death, the purge of the Soviet and Yan’an groups in 1957–1959 subsequently “secured Pyongyang’s political independence from Moscow once and for all.” (pp. 229, 246)

“Kim was neither willing nor compelled to liberalize his regime. Paradoxically, he could usually count on external assistance—and actually badly needed it—yet he distrusted his aid donors, resented being dependent on their good will, and did his best to prevent them from establishing contacts with the North Korean population. His obsession with absolute control over the society was a straitjacket that inhibited the dictatorship from introducing far-reaching changes.” Ultimately, “it was the severity of repression and the extensiveness of state control over society and cultural life that primarily distinguished Kim Il Sung’s despotism” and “the repressive character of the North Korean regime cannot be blamed on external factors alone.” (pp. 258–261, 264, 266)

It was rough out there…

Wild to get quotes like this…

Dresses fade after first wash in DPRK…

Mao handed over territory at a rapid clip to DPRK in return for friendship points.

Grab Bag

Will in the World: How Shakespeare Became Shakespeare, Stephen Greenblatt, 2004. Great reading books where the authors are masters of their sources. Greenblatt isn’t a genius literary critic but I really appreciated the historical grounding that reading this book gave the plays.

My favorite extended beat was how his marriage informed how he portrays relationships in his plays. Will was eighteen and Anne Hathaway was twenty-six and pregnant. The plays kept working through it.

And Will? Through the centuries eighteen-year-old boys have not been famously eager in such situations to rush to the altar. Will might, of course, have been an exception. Certainly, he was able as a playwright to imagine such impatience. “When and where and how / We met, we wooed, and made exchange of vow / I’ll tell thee as we pass,” Romeo tells Friar Laurence on the morning after the Capulet ball; “but this I pray, / That thou consent to marry us today.”

Romeo and Juliet’s depiction of the frantic haste of the rash lovers blends together humor, irony, poignancy, and disapproval, but Shakespeare conveys above all a deep inward understanding of what it feels like to be young, desperate to wed, and tormented by delay. In the great balcony scene, though they have only just met, Romeo and Juliet exchange “love’s faithful vow” with one another. “If that thy bent of love be honourable, / Thy purpose marriage,” Juliet tells Romeo at the close of the most passionate love scene Shakespeare ever wrote, “send me word tomorrow.” When she knows “Where and what time thou wilt perform the rite,” she declares, “All my fortunes at thy foot I’ll lay, / And follow thee, my lord, throughout the world.”

In the Tempest near the end of his career, as the father of two grown daughters, he rails against premarital sex far louder than the scene seems to call for.

The intensity of the dire visions of premarital sex and its consequences may have had much to do with the fact that Shakespeare was the father of two growing daughters. His most explicit warnings about the dangers of premarital sex take the form, in The Tempest, of a father’s stern words to the young man who is courting his daughter. Yet in Prospero’s lines from this play, written late in his career, there is a sense that Shakespeare was looking back at his own unhappy marriage and linking that unhappiness to the way in which it all began, so many years before. “Take my daughter,” Prospero says to Ferdinand, and then adds something halfway between a curse and a prediction:

If thou dost break her virgin-knot before
All sanctimonious ceremonies may
With full and holy rite be ministered,
No sweet aspersion shall the heavens let fall
To make this contract grow; but barren hate,
Sour-eyed disdain, and discord, shall bestrew
The union of your bed with weeds so loathly
That you shall hate it both.

On to the industry. The most underrated fact about Shakespeare is that he took equity. Playwrights mostly sold scripts for a flat fee and mostly died broke. Shakespeare saw that and knew he had to found a company with its own theater in order to improve his lot, and ended up retiring to one of Stratford’s best houses.

Shakespeare not only had a share in the profits from all the company’s court and public performances, but as part owner of the Globe, he also received a portion of the rent that all of the sharers paid (that is, he was in effect in the happy position of paying rent to himself). Imagination, entrepreneurial skill, and unremitting labor had made him a wealthy man; he had, as Juliet’s nurse says, thinking of the sound of coins in money sacks, “the chinks.”

The playhouses shared a neighborhood, an architecture, and an audience with the bear gardens.

Shakespeare’s imagination was excited as well by the less innocuous amusements of the suburbs. Henry VIII bequeathed to his royal children a love of seeing bulls and bears “baited,” that is, penned up in a ring or chained to a stake and set upon by fierce dogs. The bulls—on occasion “wearied to death” for sport—seem to have been more or less anonymous, but the bears acquired names and personalities: Sackerson, Ned Whiting, George Stone, and Harry Hunks (the latter blinded to increase the fun). The game was something of an English specialty—in their travel journals foreign tourists frequently noted that they took in the sight, and Queen Elizabeth treated visiting ambassadors to it. [The UFC of their times!]

“Be there bears i’th’ town?” asks the asinine Slender, in The Merry Wives of Windsor; “I love the sport well.” Shakespeare clearly visited the bear garden in person—he had professional reasons to be interested in what crowds were excited by—but evidently he was less wholeheartedly enamored. The sport, he saw wryly, served to make the Slenders of the world feel more like real men. “I have seen Sackerson loose twenty times, and have taken him by the chain,” Slender boasts. “But I warrant you, the women have so cried and shrieked at it that it passed. But women, indeed, cannot abide’em. They are very ill-favoured, rough things.”

Elizabethans perceived bears as supremely ugly, embodiments of everything coarse and violent, and Shakespeare repeatedly echoed this view, but he also grasped something else: “They have tied me to a stake. I cannot fly,” says Macbeth, his enemies closing in around him, “But bear-like I must fight the course.”

Theater also competed with public punishment as entertainment, and playwrights knew their audience could judge the props against the real thing.

London was a nonstop theater of punishments. Almost daily he could have watched the state brand, cut, and kill those it deemed offenders. London’s many established punishment grounds did not exhaust the locations of these spectacles: in some cases of murder the offender’s right hand was cut off at or near the place where the crime was committed and the bleeding malefactor was then paraded through the streets to the execution site. Such spectacles were virtually inescapable for anyone who lived in the great city.

Lavinia’s ghastly fate in Titus Andronicus—her hands lopped off, her tongue cut out—would have been easy for Elizabethan actors to represent in graphic, realistic detail, for they had seen such things performed in the flesh on scaffolds in the suburbs, near the playhouse. And when Shakespeare’s characters displayed the bloody heads of Richard III or Macbeth, members of the audience could easily have compared the simulation with the real thing.

The moralists’ bill of particulars against the theater functioned mainly as advertising. John Florio’s English–Italian phrasebook treats playgoing as simply what Londoners do on holidays:

The theatrical world Shakespeare found his way into was volatile, speculative, competitive, and precarious. The stage had vociferous enemies: the theaters, preachers and moralists charged, were temples to Venus and other devilish pagan deities; respectable matrons who went innocently enough to watch the plays were quickly lured into lives of licentiousness; men were sexually aroused by seductive boy actors; the Word of God was mocked and piety held up to ridicule; grave authorities were brought into contempt; seditious ideas were planted in the minds of the multitude. Go to plays, thundered one irate minister, John Northbrooke, “if you will learn how to be false and deceive your husbands, or husbands their wives, how to play the harlots to obtain one’s love, how to ravish, how to beguile, how to betray, to flatter, lie, swear, forswear, how to allure to whoredom, how to murder, how to poison, how to disobey and rebel against princes, to consume treasures prodigally, to move to lusts, to ransack and spoil cities and towns, to be idle, to blaspheme, to sing filthy songs of love, to speak filthily, to be proud. . . .” The catalog of vicious lessons continues breathlessly, to be augmented over the years by many other preachers. And as if this were not enough, the wickedness on stage, the theater’s enemies complained, was matched by the wickedness of the audience. At our playhouses, wrote Stephen Gosson in 1579, “you shall see such heaving and shoving, such itching and shouldering to sit by the women; such care for their garments, that they be not trod on; such eyes to their laps, that no chips light in them; such pillows to their backs, that they take no hurt; such masking in their ears, I know not what; such giving them pippins, to pass the time; such playing at foot saunt [i.e., footsie] . . .; such ticking, such toying, such smiling, such winking, and such manning them home, when the sports are ended.” It is a terrible thing, moralists sourly observed, that many who sit happily for two hours to watch a play cannot bear to sit for an hour to hear a sermon.

These charges were leveled in the name of closing down the theaters, but apart from leading to a ban on Sunday performances, they principally served, not surprisingly, to intensify the public’s interest. “Where shall we go?” wrote John Florio, in an English-Italian phrase book that he published in 1578. “To a play at the Bull, or else to some other place.” Florio was born and brought up in London, the son of refugee Italian Protestants. His little language lesson—revealing, as those in modern textbooks are, precisely because it was attempting to be so ordinary and everyday—continued:

Do comedies like you well?
Yea sir, on holy days.
They please me also well, but the preachers will not allow them.
Wherefore? Know you it?
They say, they are not good.
And wherefore are they used?
Because every man delights in them.

When Tamburlaine hit London, Shakespeare understood what league he would have to play in. John Douglas Thompson’s 2014 Tamburlaine at Theatre for a New Audience is still the coolest thing I have ever seen in a theater.

Nothing holds Tamburlaine back, no fear, no deference, no respect for the established order of things: “Emperors and kings lie breathless at my feet.” With these words and with the slaughter of the innocent virgins of Damascus, he takes his beautiful bride, the divine Zenocrate, daughter of the conquered sultan of Egypt. Then, shockingly, outrageously, the play was over, and the crowd applauded, cheering the trampling of everything that they had been instructed with numbing repetition to hold dear.

This was a crucial experience for Shakespeare, a challenge to all of his aesthetic and moral and professional assumptions. The challenge must have been intensified when he learned that Marlowe was in effect his double: born in the same year, 1564, in a provincial town; the son not of a wealthy gentleman but of a common artisan, a shoemaker. Had Marlowe not existed, Shakespeare would no doubt have written plays, but those plays would have been decisively different. As it is, he gives the impression that he made the key move in his career—the decision not to make his living as an actor alone but to try also to write for the stage on which he performed—under Marlowe’s influence. The fingerprints of Tamburlaine (both the initial play and the sequel that soon followed) are all over the plays that are among Shakespeare’s earliest known ventures as a playwright, the three parts of Henry VI—so much so that earlier textual scholars thought that the Henry VI plays must have been collaborative enterprises undertaken with Marlowe himself.

The Great Explorers: The European Discovery of America, Samuel Eliot Morison, 1978. Good for falling asleep you’ll get some great nautical dreams.

Gulliver’s Travels, Jonathan Swift, 1726.

Playmakers: The Jewish Entrepreneurs Who Created the Toy Industry in America, Michael Kimmel, 2024. Getting its own book review.

*First 200 pages of Reiner Stach’s Kafka: The Decisive Years. First biography of an author I’ve read and won’t be the last. Incredible the talent level this genre brings in.

*Motherland: A Feminist History of Modern Russia, from Revolution to Autocracy, Julia Ioffe, 2026. Full review in Q3, would love to have her on a podcast alongside Wang Zheng.

Painting As a Pastime, Winston S. Churchill, 1948. My brain might be too fried to paint or draw anymore, which means I probably really need it. Maybe when kiddo is old enough to sit and draw for more than three minutes?

Madame de Pompadour, Nancy Mitford, 1954. Brilliant beach read, will get its own review later.

The Masks of War: American Military Styles in Strategy and Analysis, Carl H. Builder, 1989.

For more books, check out Q1 2026 as well as past years’ books!

Movies

Screwball

Lady Eve and Some Like it Hot are all-time gems. Life is short you deserve to watch movies that make you smile.

The Lady Eve, 1941. ★★★★★

Some Like It Hot, 1959. ★★★★★

The Palm Beach Story, 1942. ★★★½

Christmas in Connecticut, 1945. ★★½

Airplane!, 1980. ★★★½

Film Noir

Out of the Past, 1947. ★★★★★. Well-treated in this podcast episode.

Cold War & Espionage

The Spy Who Came in from the Cold, 1965. ★★★★★

War

Battleground, 1949. ★★★½

New Releases

Bugonia, 2025. ★★★½

Sentimental Value, 2025. ★★★★

ny theater:

Henry VI Part 1 at the public made me embarrassed for the public. it’s a bad play the actors did not enliven in the slightest, the asia styling of the fights made no sense (ok, so joan of arc fights with a wushu stick...) wish i could’ve walked out faster.

The Winter’s Tale in central park I got a magic ‘warm night, raining just a little bit but not too much that they cancel the show’ night. Within 30 seconds was a night and day experience from Henry VI Part 1—these actors actually know how to make Shakespeare make sense! The midwife was commanding, and fools all excellent.

Anatomy of Melancholy by Eno River Players continues their streak of doing weird erudite stuff that makes me read alot to prep. Verdict: yes, Robert Burton deserved his own play.

Well, I’ll Let You Go brilliant acting by the leads, fun seeing a few actors I recognized at the matinee I attended be so locked in to their peers’ performances

Birthright was the play of the year so far. Skipping forward ten years every intermission hooks the audience into guessing how they all develop. Even at 3.5 hrs I could’ve done with more! I loved the ‘room goes dark when you’re checking your phone in a crowd.’ Politics were cringe to listen to but juggling six characters and their interactions was serious playwriting achievement. Having the mom character does break up/dull the confrontations considerably, would’ve liked for maybe one more dramatic blow up than the ‘oh we’re all friend in the end’ energy it leaves you with. I’ll take 9.5/10 though.

food H1 2026

little uluh—best hunan food in nyc, seriously hot, recommend the fish head, chicken lemongrass curry fish, mung bean drink (’green smoothie’), ‘stir fried double crispy’ (beef aorta + chicken gizzard--now we’re really living, and 辣子鸡。The guy who makes the milk tea also has been very sweet giving whole milk to our daughter when we’ve dropped in. Haven’t seen it full yet but it absolutely deserves to be.

Dumpling Xi--best dumplings in ny, sadly now closed. they weren’t able to offer any vegetarian dumplings for months because home base wouldn’t let them touch any of the recipes.

lin & daughters--excellent buttery noodles and chicken soup.

pizza studio tamaki—heralded tokyo pizza chain drops in, perhaps limited by nyc’s vegetables but dough a perfect nepolitan that retains its chewiness though the middle of the pie

dolar shop--hot pot is not fun and has a ceiling on how good it can be. maybe it’s a fun novelty when you’re 10 but i want people to cook for me

886--excellent fried chicken, cold sesame noodles are good days after, but st marks makes it hard to enjoy

Borgo--great service and backyard, but haven’t actually enjoyed the food as much as others seemed to

Claud--walked in to a table under a stairwell, food was 7/10

Casa Mono--best spanish food in the city, chef taught me how to make scrambled eggs like them, trick is heavy cream low flame and stirring until your wrist hurts

bakeries h1 2026

Papa D’amour--banana bread an A+, scallion bun A, ‘croissant bao’ underwhelming, corn ice cream solid, salt bread bagel a letdown, vanilla ice cream with the butter salt inedibly salty

Lysee--maybe i’ve had too much of them but the preciousness gets a little annoying. very bringable to dinner parties, you want to share bites not eat full cakes

Radio Bakery has nothing asian about it but its focaccia sandwiches are worth the nap required to eat them.

nick & sons--greatest discrepancy between ‘looks like this should be amazing’ and ‘actually not great’ i’ve come across in the nyc bakery scene

apollo bagel--best emerging bagel in the city? It puts popup bagel to shame, some real chef action going on w the cucumber dill bagel in particular. reminds me of a slightly less doughy but still perfectly chewy Bo’s Bagels of Harlem.

Competition between the U.S. and China has entered a new domain: meme warfare.

The lines between memes and reality began to blur a long time ago. The Texas judge who conducted a Zoom trial with a cat filter on his face during the COVID-19 pandemic was a bizarre premonition of darker applications to come. Now, Chinese state media is intentionally layering cute animal filters onto ballistic missiles, and the U.S. White House is splicing bombings of Iran with Spongebob clips, as both countries build up (or actively abuse) their military might. These new tactics put a filter of ease and glamour over the harsh realities of war, with lasting consequences.

The Advent of Advanced Military Advertising

In early January, as the U.S. launched the first of what would become a series of increasingly aggressive military actions from Venezuela to Iran, the White House social media team released a warfighting highlight reel. The viewer watches soldiers shoot at air, run through trees, and jump out of planes, while President Trump’s voice warns the world: “If you attack the United States of America, we will crush you without mercy.” The footage is dotted with color-adjusted shots stylized like a video game, with an upbeat action movie-esque soundtrack playing in the background. When the beat drops, rockets blast off.

Just a couple weeks prior, the Chinese People’s Liberation Army (PLA) Eastern Theatre Command put out an AI-generated montage of animals that shape shift into advanced military weapons, pointed towards Taiwan. The video, released during China’s large-scale Justice Mission 正义使命 military exercises from December 29-30, 2025, features sharks that shape shift into submarines, honeybees that turn into drones, and wolves that become robots with an uncanny resemblance to the Unitree Go2 Air Robot dog. It ends with a swarm of humanoid robots joining the march, visually invoking China’s drone swarm tactics, multiplied across ground, air, and sea domains. The final screen reads: “United forces, clustered weapons, cutting off the path to independence / 联武群戈斩 ‘独’道”.

War propaganda is nothing new. In the past, the United States, China, and countries around the world have appealed to honor, patriotism, and action movie-like depictions of strength in their military communications. But animation advancements, and now AI, have made it easier than ever to create video content that combines the real and the manufactured into a seamless product ready for TikTok or Douyin in seconds. Facing publics that generally oppose military intervention abroad, the United States and China have begun experimenting with an AI-enhanced mode of military advertising that makes war seem cleaner, easier, and less serious.

Lay of the Digital Land

The new era of military propaganda is enabled and likely in part motivated by complex social media ecosystems in China and the U.S. Official PLA accounts often release official content on WeChat 微信, the most-used super-app (messaging, social media feeds, payment, etc.) in China, reaching a large swath of society with over 1.38 billion monthly active users. The TikTok equivalent, Douyin 抖音, dominates creative, short-form video content. Xiaohongshu 小红书 (lifestyle and visual curation) and Weibo 微博 (microblogging) make up other niches. Bilibili / 哔哩哔哩 is popular among Gen Z, home to gamer and other subcultures, with Tencent’s QQ / 腾讯QQ retaining a smaller, messaging-focused youth user base. The PLA distributes across all of these in different ways: WeChat is used for official framing and long-form announcements, for example, while meme-native, remixable material is often released on Bilibili and Douyin.

In the U.S., the digital landscape is flatter and more fragmented. YouTube and Facebook are the only platforms used by a majority of Americans in every age group. The White House and Department of War post primarily to X (formerly Twitter), YouTube, and Facebook, each serving a distinct function. Just 10% of adults use X daily, but since the company’s acquisition by Elon Musk and the dawn of the second Trump Administration, it is known for politically-charged content and press attention. YouTube houses longer-form material, and Facebook retains the oldest and broadest audience across the country. Instagram primarily features reposted White House clips and photos, rather than the video content discussed below. The U.S. government’s approach to the media is also changing rapidly; in June, the U.S. Army shut down and consolidated thousands of military unit-level accounts to promote unity.

China’s Military Cuteness Campaign

Traditionally, Chinese military communications have stressed loyalty and honor, tied to the goal of national rejuvenation after a period of turmoil during the 1900s, from the collapse of the Qing dynasty to the Chinese Civil War. Previous PLA propaganda videos frequently showcased new weapons capabilities. During the PLA’s August 2022 exercises around Taiwan, for instance, the Eastern Theater Command released a music video of soldiers singing, “We’re just waiting to be summoned…towards victory, no surrender, we vow to defend every river and mountain in our motherland“. This was accompanied by footage of missile launches. Individual PLA recruitment videos have paired patriotic framing with displays of military hardware, like in this 2017 video that set footage of military equipment and combat to lyrics about leaving the teenage lifestyle behind to grow up and protect the country. Academic analysis of PLA recruitment ads from 2013–18 similarly found that alongside traditional appeals to recruits’ sense of patriotism and collective duty, contemporary PLA ads also promote military service as a vehicle for personal development and self-fulfillment.

Since last year, however, China’s military has been using playful memes to make weapons and warfare appear “cute”. One tactic pairs military technology or operations with trending soundtracks originally released by popular artists with no connection to the military. In a short video recapping China’s 2025 military parade, Global Times overlaid cartoon peach-pink pig emojis onto Dongfeng (DF-26D) ballistic missiles, the icons swinging left and right across the frame as the missiles march forward. At the end, a burst of 3D hearts and heart-eye emojis floods the screen, in a visual flourish more characteristic of fan art than a weapons showcase. The accompanying soundtrack, “Mommy Said I’m a Pig / 妈妈说我是一只猪”, features simulated children’s voices over a lively drum beat, with lyrics captioned in puffy, candy-colored fonts:

妈咪说我是猪，我是猪就是猪，那也是妈咪养的可爱猪，我妈咪养的好哇，养的圆嘟嘟, 油光水滑毛色亮，是美丽的咪猪

Mommy says I’m a piggy, I’m a piggy and that’s that / But I’m the adorable piggy that Mommy raised / My mommy raised me well, raised me round and plump / Glossy coat, shiny fur, I’m a beautiful little piggy

Similar tactics have emerged across national PLA-affiliated and local government WeChat accounts: Wuhan’s municipal government, for instance, pairs warship fire drills with a soundtrack originally written about children “bouncing and jumping / 蹦蹦跳跳”. The caption reads “Netizens [say]: my mood got better after watching this video / 网友：看完这个视频心情变好了”. The PLA’s Eastern Theatre Command also released an attention-grabbing video set to viral music during a series of military exercises in the Taiwan Strait in late 2025. Its title, “So Close, So Beautiful, Go to Taipei Anytime / 这么近，那么美，随时到台北”, is a play on a trending regional tourism slogan: “So Close, So Beautiful, Go to Hebei on the Weekend / 这么近, 那么美, 周末到河北”. The video is set to the light, bouncy tune of “One Spot,” a popular song by Chinese Kpop artist NINA (Shen Jianrun 沈佳润). If you simply read the captions – in the same puffy font style as the piggy video – you might think a travel vlogger had posted a lighthearted ad for Taipei’s beauty.

“Wow, Taiwan Province seen through the clouds is so beautiful / 哇 云中的台湾省太美了” , as a fighter jet watches the island from above.

“Bro, you showed up too! / 兄弟，你也来了”, as another jet arrives.

“Hey! There are also friends who came by boat / 哟！还有坐船来的朋友”, as a fleet of warships crosses the screen, heading towards Taipei.

Beyond viral soundtracks, military communications accounts have been quick to adopt other trends from Chinese online culture. The aesthetic of Taobao daihuo 淘宝带货 – in which livestreamers demonstrate products live in rapid succession, highlighting technical specs with lively commentary – has been repurposed to explain the special design features of PLA uniforms, gloves, and masks. The phrase “Keep the spark going / (续火花, which typically refers to maintaining an online friendship streak on Douyin, now describes the continuous firing of machine guns.

The United States’ Video Gamification of Violence

Like China, U.S. military advertising has historically framed warfare as serious and honorable. Recruiting campaigns have focused on citizens’ sense of purpose, duty to their country, and desire for financial stability, at times adapted to target specific groups such as women in the 1960s seeking a more independent lifestyle. More recently, the U.S. Army under the Biden Administration tried to improve recruitment with an animated video “Emma” that highlighted the story of a real soldier who grew up with two moms, marched for LGBTQ+ rights, and found her calling in military service. The video and related animated series about service members from diverse backgrounds faced intense media backlash. But the tone of the videos stayed serious, focusing on personal bravery and national duty.

The Trump Administration has left these norms behind. White House and Department of War (DoW) social media campaigns over the past six months make real violence into a game, in real time. As described at the top of this article, one common approach pairs an action-movie or hardcore electronic soundtrack with recycled clips from normal training videos originally posted by one of the military services. Last November, the DoW tweeted a recap of a weapons training and ordinance disposal drill in Montana, which consisted of large dogs and soldiers running around corners and shooting into the sky, set to a country rock soundtrack. Similarly, a separate 40 second reel is captioned “This is what strength looks like….”, with shots of men in camouflage gear running, shooting, and staring into a glowing sunset, effectively romanticizing military operations, as an Instagram influencer might do with a montage of beach walks and acai bowls captioned “This is what self-love looks like…”.

The DoW and White House have also begun to incorporate animated and possibly AI-generated content. In May, the DoW put out a video where War Secretary Hegseth makes the case for America to “build the weapons of the future” over a series of animated visuals. It opens by bashing the Biden Administration for underinvesting in military technology, as we watch an animation of a white-haired Biden in a COVID-19 mask stumble across the screen and face-plant himself into an ice cream cone. In one frame, as Hegseth discusses foreign adversaries, an Iranian leader on crutches tries to fire a drone, misses, and falls over backwards. Everything is labelled “Golden” in shiny gold letters – the Golden Dome protecting the United States, the Golden Fleet ensuring American maritime dominance – giving the United States and its weapons a holy vibe. The video later takes an intimidating turn, discussing military AI investments as we see an autonomous army amass troops in the background. A drone targets and blows up a cave marked with an Iranian flag. The U.S. cartoon drone army targeting Iran and the Chinese PLA’s cartoon drone swarms around Taiwan are eerily similar, even the shape of the drones and animations themselves.

The most extreme example of war gamification emerged in a slew of White House ads since the U.S. began its military offensive (Operation Epic Fury) in Iran in March 2026. This approach combines popular music, television, video games, and sport reels with real aerial footage of bombs hitting their targets, using simple animation techniques or AI to manipulate footage. Like the PLA’s animal filters on a missile, this series of videos appears to try to make an unpopular war seem “light”, like a game that the U.S. will inevitably win. The White House released the first clip on Twitter on March 5th, in which the viewer peers through the electro-optical guidance, targeting and blowing up two trucks, as the TikTok viral rap track “Bazooka” loops its hypnotic “Kablow, kaboom…” in the background. A few hours later, a second video showed a similar bomb’s eye view of buildings being loudly obliterated – cutting straight to a famous meme of Spongebob saying “Wanna see me do it again?” (Originally, the line referred to Spongebob’s superhero alter-ego running to a mountain and back). The sequence loops over and over. These two videos collected 9.4 million and 3.8 million views respectively, compared to the aforementioned weapons training drill from November, which received only 188k views.

In addition to popular media, the White House has used actual video games to attract viewer attention. Call of Duty and Grand Theft Auto (GTA) both featured as eye-catching lead-ins to more footage of targets being hit. The play on GTA features the protagonist’s catch-phrase as the game begins (“Ah s***, here we go again”). After the viewer sees a real target in Iran get hit, the text “WASTED” flashes across the screen – which appears in GTA when your character dies. In a similar format, the White House paired baseball strikes and football touchdowns with missile footage, timing key tackles or swings with each explosion. One week later, the White House combined the same footage of bombings with Wii sports from tennis to archery, accompanied by the classic playful Wii sound effects, reminiscent of a doctor’s waiting room. This breezy background numbs the viewer to the weight of bombs destroying real buildings in a real country, with real people. The next clip took the meme-ification a step further: angry-looking bowling pins march towards the viewer holding a sign that reads “WE WON’T STOP MAKING NUCLEAR WEAPONS”. We cut to a bowling alley, where the text above the pins explicitly reads “Iranian Regime Officials”. A bowling ball painted with an American flag rolls towards the pins, country music blaring, and the scene cuts to footage of bombings as the pins get bowled over.

What Do the Memes Mean?

Making military propaganda short-form, animated, and blended with other “civilian” popular media trends seems to be working. Among Chinese netizens, these memeified videos are generating significantly more attention and positive response than previous PLA content. As of May 2026, the Global Times piggy video, originally published in September 2025, has over 100k views – more than 100 times that of other clips released during a similar time frame. The “So Close, So Beautiful” video from the PLA’s Eastern Theatre Command received 96k likes on WeChat alone, compared with 3k likes on more mundane clips of the same military exercise. It also gained wide traction on its BiliBili (978k views), and sparked discussion on Zhihu and Weibo. The video was picked up immediately across Taiwan’s major outlets.

In the United States, mainstream media outlets (e.g. CNN, Wall Street Journal, New York Times) are more critical, expressing unease at the lighthearted feel of these videos. Right-wing media outlets are relatively positive, especially when comparing Trump’s approach to the Biden Administration’s “woke” advertising. However, one former servicemember who had critiqued the Biden Administration’s diversity ad campaign reacted to the new Trump approach, saying “Most guys I talk to would not send their sons out for a 20-year career in this”. His comments got 35K likes on YouTube, demonstrating that making light of military life does not appeal to everyone. At the end of the day, however, the approach is still increasing clicks. The White House video of the angry bowling pins got 103M views, compared to 300K-400K views for other non-gamified content released in March 2026.

It is possible that the lack of pushback from Chinese netizens results from broader censorship on Chinese social media platforms, while there is still more room to express dissent in U.S. media channels. But the efficacy of meme-ified content in the attention economy is obvious. As AI video generation becomes more common, it will enable increasingly meme-like and cartoon-like content. After the U.S. began the war against Iran, for example, CCTV put out an AI-generated video “流金谷恩仇录 (The Feud of the Golden Valley)” capitalizing on growing global opposition to the war and criticizing military interventionism. The video portrayed White Eagles (the U.S. and allies) trying to monopolize the oil-rich Flowing Gold Valley, until the defiant Persian Cat (Iran) leads a fierce resistance against their tyranny. CCTV says that the video simplifies geopolitical dynamics into a widely accessible fable by combining AI-driven animation with the conventions of classical Wuxia (martial arts) narratives. Both the historical elements and the blatant personification of American power seek to critique U.S. actions in the world and advocate for a transition toward a multipolar global order.

The impact of cute, gamified military advertising mirrors the impact of autonomous and AI-enabled weaponry itself. AI weapons are pitched by new-age defense companies like Palantir and Anduril as making war more targeted, faster, and less bloody. In reality, the costs are often the same or worse for targeted populations, the systems are unpredictable, and military applications of AI are lacking in legal accountability. Similarly, memeified military content gives war a veneer of fakeness and lightness, making it easier for people in China, the U.S., or elsewhere to feel detached from the human impacts of their governments’ choices to engage (or not engage) in military conflict or arms buildups. Experts have characterized these media tactics – “vibe patriotism,” as one veteran put it – as a strategy aimed to desensitize publics to violence and dehumanize victims.

The G2: Meme Convergence

After the Trump-Xi Summit in May 2026, the U.S.-China “G2” popped up across news feeds, as the leaders’ striking similarities in personalistic leadership styles were on full display. As evidenced by the recent tone shift in both DoW and PLA content, these similarities extend beyond personalistic leadership. So what explains the convergence of the two powers’ military advertising models?

First, neither public has much desire for war. These advertising approaches could be a similar effort to manufacture consent, stemming from different domestic causes. The U.S. public is war-weary. In fact, it has voted for candidates that promised, with poor follow-through, to get the United States out of wars in every election since (at least) the 2000s. China, on the other hand, has a noninterference brand to maintain (more brand than practice, given its support for Russia and South China Sea military actions). CCP leadership may be trying to get ahead of potential backlash to continued investment of public resources into military modernization goals, on the nuclear, traditional weaponry, and advanced technology fronts. A lighter public view of war could also preempt backlash if the Chinese government were to decide military force is necessary to achieve control over Taiwan in the future.

Second, these strategies may simply be the best way to get clicks in today’s digital environment. The public wants short-form content and Substack posts, and the two most powerful militaries are responding with audience-friendly content created in the most rapid, cost-effective way possible. The utility of these tactics is evidenced by their use across government entities, like the Department of Homeland Security’s Twitter short, in which ICE officials make arrests with flashy captions to the tunes of popular Arctic Monkeys song “Do I Wanna Know”; or the Chinese Communist Youth League’s use of memes in routine educational campaigns.

Despite different political systems and media environments, the question both the U.S. and Chinese governments are trying to answer is how to sustain domestic appetite for investment in military power – and, ultimately, for its use. The meme-like, aesthetically appealing content shapes recruitment conditions and threat perception among domestic audiences in both countries, creating more favorable conditions for leaders to be able to take military action. As armed conflict spreads and deepens across the world and military technology continues to advance, undermining the seriousness of war could further limit the ability of citizens to accurately evaluate and exert any influence over the decision to go to war. This is particularly concerning in the near term for the United States, with a track record of frequent military action and a system that is intended to be responsive to an informed public. The United States and China are again at the cutting edge; but in the race to meme-ify modern war, nobody wins.

See the full database of downloaded videos released in Chinese and U.S. media outlets here.

Our conversation covers:

• Why even a small amount of AI-driven unemployment could upend national politics, especially if agents start replacing junior white-collar workers and breaking the career pipeline,

• What policymakers can do to protect workers without calcifying the economy, how to collect better data on AI and labor, and whether to subsidize junior hiring,

• The case for funding the AI transition with corporate income taxes instead of token taxes, and why offering the government AI equity or “golden shares” would backfire,

• Data center backlash, strategies for winning hearts and minds, and options for friendshoring AI infrastructure,

• The political economy of safety — why pausing AI development wouldn’t work, but a coordinated slowdown might,

• The coming reckoning between open source and the national security state.

Plus, Kafka on bureaucracy and happenstance, how Anton’s years in esports shaped his career, and why new ideas seem to travel more easily through American policy discourse than European institutions.

Listen now on your favorite podcast app.

AI at the Ballot Box

Jordan Schneider: Am I right? Are things getting weird?

Anton Leicht: There are several different trends happening all at the same time, and they all seem to be compounding — things are getting real. Parts of the US government are waking up to the cyber implications. There are also a lot of incentives leading into the midterms to find very visible symbols of the things that people on the populist flanks of both sides care about, and it turns out AI is one of them. On top of that, AI is being adopted very quickly. These trends all compound to the point where a lot more newspaper pages and media coverage are full of AI news, and it just feels like it’s all getting very big.

It’s still unclear whether this all comes together into one big “AI is a big thing” story, or whether these are actually three separate stories that will die three separate deaths. But it’s happening.

Jordan Schneider: AI is becoming a national issue that people campaign on and vote on. I remember the 2024 Trump-Biden debate, where one of the things I was watching for — instead of senility — was how much they would talk about AI. Export controls came up once, but it was basically not a thing. It was something on the horizon that people were vaguely aware of.

We can look at all the different streams you mentioned — the labor impact, the stock market, inflation, cyber risk, whatever — and it seems reasonable to assume that at least over the next three months, we’re going to pick up steam, not dial it down. In which case, politicians will want to say they’re going to do something about it. The question then becomes — what?

Anton Leicht: They don’t really know just yet. I wouldn’t be surprised if the midterms were still fairly quiet on this, because a lot of policymakers on both sides are still hiding. They don’t really know, first, which way the wind is blowing, and second, what the specific policy asks should be. They’ll let the midterms happen and let some of the discussion unfold. Then, once they see how salient the issue was in post-midterm polling — once they realize people really cared about this in that race and that race, once they have a view of what the political spending and public opinion looked like — they’ll go into 2027, into the next Congress, realizing they absolutely have to do something. They can’t go into another election cycle having done nothing about this. That’s going to be the sentiment on both sides, and then we’ll be in a window where actual policy that gets real national attention actually happens.

Jordan Schneider: Let’s make this podcast proactive and give people an agenda they can campaign around — or at least use to make the world and our futures a better place.

Anton Leicht: We can start with labor impacts, because that is realistically one of the more salient policy areas. With labor markets specifically, there are two questions. First, are there things that actually need doing? Second, are there things that just need doing for political reasons? There is going to be some political need for action on this — and we’ll have to find the least bad policy that addresses it.

There are things we should substantively do regardless. We should be better at knowing what is actually happening in the labor market as AI affects it. We have very fuzzy, spotty data on this. It’s not very good or comprehensive. A lot of the most important and interesting data sits inside the labs. They have very good statistics on usage and diffusion, and on the gap between how much AI is being used for augmentative purposes versus more automation-heavy workloads. That is all extremely useful data for making any kind of responsive policy, but we don’t really have it — only the labs do. Finding some way for governments to access this data and make decisions based on it seems obviously valuable. Then there’s the separate question of actually dealing with the labor market impacts.

Jordan Schneider: Let’s start with statistics. This is classic ChinaTalk — three minutes in and we’re already on the Bureau of Labor Statistics (BLS), which has been beaten up fairly embarrassingly over the past year and a half. The cadence is slow, and the tools aren’t tuned to answer these questions. When you look at the productivity data, it doesn’t tell you anything like “Opus 5 is doing more than Opus 4.8, so we should have some labor intervention.” Your point stands — having a tactile feel for what people are actually using their API calls for is at least as relevant as lagging indicators like who’s getting hired and who’s getting fired.

Anton Leicht: Definitely — data from API calls is simply less lagging. You get a sense of whether these workloads are structured so that they take on an entire task profile analogous to what people currently do as a job, or whether usage is extremely spiky, with all the workload going toward very specific tasks. In the latter case, you could imagine the labor market reorganizing around the tasks being carried out by AI systems, with people filling in the gaps in the jagged frontier and doing the things AI isn’t good at. The alternative would be agents being deployed as full-stack replacements for the workers’ task profiles.

There’s a bunch of data along these lines that gives you some idea of what more adoption would even mean. If governments and policymakers choose to hasten adoption right now, would that push us toward a more competitive workforce that is better augmented by AI tools? Or would hastening adoption mean outright displacement? We just don’t know.

Depending on how these questions shake out, there are very different policy imperatives for dealing with possible labor market impacts. If the workloads are augmentation-heavy, you want to hasten adoption now, because you want to make the workforce resilient and as productive as possible so it doesn’t get shocked into much higher unemployment later, when other competitors have become more competitive. But if adoption right now already leads to high amounts of automation and displacement, then policymakers might have some motivation to slow things down a little.

We have only spotty data on this from some of the labs. If you look at Anthropic’s Economic Index, they have good information — sometimes they volunteer it, sometimes they share it with internal researchers, sometimes with external researchers. But that’s data from one lab. It’s not standardized with data from other labs, and we don’t really know whether Anthropic’s consumer base skews in a particular direction. We need to find a better way to aggregate this data and make determinations based on a larger share of the actual AI usage market. We’re very far from that.

Jordan Schneider: Even when you try to break it down — is this just making workers more productive, or is it drop-in replacement? — the answer is going to be both. Even if it’s a drop-in replacement, that doesn’t necessarily mean everyone will lose their job. There are second and third layers of analysis, and then you have to price in what happens if these models are much better in six months and what that means. With all those caveats, data is good — we’re fans of more data here. What tools do you think smart legislators should be putting on the table?

Anton Leicht: What can we actually do about the labor market impacts once they arrive? The question is what you want to achieve. There are plenty of tools that respond to the immediate political need of avoiding widespread displacement. If you talk to very AI-pilled people, they will tell you about headline unemployment numbers north of 10 or 20 percent. In political terms, that is a crazy world to be in. I don’t think there has ever been a functioning Western democracy with 10 or 20 percent unemployment without things actually going to hell. These are crazy numbers, and if we’re in that world, this isn’t even a prospect policymakers would seriously entertain.

But there’s a scenario one step below that — say we get an uptick of 1 percent unemployment. That is something we might realistically expect during a transition phase, even one leading toward a much more productive and much better economy. We have definitely seen upticks into high single-digit unemployment in the wake of technological transformations. In some ways, that’s a good trend, because it means people will potentially end up in better jobs, the economy gets more productive, and we get better growth. There are all these high-level macroeconomic reasons why you might welcome this kind of trend and say that some intermittent higher unemployment is a feature, not a bug — a necessary step of creative destruction on the way to that better economy.

That’s obviously politically unsustainable, though. We’re not going to have two or three years of 8 or 9 percent unemployment where people say, “Creative destruction sounds great to me — I’m very optimistic about the shape of the future economy once this all settles down.” You need some way to smooth over the disruption enough that you can still reach this better equilibrium without the displaced revolting and without the political economy going completely off the rails. That’s a very different question from asking whether we want to stop this altogether.

If we stay with the question of how to smooth this over, most of it comes down to compensating the specific losers of this transition. That’s really hard, because in past technological disruptions, what we’ve mostly tried to do is take, say, the 55-year-old displaced steelworkers and give them wage insurance or early retirement. It didn’t go perfectly, but there is something you can do about having someone retire ten years earlier. Purely financial compensation works reasonably well at stabilizing their situation and giving them the sense that, while this is unfortunate, it’s not an existentially terrible situation. You can’t do the same thing when the entire pipeline of ambitious young white-collar workers breaks down.

The Political Tipping Point for AI Displacement

Jordan Schneider: When does unemployment become politically salient? For all the talk of AI taking jobs, US unemployment is pretty low — and it’s been low for long enough that a spike will be a big shock to the system and will freak people out. I agree with you that it won’t take 10% unemployment to alarm people.

You don’t even necessarily need a big absolute spike. You just need one interest group, one industry, or one geography to get hit really quickly, and suddenly it’s front-page news. Everybody knows an accountant — everyone has an accountant in their life. There’s no world in which the political class or democracies tolerate 10 to 20% unemployment. You asked which democracies could be fine with that. Well — Europe, kind of, for a decade.

Anton Leicht: But that was tough. It wasn’t a very stable political situation either.

Jordan Schneider: It was also multicausal, and it was a bit of frog-boiling — everyone thought it was going to get better. There was one shock, and it just dragged on for a while. These seem like very unanalogous situations, where it will only take a little bit of targeted unemployment to freak people out.

Anton Leicht: It can be really targeted, because the political incentives for exploiting this come back to the question of political salience. There are a lot of general political messages that are extremely compatible with making a big deal out of AI unemployment once we see any trace of it. There’s the entire story about tech oligarchs building huge data centers, using all this capital to build systems that displace workers and take away their dignity, livelihood, and prospects. There’s the tension between coastal elites and the rest of the country. There’s the story of tech CEOs aligning with the administration. There are so many political incentives to amplify any small instance of labor disruption that I see basically no way this doesn’t become a big, salient issue. It just needs to happen somewhere.

We’re already seeing layoffs being blamed on AI that aren’t even substantively about AI, just because it makes for a convenient story for the firms doing the laying off. We’ll see more of that. We’ll see junior hiring freezes in some domains as people start experimenting with using agents instead. We’ll see, as always happens, some companies going under and blaming it on AI and new technology. Then 20,000 people will be out of work, and the story will be about how AI suddenly wiped out 20,000 jobs somewhere, followed by big media coverage.

The particularly tricky part about AI displacement specifically is that it hits junior white-collar workers. There’s a limited extent to which the rest of the country really empathizes with displaced 55- or 60-year-old steelworkers. But there’s a much broader sense in which many people have children, or know people who want to build a better life for themselves. There’s an American-dream quality to the idea that you can study hard, work hard, and make your way into an ambitious white-collar job — that there’s a career trajectory, a channel for ambition, and a path to a better life.

If that pipeline breaks and a story emerges about widespread youth unemployment and the hopes and dreams of the younger generation being dashed, that’s an even more explosive political story, because so many people will be able to empathize with it. It doesn’t take much for this to blow up — just a few very localized instances.

Protecting the Entry-Level Job Market

Jordan Schneider: Here’s the thing — fake concerns like the water use controversy have already blown up. Fear of data centers as an abstract phenomenon is something I would not have priced in four or five years ago. The fact that this is all just latent noise driving the backlash, as opposed to some acute event, has been remarkable.

We’ll come back to data centers, though. I want to pick up on your point about college grads who can’t find jobs. The one place where the entire society has already been impacted is with middle schoolers, high schoolers, and college graduates, where suddenly the schooling experience has changed. As a parent, you’re stressed out and concerned — abstractly about the job prospect question, depending on how old your kid is, but also very acutely, because homework that used to be homework is now just something ChatGPT can write. That anxiety has already been building. To have it then manifest as, “Okay, I’m 22 and I can’t find a job,” is a very salient, anxiety-producing thing.

Anton Leicht: It is, and it’s just very unclear what we’re supposed to do about it. There’s a real sense in which junior workers don’t add that much — they don’t do that much good work. The general logic is that they do some of the more menial, repetitive tasks — tasks that AI agents are genuinely very good at — and firms hire them as an investment in broader talent pipelines. They keep them around, and a few years later, they’re useful.

But the basic math of hiring junior white-collar workers still has to hold. They need to bring something to the table so it’s worth it for a firm to hire them at all. Otherwise, it’s just too economically enticing to cut costs, skip the hiring, and hope that someone else will hire and train entry-level workers instead.

There might even be a world where this isn’t actually rational in terms of company incentives, because firms lose out by failing to build the pipeline into mid-level positions — the five-to-eight-years-of-experience positions they will really need to fill in the future, especially once there’s more and more AI agent capability to be orchestrated by that middle layer. But they may not build these pipelines, because their short-term incentive is to skip the junior hiring — “Well, we can just use the agents for that instead. We can save this entire hiring round.”

There’s a real concern that this is simply a coordination failure across the economy. Everyone has short-term incentives to skip junior hiring for a year or two, and then we suddenly lose out on educating and building out an entire generation of junior white-collar workers who can’t find jobs. They never turn into those experienced five-to-eight-year workers who could sit above the agents and do the orchestration and augmentative work.

Jordan Schneider: Is your proposal, then, to subsidize the 22-to-27-year-olds — by changing the tax code or something — so they’re less expensive to hire?

Anton Leicht: It’s tentatively the least bad version of fixing this problem that we might be able to come up with. There are still things about it that I don’t particularly love, because the baseline concern is that we don’t want to calcify current economic structures too much. Every subsidy, wage guarantee, and jobs guarantee carries the risk of brute-force subsidizing a very specific economic structure — one that works at one level of technological deployment but doesn’t work at the next. We might want the economy to transition naturally to having very different jobs for junior workers, or perhaps a very different concept of what it means to be a junior worker — when people start doing productive work, and what education looks like before that. We want to leave room for that.

Jordan Schneider: As a veteran of the German political ecosystem, can you draw the comparison? What’s the failure case American politicians should be trying to avoid?

Anton Leicht: The failure case is that we completely locked down the labor market. We made it massively calcified by giving everyone extremely strong labor protections, which means they basically can’t be fired. That in turn means there’s a very high bar to hiring anyone, because once you hire them, you can’t really let them go. It becomes very hard to start new speculative firms, because once those firms start struggling, they can’t actually lay off the people they’ve hired. They’re structurally disincentivized from being particularly ambitious and hiring quickly. It becomes difficult to reallocate talent within your economy to new areas of business and commerce that might emerge as the economy expands around what’s now possible — say, now that there are a lot of very intelligent AI systems. You’re just stuck with the specific economic structure you had when you started handing out all these labor protections.

That works well if the economy is moving slowly — if what you’re doing is iterating on a well-established paradigm. In the German case, we build cars, and we gave strong labor protections to everyone involved in building cars. Those workers are very secure and very safe, and they keep iterating on building marginally better cars. There are basically no new firms — nearly all the economically important firms in Germany are almost 100 years old. There aren’t many dynamic new entrants. That works in a low-technological-pace, high-marginal-iteration environment.

But then technological disruption comes along. Now we’re doing EVs instead of the legacy car industry, or it might be good to pivot the industrial base toward defense production. That turns out to be very, very difficult if the labor market is too calcified. You get a competitiveness problem later on, and at that point the labor protections don’t help anymore either, because the firms are simply no longer competitive. They don’t sell any cars because the EVs are better and cheaper. The firms go under, and everyone gets fired anyway, because the firm is just no longer there. You don’t actually get the protection in the long run.

That’s the failure mode you want to avoid — but you also want to avoid the failure mode where no one has work because everyone is scared of hiring, thinking they can cut corners and save. That’s where the subsidy idea comes in.

Jordan Schneider: The idea that 22-to-27-year-olds are the people we need to protect is interesting, but why not just subsidize humans? Why don’t we cut payroll taxes or something to give humans a better shot relative to AI?

Anton Leicht: We should also do that. We should at least even out the tax incentives. There’s a miscalibration in the fact that if spending on AI agents and AI tokens directly competes with hiring, the structure of the tax system means that CapEx on AI buildout — and maybe even agent spend — is much more favored by the tax code than hiring humans and paying payroll tax for them. That seems like a concerning asymmetry, and it should be fixed — not necessarily to actively boost humans over AI systems, but just to make it not irrational to hire a human for the same thing that a million tokens could do.

The junior hiring problem is something we should be more concerned about ahead of time, because of this pipeline-breakdown effect. You don’t only get the incidental youth unemployment — you lose the talent development that would enable those workers to do better jobs later.

If you increase unemployment by one or two percent across the board because everything is getting more efficient, more people are getting laid off, and there’s creative destruction, that’s not great — and we’ll probably find political responses as it happens. But that’s an economy-wide effect that requires complex handling at an economy-wide level. There isn’t really anything you can do about it surgically. It’s going to be a broad-based economic and political trend that’s very hard to get ahead of.

It is slightly easier, however, to get ahead of this very narrow, specific failure mode. First, these systems seem uniquely, specifically good at doing junior white-collar jobs. Second, there seems to be a strong incentive — and it seems uniquely easy — for firms to skip junior hiring rounds specifically, as opposed to laying off people with more experience, just on the hunch that AI agents might come in. Third, there are long-run consequences of losing out on this generation’s talent development for a few years. Those are features that make this problem distinct.

Jordan Schneider: But don’t we want them all to turn into plumbers anyway? Aren’t we slowing our plumber-nurse-schoolteacher transition?

Anton Leicht: I’m not sure how many jobs the plumber-nurse-schoolteacher-data-center-electrician transition can really absorb in the long run. That’s a big bet. Qualitatively, it sounds great. Quantitatively, I’m not sure the numbers actually work out. At some point you’ll have built the data centers and most of the infrastructure, and I’m not sure the math on that really works.

Jordan Schneider: But by doing this, we’re taking another bet — that there will be white-collar jobs in the future that we want to exist, right?

Anton Leicht: It might be a bridge to nowhere in some sense. You start subsidizing, and then once the subsidy runs out, workers don’t get hired because it turns out you don’t need them anymore. The layoffs start creeping up the job ladder until they reach more and more experienced workers. That’s possible — but in that case, we have bigger problems anyway.

In that scenario, we’ll have very widespread unemployment, well into the double digits. We’ll see the kind of broad displacement that some of the more jobs-doomer types predict, and then we’ll need an on-ramp into actual broad-based redistribution. Having figured out a comprehensive subsidy system that keeps jobs around as a vehicle for delivering what is basically redistribution isn’t the worst infrastructure to set up ahead of time, if that’s the future we’re heading for.

If you ask smart people who do a lot of polling and thinking about the best way to deal with a world where white-collar jobs actually don’t make sense to pay for anymore, not all of them will say it’s just going to be UBI and handouts. A lot of them will say it has to be something like a jobs guarantee, where you wrap the redistribution into some version of a task that is at least marginally useful. You can imagine a world where a marginal job subsidy gets increased and increased until it starts resembling a jobs guarantee- based redistribution scheme. Building the infrastructure for that through a junior hiring subsidy is not the worst on-ramp to that endgame, if that’s where we’re heading. I still hope it’s not, and I still think it’s not — but if it is, I don’t think we’ll regret having spent time building that infrastructure.

Jordan Schneider: The other parallel to subsidizing young white-collar workers is the question of guild protectionism. We talked earlier about how some group is going to be hit hard and fast at some point. Is this a model for that? Because that seems like a failure mode to me — creating some protection for the accountants or whoever it happens to be.

Automation, Lobbying, and the Race for Legal Protection

Anton Leicht: We’ll see almost a race to the bottom on this. Many domains and job profiles are going to make the case that they specifically shouldn’t be displaced by AI. Some professions will end up with fairly arbitrary human-in-the-loop laws. In some professions, we might substantively have a preference for humans carrying them out. Then there are jobs with a sufficiently strong lobby that accounting simply has to be done by a human, or lawyer tasks have to be done by a human — and then that expands to a bunch of other domains. Many jobs and professions will lobby for these specific carve-outs from automation, essentially by asking for legal protections.

It will be hard to sustain that, though, because at some point you get vertically integrated competitors that don’t offer exactly the same kind of service. There will be vertically integrated, essentially AI-only firms providing not quite legal advice, but almost legal advice. There will be a lot of lobbying to keep those out of the market and to ban them — but then there will be open-source versions.

As soon as capabilities exist that can do similar things to what these professionals — accountants, lawyers, whatever — do, it will be very difficult to keep people from using this abundant capability that will be available somewhere. It will be hard to keep them from paying for it. Yes, you can protect employment in the legacy pathways through which these jobs and professions are carried out, but in practice, people will find shortcuts. There will be vertically integrated competitors. In the long run, I just don’t see how it works out if this kind of intelligence is sufficiently abundant.

Jordan Schneider: Speaking of vertically integrated competitors — my wife, for whom I sent out the newsletter seeking a co-founder, has now found one. They are looking for partners, investors, and advisors, as well as potential hires, to take on the incredibly inefficient commercial insurance broker ecosystem.

Anton Leicht: There you go. Many such cases.

Jordan Schneider: Reach out — jordan@chinatalk.media. We’ll put you in touch if you want to be a part of that.

Anton Leicht: If you want to accelerate the atrophying of the legacy.

Jordan Schneider: Your take, then, is that whatever these protections turn out to be, they won’t truly stand in the way of productivity growth?

Anton Leicht: They’ll delay things, and there will be enough risk-averse people who still go with the legacy firms. But if it’s inefficient to keep humans in the loop, these firms will operate at a massive price disadvantage in providing whatever services they offer. Yes, there are a few select examples of highly protected, guild-like structures. But in the long run, the efficiency gains will be so brutal that the next best outside service will win out.

Jordan Schneider: The way this story happens is that the underlying thing gets banned, right? The underlying thing that’s ten times more productive gets banned — not just that you have some protection scheme.

Anton Leicht: You start with the protection, but then you need the lobby to actually stop the alternatives that are popping up everywhere as well. It’s not enough to have the law on the books that says only humans can be lawyers. You also need to crack down on every application of a chatbot that says, “Well, this is not legal advice, but if it were legal advice, I would caution you against doing this and advise you to do that. If you want to go to court yourself, you might want to say these things — but again, this is not legal advice.” You have to stop the models from saying that, and then you have to stop the clever prompts from extracting the same information from the models.

Jordan Schneider: Then you have to stop the open-source models.

Anton Leicht: Right — good luck. At some point, this is just not going to work. The accounting case is the same. You would have to stop the models from being able to process your Excel sheet. At some level of abundance, it’s simply not going to work.

Optimal Taxes and the Trap of Golden Shares

Jordan Schneider: Your other suggestion is raising the corporate income tax. Setting aside Ireland and everything else — why do you like that intervention?

Anton Leicht: We shouldn’t raise the corporate income tax just because I’d like to raise it. The underlying problem is this: if we expect some of the revenue from taxing labor to go away because there’s going to be disruption to the labor market, then it follows that we need the money from somewhere else — especially if we think we need to pay for things like junior job subsidies, or whatever your favorite intervention is. It doesn’t even need to be that. It can be job guarantees, or retraining programs if you think that’s a good idea. Those are also expensive. No matter how this shakes out, if there is going to be some amount of labor disruption, we’re going to need money to pay for the response, and we currently don’t have it. As one leg of this, we’ll have to find some taxation structure that gets us the revenue to pay for whatever we might want to do.

There are a lot of bad ideas about how to do that. The default case for corporate income tax is that it’s a system that already exists. It’s reasonably non-distortionary compared to more targeted taxes, and it’s reasonably effective at capturing value wherever it accrues. It’s agnostic about whether the model developers are the ones who make all the money, or the chip designers, or the adopters. There’s even some initial precedent for getting an international version of this to work. The G20 minimum-floor arrangement didn’t actually happen, but it got pretty far along, and it’s not absurd to think we might get something like it done if AI makes the issue much more pressing. It’s something we have, something that works fairly well, and something we know doesn’t actively backfire. That’s a test that basically no other idea passes.

You could instead do asymmetric taxation of AI specifically. One idea people have been talking about a lot — the Steyer campaign and the Bores campaign have both raised it, for example — is a token tax. The more AI tokens you draw, the more you get taxed. The problem is that it hits the ambitious adopters the hardest and the legacy firms that do the least adopting the least. It’s not clear that’s what you want if you’re worried about AI labor disruption, because what you care about is the ratio within firms. You have a firm with a bunch of human employees, and what you don’t want is for them to reduce the human headcount and increase token spend instead.

Anton Leicht: You’d be very happy if firms kept all their employees while also spending heavily on tokens to augment those workers — keeping people employed, making the firm much more competitive, and ensuring it isn’t as susceptible to disruption from more AI-integrated firms in the future. As long as a token tax only cares about how much you spend on AI, it actively disincentivizes firms from moving toward this augmented, adoption-forward posture. Instead, it incentivizes them to pursue the protectionist interventions you’ve talked about, or to simply hang on for as long as possible and eventually go under in five years because they’ve become uncompetitive. There are a bunch of misaligned incentives in AI-specific taxes.

There are several other proposals along these lines with similar features. The unifying theme is that whenever you try to target AI specifically, you also hit augmentative uses of AI that might be very valuable in staving off the worst version of the labor market impacts. That’s the category we have to trade off against — and in that comparison, the corporate income tax wins.

Jordan Schneider: The corporate income tax is tricky, though, because you generally tax things you want to disincentivize, and companies growing isn’t like cigarettes. But I take your point if we’re comparing it to the alternative — say, taking a 10% stake in OpenAI.

Anton Leicht: What we’re currently taxing — what we’re trying to replace — is labor, and we also don’t want to disincentivize people from working. We’re already taxing a desirable activity purely for revenue reasons, so on that trade-off, we’re fine.

If the alternative is the stake, that seems like a horrible idea. There are two versions of the stake proposal — a few different axes here —, but the first question is whether the government takes stakes specifically in AI developers or in a broad portfolio of AI-related firms. If it’s only the AI developers, you have a regulatory capture problem. If your tax revenue — or the number showing up in your Trump accounts, or whatever — depends not on the economy doing well, but on a very specific subset of the economy doing well, namely these AI developers, you create very specific incentives in how you make AI policy. You want the revenue accretion and value capture to happen at the level of the AI developers specifically, so you don’t want it happening at the level of the infrastructure or at the level of the adopters. Think of the “little tech” joint letters, right? This is exactly the kind of regulatory capture people have actually been worried about over the last few years. If there’s a specific fiscal incentive for the government to ensure that OpenAI and Anthropic capture the revenue and no one else can, that seems obviously horrible.

Jordan Schneider: It’s already strange enough for a president to be indexed on how the stock market performs. I’d much rather have that than a president indexed on this little Trump account thing — a basket of four companies from a single sector.

Anton Leicht: What you might say then is, “Well, then we do the better version of the stake — we get a fairly diversified portfolio, stakes in all kinds of AI companies.” But a lot of the startups, a lot of the swift adopters, aren’t actually listed in any way. You won’t actually have the Trump accounts guys running around writing angel checks in Silicon Valley. That’s probably not going to work.

Jordan Schneider: I’m not sure the Schneider household insurance broker’s competitor is quite open to the economic defense Deal Team 6 coming in for a 5% stake.

Anton Leicht: That probably doesn’t work. If you diversify enough — if your measure for what goes into the stake is “we’d like a diverse spread of firms that might plausibly benefit from advanced AI systems diffusing into the economy” — then in the limit, surely that’s just the stock market. We don’t actually know all that much about where the value goes. You end up with an imperfect simulacrum, a proxy of the stock market that you rebuild in your Trump account and call an AI stake.

It has chip designers. It has a bunch of services firms that have deals with AI labs. It has the AI labs themselves. It has pharmaceutical firms that might benefit from AI R&D, and manufacturing firms that might benefit from industrial iteration. At some point, you’ve just recreated the stock market from first principles and taken a stake in all of it.

In the limit, this collapses to one of two things. Either it’s a stupid sovereign wealth fund set up ad hoc, without sufficient investment structures and investment targets that actually work to give you long-run access to capital. Or it’s a galaxy-brained recreation of the corporate income tax that we’ve been talking about anyway, because it captures value created by all companies. We might as well just do that.

Data Centers and Self-Inflicted Backlash

Jordan Schneider: Let’s talk about data centers. My premise is that this is the companies’ fault. There is so much money in keeping this train going that spending the extra $100 million to make a data center quiet, the extra $100 million to make it pretty, the extra $100 million to put pickleball on top of it so it’s a fun place to be — so that you actually want it in your neighborhood — makes the current situation an incredible series of own goals on the part of this entire ecosystem.

The fact that you see folks blaming the backlash on Chinese propaganda is embarrassing and incredible cope. Shout out to the Chinese for tweeting four things, but that is not where this is coming from. You’ve started to see a little movement — they’re promising your electricity bill won’t go up — but where do you think this leads, and what’s the right response?

Anton Leicht: There are always these stories about how a neighboring city is building a lot of data centers, and incidentally, their new high school also looks pretty nice. On the local level, the deals look pretty favorable to some of these communities. The recently announced OpenAI data center in Georgia comes with a lot of very specific investment in the local communities where the data center is located.

Jordan Schneider: There was one in Michigan where the community got its rec center upgraded. And that was it. You could 10x this stuff, you could 100x this stuff, and these are rounding errors of rounding errors.

Anton Leicht: You say that, but at some point people also have some version of an immune response to massive investment by external firms into their community. You can’t just buy people 100 pickleball courts and expect them to say, “This is great. We just love the fact that these Silicon Valley firms are coming in and spending all this money.”

Jordan Schneider: Fine, maybe not pickleball — maybe it’s data center UBI. All you need is one town, right? You can just ask, “Who wants $20,000 a year for the next 10 years for every household?”

Anton Leicht: On principle, I’m completely with you. The hyperlocal resistance you can probably buy off in a bunch of ways, and that’s probably the way to deal with it. It’s not even malicious — you just let the towns compete for who wants the subsidy.

It’s not quite as easy, though, because not every town can host these facilities. There’s a limited selection, and companies really want the sites with the grid connectors and fast time to power. You can’t actually do it everywhere, so it’s not a clean race to the bottom in the same way it is for corporate local taxes or whatever. But yes, you can do that — that works for the hyperlocal resistance.

I’m not entirely sure it works for the state or federal level resistance. Federal-level resistance is just big-ticket politics. State-level resistance is in a weird spot in the middle — it’s salient and local enough because it’s about something being built in the state. State policymakers have a physical anchor for the things they’re talking about, and it connects to the big federal conversations because it’s about money, foreign influence, and things coming in and being built against the democratic better judgment of the community. That makes a very compelling story on the state level specifically.

Jordan Schneider: It’s an interesting inversion, because generally you have governors and statehouses handing out tax subsidies to bring in investment. No one is screaming all that much about whatever factory shows up in town. This is sensitive not because of the buildings themselves, but because of what they represent.

Anton Leicht: I would be curious to see to what extent that is the case. There was also backlash about Amazon logistics centers being built. Some of the rhetoric is simply, “People are building big things in my neighborhood, and I don’t like them very much.” That’s one part of it.

The other part is people projecting their general concerns — not only about AI itself, but also about the things AI represents to them, like the strange ways tech is changing their lives. They’re projecting that onto the data centers.

There’s also a sense of people being disturbed by the sheer amount of money and the scale of these projects. There’s a fundamental suspicion — so much money is coming in, they’re promising all these really big things, and the facility takes up a third of our town. It’s this huge deal.

Jordan Schneider: But that doesn’t happen for a car plant, right?

Anton Leicht: Because people have rich context on what a car plant is, what it’s used for, and why it’s valuable. They have experience with the fact that it brings jobs. But maybe this also has to do with the speed at which data centers are coming up all across the country. There was never a time when this many car plants got built. If we had a huge boom where dozens of gigawatts of car manufacturing went online all over the country, I’m not entirely sure we wouldn’t also get a similar level of backlash. It’s just going very fast.

It’s a really big thing moving very quickly, and there’s always going to be a homeostatic friction effect that pulls against that to some extent.

Obviously, it’s going to be very hard to build here, which is why people are thinking about where else one could build — somewhere outside the US.

Jordan Schneider: What’s the national power net assessment if the data centers get built in Alberta and Guadalajara?

Anton Leicht: It depends on where exactly you build them. If you build them in allied countries where you have favorable access deals, you actually get to pull in those countries — maybe even get these allies to provide some amount of fiscal backstop or fast-track the permitting. That works pretty well because you reduce some of the political backlash risks, you diffuse some of the political costs, you diffuse some of the capital costs, and you just get better sites.

We don’t have that many sites left in the US that have very fast time to power. We’re running out of behind-the-meter turbines very fast. We’re running out of stable grid connections for quick-to-build data centers pretty quickly. There’s been media reporting on the build-out in Australia, for example — it turns out we can build a few gigawatts of data center capacity in Australia pretty quickly. If the Australians are good with that, and if they like the things that come with the deal — the actual physical investment, maybe something on security integration and favorable access to the frontier models — that seems pretty good for both sides.

It gets less good when, A, you plan to screw over these allies at some point in the future, because then they might decide they want to use their data centers as leverage to stop you from screwing them over. If you’re going to pass eleventh-hour export controls, if you’re going to come up with non-transparent access schemes, they will be incentivized to say, “Well, you can do this, but then you can’t use the GPUs you have on our shores anymore.” To the extent that you’re planning to do that, it’s not great to have them anywhere else.

It also gets more difficult if you put them in countries that are more susceptible to being bombed by a lot of drones, or hosted by countries that might give incidental, complicated access to China. You have to be careful about how good your bilateral relationship is, how high your willingness is to give actual access and treat fairly with these allies, and you also need to be careful about not putting them in very volatile locations or countries that might not have your best interests at heart. But if you put them somewhere like Australia, it’s done.

Subscribe now

Jordan Schneider: It’s funny — there’s an understanding that rare earth production is dirty. But these data centers just buzz a little bit, and if you spend money, they will stop buzzing. Maybe there’s a gas turbine in one, but that’s also a fixable thing over a six-to-twelve-month horizon. To open yourself up to all of these second-order naughty questions in 2029 just because you’re cranky is a bit of a bummer.

Anton Leicht: From a global perspective, it sounds like pretty good news for the world that you can’t actually concentrate all this capacity on the shores of one country.

This is actually rather nice — we can diversify the leverage over this incredibly world-changing technology across the allied world, and then we’ll have some grounding for building an alliance structure that can last a little longer. I feel good about that, but I understand that from the narrow US perspective, the alternative is simply to put all of them in Abilene and be fine with that.

AI Safety vs Political Inertia

Jordan Schneider: We’ll have you back for a middle power show. Let’s talk safety for a second. Where do you see the political economy of safety showing up?

Anton Leicht: I’m skeptical that this is going to be a big-ticket political issue, because it’s structurally very difficult to come up with ways in which AI safety issues — those catastrophic risk issues — become salient. Cyber still feels pretty abstract to me. There is no obvious pathway to a seriously salient but non-catastrophic warning shot on biological misuse. Most of the really scary loss-of-control scenarios that would actually shake people into action are going to be so catastrophic that we should do safety policy before they happen. There are very few scenarios where you get these marginally increasing warning shots and a smooth gradient into high awareness of safety issues.

Even if you get that, it’s going to be very difficult to turn that political awareness into sophisticated safety policy, because to my mind, sophisticated safety policy is extremely hard to do. You can come up with a bunch of really bad ideas that sound like they improve safety for the time being, but they don’t actually.

One of the most obvious tensions people talk about here is internal versus external deployment. You can throw a bunch of rules on AI systems that are being deployed externally. You can nerf them in a hundred different ways, add all the classifiers, make them extremely difficult to access — all these things. It sounds like you’ve wrapped it all in bubble wrap and made the systems very safe. But the actually scary safety scenarios always involve others gaining illegitimate access to models that are only deployed internally — people stealing model weights and using them, terrorist groups stealing model weights, adversaries stealing model weights — or simply loss of control over internally deployed models, as in the case of the Hugging Face incident.

A lot of knee-jerk safety policy just doesn’t do much about safety structurally. It pushes things more toward internal deployment, more toward less predictable, less iterative processes. I’m worried about this idea of riding the wave of political salience up until we get good safety policy. The good news, though, is that I’m not sure we need that, because there are good policy interventions on safety we can make right now. All we need for them is some almost technocratic, policymaker-level awareness.

Anton Leicht: We are getting that. The cyber stuff is freaking out senior officials — people who actually have some influence over what’s going on. It’s not just an AI policy issue anymore. By all accounts, Scott Bessent is super involved, Susie Wiles is super involved. This is a big-ticket issue because people are genuinely worried about the cyber element.

It’s unclear to what extent they’ve generalized from “the cyber thing seems real” to the broader notion that, because the cyber thing is real, we should also take seriously the idea that these models are going to be capable in somewhat unforeseen and somewhat scary ways — and therefore we should take this entire risk portfolio seriously. But that’s still an easier jump to make than starting from “these models are just normal software products.” There is almost a neat consensus that some of these catastrophic risks — especially on the cyber side, but increasingly the other ones too — need to be dealt with. The next question is how we actually do that, and that turns out to be tricky.

Kafka on Bureaucratic Decision-Making

Jordan Schneider: I’ve been reading a fantastic Kafka biography called Kafka: The Decisive Years, and I had ChatGPT compare all of Anton’s writing with the Kafka corpus to find us relevant quotes.

On decision-making, a line from The Castle:

When an affair has been under consideration for a very long time, and even before assessment of it was complete, it can happen that something occurs to settle it, like a sudden flash of lightning at some unforeseeable point, and you can’t pinpoint it later. It’s as if the official mechanism could no longer stand up to the tension and the years of attrition caused by the same factor, which in itself may be slight, and has made the decision of its own accord with no need for the officials to take a hand. Of course, there has not been any miracle, and certainly some official or other made a note of the matter concluding the case or came to an unwritten decision. But at least we here can’t find out, even from the authority, which official made the decision in this case and why.

Anton Leicht: This is not actually as much of a stretch, because there is a sense in which AI policy just happens in emergent ways — the discussion is in the air, things are in the water, and then suddenly these fairly random exogenous shocks happen that don’t really change much about what we should believe about these systems. If you look at all the expert predictions, they’re roughly on track. But then the shock galvanizes the entire discussion and makes people react very specifically to that one particular instance. There is a randomness to that — it’s not very structural and not very predictable.

For example, the AI governance situation we now have in the US — based on the executive order and the ways the labs are currently working out what that order means — is dramatically over-indexed on cyber as a threat model, simply because the incidental thing this all galvanized around was the fact that Mythos had these cyber capabilities. In a very real way, the framework is over-indexed on cyber. The idea that you have lead time to do defensive hardening, the idea that ONCD or the cyber authorities should be deeply involved — all these ideas make a lot of sense for a model like Mythos, where you want to give defenders the model first so they can patch all the holes in the code, and then you can release the model more broadly because the vulnerabilities are closed.

But this fundamentally doesn’t make sense for a lot of other AI risks. It makes sense specifically for cyber. It does not make sense for biorisk, for example. What kind of hardening is anyone going to do with 30 months of lead-time access to a Biomythos that is very good at finding new pathogens? What are you going to do? You’re not going to build 100,000 vaccines and stockpile 500 million doses of all of them before you release Biomythos and Boko Haram finds a way to make a pathogen based on it.

Jordan Schneider: That gets at the crucial thing — you mentioned that priors don’t apply. If you squint at a Democratic platform from 1996 and then look at the Democratic platform from 2026 and the range of policy options being put forward, it’s very predictive. It was pretty clear where people’s priors were, within the normal bounds of what economic shocks you would get and what the reactions would be. The same is true on the GOP side, at least pre-Trump — and even within Trump, there are threads you can see in the water that led to his views being where they are.

The fascinating thing about the Mythos moment is that it showed your priors are not everything. They factor in, but only 20 or 30 percent. We had a fascinating case study of an administration that could not have been more explicitly and publicly committed to doing nothing about this — and all of a sudden they did something in an incredibly intrusive and dramatic way, just because they got freaked out by this thing that no one predicted.

Anton Leicht: In some ways, it’s just a logical reaction to the fact that no capacity was built beforehand. There was no minimal-intervention way to deal with this because the capacity wasn’t there, the expertise wasn’t there, and awareness among many of the principals wasn’t there. If you close your eyes to something that is slowly happening — and happening more and more — for long enough, then at some point you will overreact the first time it really breaks through. That is structurally more likely to happen.

But it’s also just about who is incidentally in the room and who is interested in these questions. It was not a foregone conclusion that — by all accounts and all media reporting — Scott Bessent would be extremely involved in these AI policy conversations. That was downstream of the incidental fact that a lot of banks got freaked out about cyber risk and reached out to Treasury, and things spiraled from there. There are a bunch of process stories like this that come down to incidental happenstance — who gets wind of something first, who gets freaked out first, what happens first. Who knows how it’ll look next time and who’ll be involved? We really do not know.

The Case Against a Pause 因噎废食

Jordan Schneider: To borrow from Kafka’s The Zürau Aphorisms — the US did not have a cage in search of a bird. China currently does. But that’s for another episode. Let’s talk about the idea of a pause.

Anton Leicht: What exactly are you going to pause, and who is going to pause? I don’t want to strawman this, because there is a good version of it — a marginal, coordinated slowdown that moves some capacity away from training and rapidly improving models toward something broader, more diffusion-focused, more inference-focused. You change some of the compute allocation and make the pace of progress a little more manageable. A lot of the labs are warming up to this idea, and a lot of people worried about safety implications are warming up to it too. There could be a good version of something like that.

But consider the actual idea of a pause — shutting this down for an essentially indefinite amount of time until we’ve reached some arbitrary threshold satisfying the notion that we’ve solved alignment sufficiently to keep going. Tearing down the data centers, stripping out the chips in the meantime. That’s a really bad idea, because it’ll crash the stock market and completely derail the entire US economy.

But that’s not even getting at the heart of it, because if you’re sufficiently safety-pilled and concerned about doom, then you think, “Well, of course we’ll crash the stock market — not even a question, since otherwise we’ll all die.” The more fatalistic view is simply this: no one is going to pass policy that crashes the stock market on their watch and sends this infrastructure rally to hell. That is just structurally extremely unlikely.

Jordan Schneider: Setting aside the existential risk arguments, the more relevant domestic political trigger is 10% unemployment. At that point, responses that fall outside the 2026 Overton window start to creep into the conversation.

Anton Leicht: I’m not so sure a pause on AI development really follows from 10% unemployment. A pause on AI deployment definitely follows from 10% unemployment. But I’m not sure the fundamental thing people will be freaked out about at 10% unemployment is, “These systems are getting very good very fast.” It’s mostly going to be, “These systems are suddenly everywhere.” The classic political reaction to that would be, “We’ll just stop them from being everywhere. We won’t let them do the inference anymore. We’ll stop the rollout.”

You can introduce a lot of very restrictive policies on AI agents being deployed without pausing AI development. That’s why I think it’s going to be very difficult to build this broad coalition in favor of a pause — one that ranges all the way from people concerned about very hardcore AI safety ideas to more prosaic political concerns around everything from jobs to environmental impacts. What they all have in common is that they don’t like AI. But in the specific ways they might design a pause, and in what elements of a pause they’re interested in, they’re very different.

If you’re worried about labor market impacts, you don’t have a big problem with pushing all the development toward internal deployment instead — just stopping the systems from being everywhere. Then the labs quietly deploy all their inference internally and build better and better models, until whatever you think comes out at the end of this. Those are two separate questions.

Jordan Schneider: Once you pause, you don’t unpause.

Anton Leicht: But the question is, what do you pause?

Jordan Schneider: If you somehow pause as a labor market intervention, that never gets unpaused.

Anton Leicht: It never gets unpaused in terms of labor market deployment, but the question is whether this just crashes the investment and crashes the stock market. That’s still the most likely outcome. The idea that you get exactly the policy you designed out of a pause, in as volatile a CapEx cycle as what’s happening with the data centers, seems a bit too much.

Jordan Schneider: In an AOC presidency, there are weird arcs to this. What is the interesting conversation to have about a pause, then?

Anton Leicht: The question is whether there’s a version of slowing things down that you can actually make work domestically, and then build a deal around. The domestic question is what the coalition asking for this looks like — how broad is it, and how many people agree that slowing this down in some way is worthwhile?

You can pour a lot of the broad anti-AI sentiment into the idea that America should slow down AI development as it’s happening — for instance, by making it much harder to build data centers. The current moratoriums being discussed at the policy level aren’t pauses on AI development. They’re pauses on data center construction specifically. You can definitely do some of that, and you could probably rally a fairly broad populist base behind it — people who dislike AI for all kinds of reasons. You’ll even get some moderates interested. They’ll say, “We see the anxieties, we’re going to slow down a little, we’re going to figure out what to do next.” You can probably assemble a constituency around that.

Whether it’s a good idea depends on two things. First, what does it do to the domestic AI industry? How much does it push development and deployment underground or elsewhere, with the risks continuing to emerge, just in a different shape?

The second part is what this means geopolitically. There are all these ideas that you’d need a slowdown treaty and would have to get China to slow down as well. Or maybe you don’t, because you’re so far ahead that slowing down doesn’t cost much — and maybe it works because China’s fast-following model will necessarily slow down as you slow down yourself.

I’m most worried about this geopolitical dimension of the pause idea. Frontier AI development is one of the few things the US does extremely well compared to China right now. If you’re sufficiently AI-pilled, you might think that one of America’s few strategic edges is the prospect of ever-quicker recursive cycles of highly intelligent AI development — something that could give you a decisive strategic edge. Maybe not in the strict “decisive strategic advantage” sense, but it would certainly be very helpful.

Jordan Schneider: It’s a nice thing to have.

Anton Leicht: It’s very nice to have. What you don’t want to do is slow down frontier AI development specifically while all the other trends that aren’t in your favor keep compounding. A narrow pause on frontier AI development is essentially an asymmetric treaty, because it asymmetrically hits the one thing the US is good at.

You would want broader concessions in exchange. The most obvious point is that the US is so far ahead right now because it has a much better semiconductor supply chain than China does. If you slow down for a few years and give China time to indigenize more of its semiconductor supply chain — maybe they make progress toward domestic DUV, maybe they get all the way to sophisticated frontier chips at scale — then you restart the race at a later point where you no longer have the chip edge. That seems strategically almost fatal.

Share

What you’d need is a treaty that also captures much of that indigenization effort. You’d need to verify a lot of industrial policy inside China, which is extraordinarily difficult —

Jordan Schneider: Which is just not going to happen.

Matt Sheehan had a recent piece with the subtitle, “AI Safety in Parallel.” If you’re going to have any cooperation, that seems like the form it takes — both countries are freaked out to varying degrees, and maybe you can share best practices. You have enough of the political class as well as the researcher class getting concerned about the potential impact and they talk about the risks.

Whatever the climate change comparison is — if there’s a little thing you can put on the top of your coal plant, that’s not something you’re going to want to keep to yourself, because everyone benefits from it. But some grand bargain seems very unlikely.

Anton Leicht: At that point, it’s mostly domestic policy. And then you also have to wonder — how many of the safety gains are you really getting there?

Open Source Collision Course

Jordan Schneider: What about open source?

Anton Leicht: I don’t see a stable equilibrium here. Most approaches to AI safety currently being entertained by the administration run through guardrails and classifiers. That’s the whole reason they eventually made their peace with the Fable release — they were satisfied the classifiers would be good enough to stop Fable from doing the dangerous things it might be capable of.

Jordan Schneider: What is the classifier-based approach?

Anton Leicht: The model figures out, “This is a dangerous request — I’m going to refuse it.” Or it reroutes the prompt to a weaker model that isn’t dangerous even if it gives you advice on cyber or bio. Everything depends on the model making choices about how to handle your request, and then deciding one way or another not to help once it determines the request is dangerous.

The entire point of open source is that you can train out those guardrails and classifiers. The premise of “the model is capable enough, but it refuses” fundamentally doesn’t work with open weights. Controllability is the only reason governments might be comfortable with dangerous capabilities being part of models at all — the idea that the models won’t actually volunteer those capabilities.

The open-source logic — and to its credit, it works for software — is that you put the dangerous capabilities out there, give them to the defenders and the attackers, let it be a free-for-all, and hope the contest is defense-favored: that defenders are quicker at hardening their software than attackers are at exploiting it. That works decently well for cyber. But there are two open questions. First, does it work for any other risk vector? We’ve talked about the bio side, and it’s not clear that it does. Second, does “we’ll just put it out there and see who’s quicker at deploying it” resonate with any national security authority?

Jordan Schneider: We just talked about the cage in search of a bird. The baseline trend is that natsec-focused policymakers want to control this kind of stuff, and the open-source logic of defense versus offense is inherently uncontrollable and distributed. That’s an obvious collision course, and I don’t see it being reconciled. Instead, we have this letter that everyone has basically been bullied into signing. It also seems unstable, because the political economy right now is clear — the frontier model makers versus everyone else — but that’s not always going to be the case. And the safety question is absolutely going to hit us.

Anton Leicht: Anthropic keeps making these predictions, which is annoying to a lot of people — but that doesn’t necessarily make the predictions wrong. Because Anthropic makes the predictions and then also advocates the corresponding prescriptions in its more legislative-focused work, people read this as a battle between different groups advocating different stances on open source. Yes, some people idiosyncratically dislike open source, and there are definitely people who idiosyncratically love it because they came up through the broader software ecosystems of the early 2000s. But fundamentally, this is a core tension between the idea of controlling very, very dangerous technology and the idea of open source — the two are at odds, and you can resolve that tension in different ways.

Some people will say that if the nation state can’t deal with the broad proliferation and diffusion of these capabilities through open source, then maybe something will have to change about the nation state. That’s a stronger case than “we would like open source models to be widely available” — it’s more intellectually honest. If your take is that the advent of very powerful open models simply means the end of the nation state as it has existed for the last seventy years, fine — we can have that discussion. I’d probably still disagree, but that’s one clean way to resolve the collision course.

What doesn’t work is insisting there’s not going to be a collision — that the trains will pass precisely by each other. We’ll have bioweapon design capability widely available everywhere in the world, and the US government will be perfectly fine with that, cheering on the defenders, buying enough masks and building enough vaccines to hope for the best. That’s the dishonest version of dealing with this. The rest we can discuss.

Jordan Schneider: The letter will be an interesting artifact even six months from now, because we still haven’t hit the threshold where open models can actually do real damage. That is going to come — there’s no world in which it doesn’t, unless the models get closed because China freaks out about them. You can tell the cyber story again, where the closed models are closed enough, everyone has access, and your ransomware apocalypse never arrives — or it only happens in Mozambique and Nigeria, and what does Trump care about those places. But that story works for nothing besides cyber.

Anton Leicht: Which means — all credit — this was great timing. The letter is pre-Mythos-level open source models, and it’’s pre any other risk vector. It comes at precisely the point in time when the open-source story is most compelling: the argument is cyber-focused, where we have some precedent of open source working, and it predates any substantial harm caused by open models — while postdating visible harm trajectories from closed ones. The closed models are above the dangerous threshold, the open models are maybe still below it, and the salient risk vector is the one most favorable to open source. That is a great time to write this letter. As always, very clever by NVIDIA — but I’m not sure it’s going to age particularly well.

Esports and Career Skills

Jordan Schneider: You were at one point very good at League of Legends — not actually on a pro team, but at that tier of being good and getting scouted.

One thing I’ve been fascinated by is that the crossover from esports to other high-prestige, well-compensated white-collar careers isn’t something you see nearly as much as I would have expected, given some understanding of just how far out on the distribution you have to be to be world-class at these games. Do you have a diagnosis for this?

Anton Leicht: Surely you sacrifice a lot to be good at these things, right? Especially if people actually pursue these pro careers, where they live in a house with a bunch of other teenagers who just play video games all day, from age 16 to 22, and do nothing else.

Jordan Schneider: But esports has been around long enough, and these careers end early enough, that there’s still plenty of time — if you are a very sharp, motivated person, which for a lot of these games you have to be — to get back on a track where you’re doing something really cool and ambitious that has nothing to do with playing video games.

Anton Leicht: Maybe, but we also don’t see that with chess players much, right? There’s a fairly specific focus. There’s an extent to which you had a lot of success in a very small pond. People who have been deeply embedded in this scene — as opposed to people who just played a little bit — develop a very idiosyncratic social style, way of talking about things, way of thinking about things, and way of relating (or not relating) to people in the real world. I know these all sound a bit like clichés, but with a lot of people who have done this in the past, you can really tell. It leads to a fairly idiosyncratic development not only of your skills but also of your way of relating to problems. I would imagine a lot of them would be decently good at applying themselves to similarly complex and fast-paced problems, but that’s not usually the bottleneck for people being good at complex jobs.

Jordan Schneider: With League, there’s the speed of processing, which honestly isn’t all that relevant in a job market, and then there’s the meta-coaching side. But there are maybe other games with pro scenes that are more relevant. Magic: The Gathering is one where you do see a handful of folks break out.

Anton Leicht: The more strategic things are — and the less they’re about hitting buttons on a keyboard very quickly — the more you’d expect at least some of this to translate. All things equal, team games are probably slightly better and slightly more promising, just by virtue of the fact that you have to do some more complex social coordination. There’s an element of team leadership in there, and that probably helps.

But if what you do is spend six hours a day sitting in a practice tool, drilling specific combinations and clicking buttons on the screen sufficiently quickly — that doesn’t really translate in any way, other than the fact that a lot of people who have the drive to put in that kind of work at age 14 probably have a general motivation structure that’s somewhat suitable to being successful in the rest of life.

Jordan Schneider: The drive thing is really interesting. There’s a Magnus Carlsen quote where he says something like, “There’s something really boring about me, because there’s something in my brain that lets me spend 12 hours a day staring at these pieces.” To navigate a workplace, to navigate a career, to be interested in varied things, and to make that pivot from one thing you’re obsessed with to another thing you’re obsessed with — you need something else going on in your brain beyond being totally fulfilled by the esports lifestyle.

Anton Leicht: As much as I’d like to make the pitch for why this is a great career background, it’s not like you get into this because you start out very ambitious and this is an obvious channel for your ambition, the way that pursuing high-prestige, competitive things from the get-go is the usual pathway. It starts with you just wanting to play a lot of video games in your room. Then, incidentally, you might discover that you’re good at it, you start iterating, and you start growing ambitious about it. But the entry point into playing a silly video game with colorful fantasy characters on the screen isn’t usually, “Well, I really want to be really good at this, and I want to be really serious about this.” It would be reductive to understand this as a general channel for an intensely ambitious 15-year-old’s mindset. It’s more that people stumble into an ambitious version of themselves once they realize they might actually be somewhat good at it.

Jordan Schneider: The other side of the spectrum is high school debate, right? That’s a genuinely competitive thing you can get obsessed with, and it has a clear meritocracy — the person who wins is probably really good at it. You can over-optimize there too, but at least it’s training your ability to think, analyze, and communicate, which is very different from an aimbot.

Anton Leicht: Incidentally, I stopped playing League of Legends once I got into university debate. The sense of optimizing within some toy environment generalizes to some extent, but some toy environments translate to the real world better than others.

Breaking into American Policy Debates

Jordan Schneider: I want to talk about your transition to America. One of the lovely things about this country is that political discussions are more permeable than I would have expected to people with different accents who showed up very recently. What has that been like for you, a year into your AI experience?

Anton Leicht: I worked in and around German politics for quite some time. When I first started writing about AI policy — much of it specifically about middle powers and places outside the US — I was surprised by how much quicker and more permeable the US environment was to that writing and thinking. The threshold for getting into conversation with people, and getting them to take you somewhat seriously based on a half-decent argument about something they hadn’t been thinking about, is just so much lower.

Of course, there’s a glass ceiling for people coming in from the outside. It’s not only the national security and clearance conversations — you also can’t speak to mainstream politics as comprehensively as Americans can. But putting that aside, the permeability to ideas you put out there is so much higher. You could call it meritocracy, or a less credentialist setup of society. It offsets so much of the standing you might have elsewhere that the pull is very hard to resist — because people actually read what you write, and you actually get to have conversations with people who have some influence over what happens in the world. You could write a similar Substack for five years in a lot of European countries and not get half as far.

Jordan Schneider: Is that a function of parties — where ideas need to come from the inside, or they don’t matter? America already has this think tank ecosystem. Is there something uniquely American about it?

Anton Leicht: There are established pipelines for where the ideas come from for the major German political parties, but there are also established pipelines for where the ideas for the major American political parties and the major wings of those parties come from. They all have their house institutions that are very credentialed and very credentialist. These pathways exist in the US too, and yet there’s still interest in ideas coming in from the outside.

Some of this has to do with the AI conversation specifically, because many of the legacy sources of information and ideas just aren’t up to speed on what’s happening in AI, which makes outside entry easier. But mostly, it’s a political culture that is much less interested in independently evaluating ideas and coming up with new, interesting ones — and much more interested in ideas that have been laundered through all the appropriate channels, signaling that this is a position you can actually take on the floor of parliament. The risk aversion toward entertaining new ideas and thinking about these things more broadly is much, much higher in Europe.

Jordan Schneider: I recently had a conversation with someone who is also writing a middle-powers piece — on economic security rather than AI — who is Spanish. It didn’t even occur to him to publish in Spanish as well as in English, because his attitude was, “Who the hell is going to read this? There’s no audience out there.”

Anton Leicht: It’s changing very slowly. You always get this catch-up effect at some point. The Fable thing changed awareness a bit in the rest of the world, and there are some conversations one can have now that you just couldn’t have three, six, or nine months ago. But there’s always a substantial lag compared to the American conversation, which is its own geopolitical problem.

Our conversation covers:

• Why even a small amount of AI-driven unemployment could upend national politics, especially if agents start replacing junior white-collar workers and breaking the career pipeline,

• What policymakers can do to protect workers without calcifying the economy, how to collect better data on AI and labor, and whether to subsidize junior hiring,

• The case for funding the AI transition with corporate income taxes instead of token taxes, and why offering the government AI equity or “golden shares” would backfire,

• Data center backlash, strategies for winning hearts and minds, and options for friendshoring AI infrastructure,

• The political economy of safety — why pausing AI development wouldn’t work, but a coordinated slowdown might,

• The coming reckoning between open source and the national security state.

Plus, Kafka on bureaucracy and happenstance, how Anton’s years in esports shaped his career, and why new ideas seem to travel more easily through American policy discourse than European institutions.

Learn more about your ad choices. Visit megaphone.fm/adchoices

• $50k prize pool for creative ideas, proposals, and writing pitches broadly related to the China-AI beat, which double as our recruiting drive for new staff researchers/writers. Deadline is August 23rd; learn more here;

• $25k prize pool for evaluation protocols that help foreign policy and national security users learn about AI models’ capabilities, limitations, and what these tools are useful for in diplomatic and strategic contexts. Deadline is September 1st; learn more here.

ChinaTalk writer Irene Zhang went to Indonesia in July. These are her notes from the trip.

Part 1: It’s So Big

A few hours before I fly to Jakarta, my friend in Singapore sends me off over durian. “Indonesia will be a military dictatorship soon. It’s so over for Southeast Asia[n democracy].”

The next morning, I put my backpack on a baggage scanner at the entrance to Grand Indonesia, a glitzy mall in central Jakarta. The security guards’ metallic breast badges flashed past my bleary eyes. It occurred to me that the last time I did this at a place of commerce was in China during Covid.

Some time in 2024, I typed “what is the easiest major language to learn” into Google. The answer, according to the internet’s polyglot hivemind, was bahasa Indonesia, a standardized form of Malay that now unites 288 million diverse peoples across the archipelago. The spelling is consistent, the phonology relatively straightforward for English speakers, and basic grammar can be grokked in a week.

Two years, a whole Duolingo course, and many iTalki lessons later, I landed at Soekarno-Hatta International Airport and managed a terima kasih to the customs officer. With more than 32 million residents spread across its metropolitan area (including the satellite cities of Bogor, Depok, Tangerang, and Bekasi — together known as Jabodetabek), Jakarta is the most populated city on Mother Earth. Indonesia is sometimes referred to as the “invisible giant.” It is the world’s fourth-largest country by population, has the world’s largest number of Muslims, and is Southeast Asia’s largest economy.

Invisibly, to much of the rest of the world, it is troubled at the moment. The Indonesian rupiah weakened around 10% against the USD over the past twelve months. MSCI considered downgrading Indonesia, an upper-middle-income nation, from Emerging Market status down to “Frontier,” alongside the likes of Burkina Faso and Oman. The war in Iran made fuel prices soar, pulling the rug out from under an economy reliant on raw materials extraction. And last year saw violent anti-government protests across the country over corruption and inequality. People threw Molotov cocktails, stormed luxury hotels, and burned down government buildings in Makassar and Bandung, leaving 11 people dead.

My taxi passes Medan Merdeka — one of the largest squares in the world at the heart of the biggest city on earth — and I imagine it filled with people, arm after arm stretching to the end of the darkening horizon.

Part 2: The Coffee Quota

Jakarta has every kind of people in abundance except tourists. Over the course of a week, I managed to check out quite a few tourist-oriented spots: the old colonial town (Kota Tua), National Monument, Glodok Chinatown, and Masjid Istiqlal, the biggest mosque in Southeast Asia. There is the odd European backpacker and Chinese businessperson, and I catch Singaporean accents in restaurants and malls. But I did not see a single foreign tour group all week.

As it is an active site of worship, Masjid Istiqlal can only be visited by joining twice-hourly ad hoc tour groups at the entrance and borrowing appropriate clothing. When I made it there one afternoon, my eyes were burning from air pollution and caffeine intake. My group consisted of two Chinese businessmen working in Indonesia whose families were visiting, a Canadian medical student on summer holiday, two Japanese men, one Japan Airlines flight attendant enjoying a day off, and me.

Istiqlal is Arabic for independence. Sukarno, Indonesia’s first president, insisted that the national mosque be built next to Jakarta Cathedral as a symbol of religious tolerance. Indonesia requires all citizens to register as members of one of six official religions (Islam, Protestantism, Catholicism, Hinduism, Buddhism, and Confucianism) on their identity cards; atheism and interfaith marriage are effectively prohibited.1 My fellow Canadian and I exchanged an uncertain look as we filled out the registration form and identified ourselves as followers of the Buddha.

The great mosque, designed by a Lutheran, speaks the architectural language of the future as it looked like in 1978. Its interior is cool, metallic, and direct, eschewing the soft curvatures stereotypically associated with Islamic architecture. A sanctuary of stainless steel that opens into the city and seems to siphon the world.

There were signs forbidding sleeping all around the mosque, but underneath those very signs were people resting on makeshift cardboard mats.

Just one more cup, I told myself; you won’t be back in caffeine heaven for quite a while. I enjoyed the best V60 pour over of my life in a shop next to Jakarta Cathedral at 4PM. The Dutch East Indies Company (VOC) brought coffee to Java in the 17th century and broke the Arabian monopoly on bitter bean juice. Robusta replaced Arabica as the dominant variety after a disease outbreak. Today, any independent café in Jakarta seems to have single-origin menus more extensive than places charging $10 a cup in San Francisco.

When I strolled around residential Setiabudi that night, I couldn’t tell if my heart was pounding from the heat, caffeine, anxiety, or the existential terror of Jakartan motorcycle traffic.

Part 3: Visibly Chinese

Basically every Grab driver I had started our conversation by asking, “Are you Chinese?”

Being “visibly Chinese” in Indonesia felt quite different from being visibly “Asian” back in North America. The latter places you into one neat bucket within the continent’s racial logic; the former is specific and economically rolecast. “Are you Chinese?”, almost always, is followed by, “Here for business?”

By Day 2, my Xiaohongshu algorithm had detected my IP address and begun filling my feed with videos of mining compounds. Over the past decade, more than $65 billion in investments flowed from China into the archipelago in pursuit of nickel. The metal is used for making batteries, and breakthroughs in those very batteries powered China’s electric-vehicle revolution. Tsingshan Group 青山控股, a Wenzhou-based private conglomerate, entered Indonesia in 2009 and is now the world’s biggest nickel producer. I developed a temporary obsession with vlogs made by its Chinese employees in remote Morowali, on the island of Sulawesi. In exchange for living the most unstimulating lifestyle imaginable on a corporate compound (and suffering under dangerous conditions and brutal labor repression, if they work in blue-collar roles), they are paid double or triple what they might earn back home and enjoy lovely tropical views on the way to work.2

The business-oriented Chinese of Jakarta, as far as I could tell, seemed to coalesce around two main areas. There is SCBD, a high-end business district in South Jakarta, where malls contain familiar offerings like Pop Mart, sauerkraut fish, and Haidilao. And then there is PIK.

PIK stands for Pantai Indah Kapuk (roughly “Kapuk Beautiful Beach,” Kapuk being the name of a nearby neighborhood). It is a stretch of land reclaimed from Jakarta’s northern coast, separated from the city by a slim canal and nearly impossible to reach on foot. Residents live in perfectly geometric compounds of beige townhouses that would not look out-of-place in California, surrounded on the outside by both strip malls and actual malls. My Grab ride dropped me off on “Golf Island.” With no pedestrian pathway anywhere in sight, I swallowed my heart into my stomach and crossed eight lanes of busy traffic to get to the other side.

PIK is generational trauma manifested as manmade landform. In May 1998, nativist rioters massacred more than 1,000 ethnic-Chinese Indonesians and enacted sexual violence against hundreds of women and girls across Medan, Jakarta, Solo, Palembang, and Surabaya. Longstanding prejudice rooted in the Dutch-colonial economic system cast “Chindos,” who constitute merely three percent of the archipelago’s population, as conniving outsiders that robbed the indigenous of wealth, and the US-backed New Order regime falsely suspected the Chinese of communist sympathies.

Today, Chinese Indonesians occupy a social position that seems truly bizarre from the outside. By some estimates, more than half of Indonesia’s private-sector wealth is held in corporations headed by ethnic Chinese, who dominate rankings of Indonesia’s wealthiest people year after year. But they are nearly completely shut out from state power, have little hope of attaining high political office, and cultural and religious discrimination persists. The plight of Chinese Indonesians continues to cast a shadow on PRC-Indonesia relations today, even as Chinese investment pours into the country.

At the center of “Golf Island” are five Chinese-style pavilions that light up at night, attracting groups of locals who pose for social media. To my old-country-Chinese eyes, they are a bit garish. And then I remind myself that those very Chinese characters adorning the archway were illegal a generation ago. It was not until 2000 that Indonesia re-legalized teaching the Chinese language.

Part 4: The Dream

“Whoosh” is the name of the high-speed railway (HSR) line between Jakarta and Bandung. Indonesia’s first, it was financed by a joint venture between China Railway Group and four Indonesian state-owned firms. A fifty-minute economy ride costs 295,000 rupiah (around 17 USD). That’s more than the amount the average Jakartan earns in a day.

Whoosh is supposed to eventually connect Jakarta with Surabaya on the eastern end of Java. Surabaya is Indonesia’s second-largest metropolitan area; Bandung, for all its college-town cool, is a resort city built by the Dutch. (Imagine building California’s first high-speed rail line only to Bakersfield… Oh, wait.) An Indonesian friend in the US tells me that her compatriots have little faith in Whoosh’s complete vision being realized. It’s drowning in debt to the China Development Bank, underutilized by locals, and too connected to the previous president Joko Widodo’s political legacy.

The experience of taking Whoosh is uncannily similar to that of taking HSR in China. Even the seats are exactly the same level of plush. My carriage was 60% empty on a Friday morning, and I was in a great mood as the train whooshed past greenery I hadn’t seen in days since arriving in Jakarta. And just like Chinese HSR, the stop is quite far from the city proper. Travellers must get off at either Padalarang or Tegalluar and transfer to a (free) feeder train in order to actually arrive at Bandung Station.

Cold War fans know Bandung as the host of the 1955 Asia-Africa Conference. Representatives from 29 countries, most of them newly independent, gathered in the West Javan city and drafted a ten-point declaration whose principles eventually birthed the Non-Aligned Movement. Merdeka Building, where the Conference took place, remains standing today. Its exhibit materials are unfortunately thin, but you can still hear a recording of Sukarno’s voice ringing down the very hallways where delegates mapped out a world order after colonialism.

Subscribe now

African-American writer Richard Wright travelled to Bandung in 1955 to report on the Asia-Africa Conference. His notes were published a year later as The Color Curtain: A Report on the Bandung Conference. The work gives us perspective on just how wide-ranging the outcomes of the conference seemed, even before its commencement:

Rumours were sorted out, accepted, abandoned. Was it true that the Japanese were going to offer a public apology for their role in World War II? Was Sir John of Ceylon going to carry the ball for the West? Was a Peking-Delhi-Cairo Axis forming? Were the Moroccans and the Tunisians and the Algerians uniting as a bloc against France? Would anyone from South Africa be at the conference to report on the racial tensions there? Would Red China launch an attempt to capture the offshore islands during the conference? What the hell did Nehru think he was doing flirting with Communist China’s Chou En-lai?

Wright, a sensitive chronicler of Black politics back home, picked up on tensions that Bandung’s ideals could not overshadow. The postcolonial developing world had hierarchies and antagonisms independent from their shared international disposition. Industrialization, manufacturing, and technology allowed some countries to climb up the ladder of economic opportunity at lightning speed, while others missed out. More than seventy years later, one of their own is now a superpower, and it — like every other superpower that has ever existed — wants others to pick sides.

Bandung’s tree-lined streets are easy on the eyes (if no more walkable than Jakarta’s), and I obstinately wandered around while mulling over whether the dream of the Asia-Africa Conference is dead.

Part 5: The Colonies

Indonesians are busy being upset over many things, most of which have nothing to do with their country’s place in modern political history. Their currency is weak, and President Prabowo tried to dismiss the problem by saying that “villagers don’t use the US dollar.” Prabowo’s signature domestic policy, a massive free-school-meals program, is mired in corruption allegations, and the meals keep giving kids food poisoning. He’s starting a new university for civil servants instead of investing in Indonesia’s underfunded public education system. Minorities face particular forms of repression: in May, authorities tried to ban a documentary film about how state-backed bioethanol projects are deforesting Papua, Indonesia’s easternmost territory inhabited by indigenous Melanesians. And last October, Jakarta’s national defense policy declared that one of the main non-military threats facing Indonesia was “the spread of LGBTQ culture.” For what it’s worth, not a single Grab driver I talked to was a fan of the current government.

When I asked a policy researcher at a Jakarta-based think tank whether newfangled AI investments in Indonesia benefit average people, I got a laugh out of him. “You have to remember that many elites are corrupt,” he remarked, “and our internet is not very good.” Batam, an Indonesian island close to Singapore, is a growing hub for data centers, with both American and Chinese companies investing heavily in hyperscaler construction. In July, Indonesia signed on to be a founding member of the World Artificial Intelligence Cooperation Organisation (WAICO) in Shanghai. Dozens of think tank papers pore over the world’s fourth-largest country’s AI policy, development, and prospects. With pretty language and prettier graphs, most of them politely sidestep the possibility that in the short term, this might not benefit most Indonesians whatsoever.

Consider one definition of digitization: making a society, a culture, and/or a polity readable for machines, and then continuing to build that society/culture/polity around the logic of such machines. It’s a process that is largely complete in the developed world, but is still an ongoing, contested process in most other places. And then consider a proposition: the extent of productivity growth possible via AI adoption is a function of the extent to which an economy is accessible for computers. This hypothesis is not the only theory that makes sense, but it is a potential future — and it is one that could undo the path to prosperity that industrialization promised to the developing world.3

Nearly 60% of Indonesia’s workforce is engaged in informal employment. Their daily work is unprotected by labor laws, and their tasks are frequently not documented as the kind of data that computers can use. In developing Asia — exemplified, most recently, by Vietnam’s meteoric rise — the path to high income still largely runs through manufacturing. Barriers to growth are old, familiar ones: weak institutions, brain drain, and exploitation by foreign powers in the name of investment. Does AI change any of that? And if not, how do you convince societies like Indonesia’s that it might matter for their future? The researchers I talked to in Jakarta tell me that frontier models are so bad at Indonesian — a relatively easy language for many humans to learn, and the official lingua franca of 288 million people — it doesn’t even cross their minds to write prompts in Indonesian when they use AI for research.4

Things could go more badly if AI evangelists’ dreams come true. Should productivity really take off in the rich world, there will be ire from people in developing nations who gave up land, electricity, and water for a nonexistent slice of the pie. The charge Silicon Valley will face then will be worse than steamrolling local democracy in small-town Wisconsin: it will be neocolonialism.

In 1879, Thomas Edison invented a long-lasting lightbulb. Today, at least 1.18 billion people still do not have meaningful access to electricity. ARPANET celebrates its 57th anniversary this year, with three quarters of humanity connected online. Only around 16% of the world has used a generative AI tool so far. How long will everyone else wait this time?

Miscellaneous Recommendations

Best food I tried: “Real” satay. Completely different from anything I’ve had in Southeast Asian restaurants elsewhere and every bit as gloriously smokey as one might imagine. Bakmi, Soto Betawi, and Sundanese cuisine get honorable mentions.

Best coffee shop I went to: Filosofi Kopi.

Best unexpected thing: Jakarta’s street art! Here’s one about cake:

Thank you to G, Z, Y, F, J, and all the Indonesian and Southeast Asian friends whose perspectives informed this little blog post. In no way can I do your beautiful region justice, but I hope it’s fun to read a foreigner’s earnest impressions!

During paternity leave, the new influx of dad energy compelled me to read close to 15 titles on the American Civil War.

The Civil War is a case study in traditional ChinaTalk themes like technological change and organizational design. It is also a uniquely well-documented war, just on the cusp of modernity, with gripping characters and human drama basically wherever you look.

To discuss all things Civil War, ChinaTalk interviewed Jeremy Neufeld, Director of Immigration Policy at IFP and author of an incredibly nerdy Civil War Twitter account, as well as Wayne Hsieh, a professor at the Naval Academy and co-author of A Savage War, an excellent new military history of the Civil War.

We discuss…

• How the Civil War played a role in America’s rise to superpower status,

• The impact of technological innovation on the war, including the telegraph, the steamship, and the railroad system,

• The Confederate invasion of Kentucky and other strategic missteps,

• Strategies for mobilizing industry and civil society for warfare,

• Differences in institutional culture between the Northern and Southern armies.

Learn more about your ad choices. Visit megaphone.fm/adchoices

• $50k prize pool for creative ideas, proposals, and writing pitches broadly related to the China-AI beat, which double as our recruiting drive for new staff researchers/writers. Deadline is August 23rd; learn more here;

• $25k prize pool for evaluation protocols that help foreign policy and national security users learn about AI models’ capabilities, limitations, and what these tools are useful for in diplomatic and strategic contexts. Deadline is September 1st; learn more here.

Why is North Korea so weird? Is it Stalinism? Maoism? Japan-style emperor worship? Ancient Korean hermit kingdom energy? No, argues Jonathan Cheng, Beijing bureau chief for The Wall Street Journal and author of The Korean Messiah: The Failed Dynasty That Shaped the Hermit Kingdom. You’ve all been sleeping on the influence of 19th-century American revival Protestantism.

To discuss why Kim Il-sung kind of thought he was Jesus — or at least wanted to convince everyone else he was — we have the honor of Jonathan joining ChinaTalk today, with Alex Boyd.

Our conversation covers:

• Billy Graham meets Kim Il-sung — how America’s pastor visited Pyongyang in 1992 and bonded with the dictator over their shared Presbyterian childhoods, while touring a birthplace resembling the nativity story.

• From “Jerusalem of the East” to totalitarian theocracy — how 22-year-old missionary Samuel Moffett brought eschatological founder energy to the wickedest city in Korea, and transformed Pyongyang into the most Protestant city in Asia, complete with jazz clubs and a Chevrolet dealership.

• The Manchurian cauldron — how the lawless frontier of 1920s Manchuria blended Christianity, Korean nationalism, and Marxism to forge Kim Il-sung’s messiah complex.

• Jim Jones and Kim — how the Jonestown cult leader became obsessed with Kim Il-sung, adopted Korean children, and traded tips with North Korean diplomats on running a closed religious society.

• Christianity as political technology — how Kim weaponized Protestant doctrine and ritual to build a state religion that has outlasted the Soviet Union by three decades.

Plus, why North Korea is best understood not as a nation-state but as a religious society, and what Hong Xiuquan and Xi Jinping have to do with any of this.

Listen now on your favorite podcast app.

Billy Graham and North Korea

Jordan Schneider: Let’s start with Billy Graham’s visit to North Korea.

Jonathan Cheng: Billy Graham is arguably the world’s most famous preacher, the embodiment of Christianity in the 20th century. They called him “America’s pastor,” hanging out at the Oval Office with every president from Harry Truman to Barack Obama. He’s famous for filling stadiums around the world with his evangelistic crusades, but he was also trying to take his message across the Iron Curtain — holding crusades in the USSR, Poland, and Czechoslovakia when that wasn’t easy.

But North Korea was a tougher nut to crack. He had a surprising amount of personal connection there — his wife actually went to high school in Pyongyang in the 1930s. His first major evangelistic trip outside the US was to the Korean Peninsula during the Korean War, and in 1973, he preached to about 1.1 million people in the center of Seoul. He had this deep connection to Korea, but until he got his invitation from North Korea in 1991, he had never actually been there. He finally visits in early 1992. That’s where the book starts.

Jordan Schneider: Can you share Billy Graham’s impressions of the country he visited?

Jonathan Cheng: Graham had been behind the Iron Curtain before, but he knew North Korea was different. It wasn’t just a random socialist country to him; he knew Kim Il-sung had been raised in the Christian church.

He flew there in the spring of 1992 and was immediately taken to Kim Il-sung’s birthplace — a humble thatched hut where the state claims its God-King was born under humble circumstances. Upon arriving, Graham remarked, “All they’re missing here are the three wise men and the manger,” obviously referencing the birth of Christ.

Graham spent several days in Pyongyang, giving sermons at Kim Il-sung University and in the two churches Kim built toward the end of his life. There, he preached his usual message: John 3:16, the Christian gospel distilled to one verse. He also gave Kim Il-sung a copy of his 1953 book Peace with God and talked with him about the North Korean leader’s Presbyterian upbringing.

When Graham came, he played a bit coy, claiming he had no idea why he was invited. Deep down, he knew it was because of Kim’s personal connection to Christianity, but he also recognized the political angle. Kim Il-sung saw Graham as someone with access to the very top levels of American leadership — George H.W. Bush during his first visit, and Bill Clinton on his second.

The Protestant Roots of North Korea

Jordan Schneider: Maybe with that as a little tableau to open us with — Kim Il-sung bonding with, of all people, Billy Graham about their shared Christian upbringing — I want to take it back all the way to the 1800s with another American preacher, Samuel Moffett. He brought out a slightly different energy than Billy Graham, but was perhaps even more influential in the arc of Korean history.

I loved the intro — the first 100 pages were really novel history to me. Can you paint us a picture of the ideology and mindset that led a 22-year-old in the 1870s and 1880s to pack up their bags for Asia?

Jonathan Cheng: Samuel Moffett is a fascinating character. Billy Graham is a bit of a teaser at the beginning of the book, but the narrative really gets going with Moffett. He was born during the Civil War in Madison, Indiana, right on the border between the Union and the Confederacy. He attended an abolitionist church and went to Hanover College, a small school known for its religious fervency. Fun fact: nearly a century later, Mike Pence attended Hanover and was president of the exact same fraternity, Phi Gamma Delta. Both were very earnest, upright, serious Christian men.

To give you a sense of Moffett’s strict Presbyterianism — he was staunchly anti-alcohol and anti-cigarettes, but also opposed to card playing, horse racing, dancing, and even rhythmic clapping in church.

The prevailing ideology for college-educated Midwesterners at the time was this immense Christian fervor. Their highest aspiration was to spread the gospel to the literal ends of the earth. And they meant it — Tahiti, Greenland, the interior of Africa, or somewhere so exotic it was called the Hermit Kingdom. Today, Korea exports Samsung and K-pop, but back then, it was hard to imagine a more isolated place. Moffett didn’t even specifically ask to go to Korea. He just said, “Send me somewhere really far away.”

The backdrop here is an America that had just reunited after the Civil War, flush with new wealth and innovations like the steam engine and the telegraph. Manifest Destiny had reached the West Coast and was spilling into the Pacific. The Christian version of this was a movement calling for “the evangelization of the whole world in this generation.” Moffett saw himself as part of this enterprise, driven by an eschatology that believed spreading the gospel to every last person on the planet would actually hasten the second coming of Jesus.

Jordan Schneider: The ethos was about bringing about the end times and going on this great foreign adventure. There’s a big founder energy part of all of this too. It’s really a zero-to-one thing, especially when you’re the guy showing up in Korea.

Other things people need to understand is that this was dangerous. These missionaries didn’t just get diseases — they got killed occasionally. When you went, this wasn’t necessarily like Peace Corps, where you do two years and you’re back. This was decades. Maybe you’d see your parents once again before they died in 30 years or something. It was very much a burn-the-boats style commitment, as opposed to the type of Mormon missionary tours or Peace Corps tours you see today.

Alex, fellow Peace Corps veteran — how did you see the echoes of this story in the ethos that you walked into?

Alex Boyd: I see the zeal — not necessarily in the Peace Corps to spread God’s light. But I also see competition. When Moffet is going to South Korea, he has his mission to spread God, to finish American Manifest Destiny or whatever. But he’s not the only one who has designs on Korea. The Japanese have also gone through their own Meiji Revolution and have the idea that they’re going to become a Western power and engage in this. I see echoes of that competition in the Peace Corps where you’ll meet other groups in China doing their own thing.

But I was curious — what’s the situation in Korea when he lands? You say it’s isolated. What’s the political situation? What’s the state of play when 22-year-old Sam arrives? Where did he arrive first — Incheon?

Jonathan Cheng: Yes, he arrived in Incheon first, the port city of Seoul. At the time, Korea was still the Hermit Kingdom under the Joseon Dynasty, which had been in charge for almost 500 years — incredibly long by dynastic standards. It was very much seen as the “sick man of Asia,” arguably even more so than China.

Everyone, including the rapidly Westernizing Japanese, sensed that the dynasty was nearing its end, triggering a scramble for Korea. Sailing there took weeks, and as Jordan mentioned, it was effectively a one-way ticket. Moffett’s farewell at his church in Indiana felt like a wake. His congregation thought they’d never see this bright, earnest young man again. But Moffett didn’t see it as sad — to him, there was no higher calling than sailing to this kingdom.

By sheer coincidence, he arrived in Incheon on his 26th birthday in January 1890. The Joseon kingdom was crumbling under uprisings and mutinies, though the newly arrived missionaries were mostly unaware of the political turmoil at first. Other Presbyterian and Methodist missionaries had been there for about five or six years, but they weren’t having much luck. They had some success in Seoul, but almost none down in Busan — it took them three years to get their first baptism.

Lacking traction, they decided to try the North. They told Moffett, the newest arrival, to head up to Pyongyang, the second-largest city, and give it a whirl.

Jordan Schneider: How do we go from one baptism in five years to product-market fit?

Jonathan Cheng: The key was finding the right target demographic. It wasn’t in Seoul or Busan, but up in the northwest, closer to the Chinese border and the old smuggling routes. It was in Pyongyang that Moffett tapped into a massive desire for his message.

He couldn’t have predicted it, though. At the time, Pyongyang had a reputation as the wickedest city in Korea — the missionaries called it the Sodom and Gomorrah of the peninsula, and I’d call it Las Vegas. It was famous for having the best saloons, the best kisaeng (dancing girls), and incredibly belligerent residents known for street brawling. Hundreds of men would gather, split into teams, and hurl stones at each other. The missionaries noted how crazy it was that Koreans were so fascinated by this deadly sport. Spectators brought the same enthusiasm to this deadly sport that Americans brought to baseball.

Jordan Schneider: In 1900, you had football fatalities too. They shouldn’t have been casting too many aspersions.

Jonathan Cheng: True. The concussion risk was very high back then for sure. But it wasn’t a very promising place to bring the Christian message. Yet it turned out to be the turning point.

The Rise of Christianity in Pyongyang

Jordan Schneider: By 1927, we have a quote that Pyongyang has a Chevrolet dealership, three-story Western houses, massive churches towering over straw huts, jazz clubs, and factories churning out socks and rubber products. “The most Protestant city I’ve ever been in,” writes a pastor’s son. How on earth do we get from A to B, Jonathan?

Jonathan Cheng: It starts with Moffett. Almost immediately, he found a huge hunger for his message that didn’t exist in Seoul or Busan. He was overwhelmed and quickly called for reinforcements, telling them, “I don’t know what’s going on here, but everybody wants to know about this Christian gospel.” They established church after church and could barely keep up with the demand.

Then the First Sino-Japanese War hit in 1894-95, and Pyongyang was utterly flattened. Because Moffett stayed until the bitter end and was among the first to return, locals began to see him as a deeply committed, almost prophetic figure. People packed his churches faster than he could build them. In the ruins of the city, he bought up about 120 acres of land right in the middle of Pyongyang.

So, a city now famous for statues of Kim Il-sung was, at the turn of the 20th century, filled with Presbyterian institutions. Moffett founded churches, the first modern college in Korea, and numerous schools — including one Kim Il-sung would later attend and the high school Billy Graham’s wife went to. He even built the largest Presbyterian seminary in the world, plus schools for the blind and deaf, and a YMCA.

Pyongyang became known as the “Jerusalem of the East.” Church leaders worldwide flocked there, stunned that while missionary work had stalled for centuries in China, India, and Japan, Pyongyang transformed into a massive beacon of Christian piety in just a few years.

Alex Boyd: One of the fundamental parts of that explosion of growth was this charismatic emotional aspect of Christianity there. Where did that come from? Moffett didn’t like rhythmic clapping. Presbyterians are sometimes derisively called “the frozen chosen.” This anticipates what we see in North Korea today — these videos of wild clapping and profuse crying at the appearance of the great leader. What’s the origin of this charismatic emotional aspect?

Jonathan Cheng: You’re absolutely right. The Presbyterians were a buttoned-up bunch and certainly didn’t intend for this highly charismatic, emotional strain of Christianity to flourish. Yet, that’s exactly what happened during the Great Pyongyang Revival of 1907.

The missionaries had hoped for a revival, but they didn’t bank on an outpouring of mass hysteria. There were incredible scenes of new Korean converts rushing to the podium, slamming their heads on the floor, and publicly confessing to every sin in the Ten Commandments — from theft to adultery to murder. This fervor swept across the Korean Peninsula and up into Manchuria.

Unable to restrain the emotionalism, the missionaries eventually just threw up their hands and said, “We just need to go with it. This must be what God wants.”

There are many theories as to why this happened. Korean Christians often point to the influence of native religious traditions like shamanism, Buddhism, and Confucianism. Christianity naturally takes on a local flavor wherever it spreads, and the missionaries noticed this difference immediately. They were astounded by pre-dawn prayer rituals where converts would wake up at 4:00 a.m. to pray for hours, or how they would memorize entire books of the Bible and recite them at will — a practice missionaries attributed to Korea’s scholarly Confucian roots.It was incredibly impressive and again, just blew the minds of these missionaries.

Christian Revival and Korean Nationalism

Jordan Schneider: We have to call back to Mike Pence. It’s interesting because you do have bits and pieces of this in American religious history, but you didn’t quite live through it if you were born in 1860 and left around 1885. All of a sudden, you’ve been running your very prim and proper boarding school, trying to have everyone be buttoned up and tight-laced. Then you have this charismatic, ecstatic explosion.

It just goes to show you don’t really know what’s happening — you’re just along for the ride. Ee ja nai ka, the Japanese phenomenon of spontaneous dancing in the street that happened towards the start of the Meiji Restoration, is the one thing this reminds me of.

But maybe we can layer in Christian revival with Japanese attitudes toward this. They’re pushing Shintoism and ultimately get a little freaked out by all this Christian energy. How do those two strands interact in this period?

Jonathan Cheng: This period — right around when Kim Il-sung was born in 1912 — is also when Korean nationalism started to rise as a response to Japanese imperialism. Japan saw Korea as a foothold to the Asian mainland, and as they asserted control, Korean nationalism was born, with much of it incubating inside the Christian church.

Usually, Christianity accompanied Western imperialist gunboats. But in Korea, the Japanese were the imperialists bringing Shintoism. So Christianity became a patriotic bulwark and a vessel for nationalist hopes. The Japanese shut down almost all civil society in Korea, but they left a carve-out for the Christian church. They wanted to prove to Washington, London, and Paris that they were an “enlightened” colonial power, which meant not crushing the West’s most cherished institution.

Kim Il-sung’s father, Kim Hyŏng-jik — who now has his own cult of personality in North Korea — was a devout Christian nationalist. There’s a scene in my book where he performs a blood ritual at the tomb of Ki-ja, a Chinese nobleman who was seen as the progenitor of the Korean nation. He and his allies pierced their skin, wrote vows to fight for independence, and prayed explicitly in the name of Jesus Christ. They believed their fight against Japanese imperialism would happen under the banner of Christ.

The Japanese knew the church was an antagonistic vehicle for nationalism, but couldn’t crush it without angering the West. Meanwhile, families like the Kims flocked to the church. I believe it was a genuine embrace of the faith, but it also dovetailed perfectly with their political goals. Uniquely, the church provided a nationwide network where like-minded people could gather and organize the fight for independence.

This is the dynamic you have here on the Korean Peninsula at the beginning of the 20th century. The Japanese colonial government knows that the church is perhaps this antagonistic vehicle

Kim Il-sung’s Christian Upbringing

Jordan Schneider: Kim Il-sung is born into this context. What was his childhood like? Was he raised Christian? Was he raised as a freedom fighter?

Jonathan Cheng: Kim Il-sung was definitely raised as a Christian and a freedom fighter, though his family wasn’t socialist; they were conservative Christian nationalists.

His father raised him in the church, but his mother, Kang Pan-sŏk, was perhaps an even bigger influence. She was the daughter of a church elder and served as a “Bible woman” — a role created to give women leadership within the Presbyterian Church. At a time when Korean women were almost universally illiterate and often not even given formal names at birth, she traveled from village to village teaching women to read using the Bible. (Her name, “rock,” was actually a baptismal name, like Peter, the rock upon which the church was built).

Kim Il-sung was effectively a third-generation Christian. It’s fraught to label him a believing Christian later in life, and while there’s unverified evidence he was baptized, there’s no question he was deeply steeped in the culture. He attended church every Sunday, taught Sunday school, learned to play the church organ, and acted in church plays. As a teenager, he even lived in a pastor’s home and attended a missionary school started by his grandfather and Samuel Moffett. He knew his Bible intimately and quoted scripture to impress listeners for the rest of his life.

So yes, he grew up in a Christian bubble, and his father seamlessly fused that faith with militant anti-Japanese nationalism.

Jordan Schneider: Let’s layer in the Soviet-ness of it all. How does communism end up appealing to him? Maybe we could do a little historiography of how he ends up backdating a lot of this stuff. When thinking about the different strains of influences, maybe do a compare and contrast while we’re still thinking about the early years.

Jonathan Cheng: He’s born in 1912, and the big seminal moment in Korean nationalism, in modern Korean history, is March 1st, 1919. That is the July 4th, 1776 of Korea. That is the day they unveiled their Declaration of Independence. They had 33 signers. These 33 signers of the Declaration of Independence are now the pantheon of Korean nationalists, and half of them are Christian, most of them Christian pastors.

It’s written in very lofty language, as declarations of independence perhaps should be written. It invokes explicitly the will of God, and a lot of Americans who read this document talk about how it taps into America’s own Declaration of Independence — that lofty language. Kim Il-sung is six years old.

The other point to make here is that because it’s 1919, this is right at the end of World War I. This is right when Woodrow Wilson, who I will note was the grandson, son, and nephew of Presbyterian ministers, is in Paris. He is trying to negotiate and trying to win this principle that he’s made such a big part of his presidency, such a big part of his image — these 14 Points, this call for self-determination of all nations. He’s making this call in the context of his own Christian upbringing and his own Christian principles as the President of the United States.

We know what happens to this Wilsonian promise because after 1918, after 1919, not all small nations and peoples are given self-determination. Colonialism and imperialism still have a very strong hold. It turns out that the UK, the French, the Italians, many of these great powers aren’t so willing to give up their colonies. Japan also says, “Well, why do we need to give up ours?”

You have the collapse of this Wilsonian idealism that’s very much rooted in these Christian principles, and you have many people like Kim Il-sung’s own father — so many Korean Christian nationalists who had put their hopes in the Wilsonian basket — find that they need to find another basket.

Many of them look to Moscow, to what happened there in 1917, just two years prior, where you have Vladimir Lenin succeeding with his calls for self-determination. There’s a nice historical footnote here that this idea of self-determination for all nations really originates with Lenin before Wilson picks it up. When Wilson’s promise implodes, many of these nationalists around the world — not just in Korea but all over the world, including in India, including in the Eastern European states, the Balkans — turn their eyes instead to Lenin and to the USSR that is beginning to emerge.

Alex Boyd: Every American is like a temporarily embarrassed millionaire, right? People in colonized countries in the 1990s were grasping for technologies, all temporarily embarrassed people’s leaders. But what’s really key about Kim Il-sung is not only that he was born in this Christian tradition, not only that he was looking to the Soviet Union, not only that he a nationalist, but he’s also in China.

He spent most of his time in Jilin and in occupied Manchuria. What does he learn there? What does he learn from China as he witnesses the disintegration of a past social order and the imposition of a Western social order by Japan, which is revolutionary enough in itself?

The Korean and Chinese Independence Movements

Jonathan Cheng: That’s a great question. March 1st, 1919, comes before one big seminal moment in Chinese modern history — the May 4th Movement of 1919 that begins shortly thereafter and is inspired in many ways by what happens in Korea.

You have the assassination by a Korean Catholic convert of Itō Hirobumi, the four-time prime minister of Japan and the first resident general of Korea, when Japan makes Korea a protectorate. That shooting happens in Harbin at the railway station, but the trial happens in Dalian. We have Liang Qichao, one of the great late Qing reformers, who’s there in the audience, in the gallery for the trial — very much a fan of this Korean assassin.

You see those resonances between the two independence movements. Then you have the provisional government of Korea set up in Shanghai, in the French concession, in 1919, just around the corner from what is today the site of the first meeting of the Chinese Communist Party, which happens in 1921. They’re literally just a stone’s throw from each other, and both of those buildings you can now visit in the Xintiandi complex in Shanghai today.

So there are these two independence movements with great parallels between them. Many Korean socialists who begin as Korean Christian nationalists abandoned Christianity or abandoned the US.-led approach to seeking independence and embraced socialism and Bolshevism. Many of them end up going with Mao to Yan’an as well.

You have a lot of resonance. Part of Mao’s reason for getting involved in the Korean War is also partly because of this affinity, fraternity between the Korean and Chinese independence movements.

Kim Il-sung spends a lot of his adolescence, a lot of his teenage years in Manchuria, specifically in Jilin Province, and very specifically in Jilin City.

His Christian upbringing continues to resonate because he lives in the home of a Korean Methodist pastor who is one of the leaders of the Korean independence movement. This man, Son Jeong-do, was a Methodist pastor, but he was also the first speaker of the provisional Korean legislature. He’s a fascinating guy. Kim Il-sung, in his memoirs, calls him “the savior of my life.” He uses this language very specifically and identifies himself explicitly as a pastor. He goes out of his way to say he was a Methodist pastor, and his patriotism — his love for Korea — was not something that was separate from his Christian faith. They were one and the same.

He has two boys. One of his boys later becomes the founder of the South Korean Navy and the defense minister of South Korea. Kim Il-sung is growing up living in the home of a Korean pastor in Jilin City, sharing the same roof with the man who is going to be his antagonist in the Korean War. The younger brother in this household later becomes a doctor in Nebraska who takes the opposite side of the trade and ultimately embraces Kim Il-sung, looking up to him almost in a worshipful way. He ends up embracing Kim Il-sung in Pyongyang at the end of their lives in the 1990s and is buried in the Patriotic Martyrs’ Cemetery outside of Pyongyang.

You have this incredible period where you have this young, newly orphaned Kim Il-sung. He’s just a teenager. He’s living in the home of a Korean Methodist pastor. I should note that he’s literally living in a church because the pastor’s annex, the residence, is literally just a part of the church. He picks up a lot. He learns a lot about how revolution is done here, about how nationalism is done. He does end up drifting away from the church during this period and drifting towards Bolshevism, socialism, eventually what you would call Stalinism.

Share

Jordan Schneider: Mao was 55 in 1949 with the founding of the PRC, but Kim Il-sung was only 36 when the DPRK was founded in 1948. Let’s tell that story. How does he end up being the one that the Soviets choose to be the figurehead? And how does he go from figurehead to obviously something a whole lot more than that?

Jonathan Cheng: The Soviets came roaring down from Siberia through Manchuria after the atomic bombs fell on Hiroshima and Nagasaki. The Japanese surrendered much faster than the Americans had expected, even though it was the Americans who dropped the bombs. These were fateful days — Stalin moved so quickly into the Korean Peninsula that the Americans had to scramble to make a deal, asking the Soviets to stop at the 38th parallel, causing Seoul to fall in the US zone and Pyongyang in the Soviet one.

The Soviets needed to establish an administration quickly. They knew they needed a Korean at the head — a Russian leader of North Korea wouldn’t work. Their first choice was the most popular political figure in the northern half of the Korean Peninsula: Cho Man-sik, a conservative Christian church elder decades older than Kim Il-sung. He was a genuinely popular, charismatic figure known as the “Gandhi of Korea” because he wore traditional Korean robes and advocated for nonviolence.

The problem with Cho was his stubbornness. He didn’t want Soviet-style rule — he wanted American-style governance. He advocated for capitalism, electoral democracy, freedom of the press, freedom of assembly, and other American ideals. The Soviets removed him and needed someone else.

They knew they needed someone Korean, but they didn’t trust the Koreans who had spent significant time in Moscow, partly because of their close ties to the Communist International, which Stalin distrusted.

Why did Stalin pick Kim Il-sung? We’re not entirely sure, but he was certainly the right man at the right time. He was young and genuinely charismatic. He spoke excellent Chinese because he’d grown up primarily in China, and he spoke some Russian. He was pliable, and he identified as a socialist, though he never finished high school. His grasp of Marxism-Leninism or dialectical materialism probably wasn’t very deep, but he was definitely on that side of the ideological ledger. As a soldier in the Red Army, he was perfectly positioned.

Initially, the Soviets considered making him deputy head of police in Pyongyang or perhaps defense minister. Neither outcome materialized because they needed someone new, and Kim elbowed his way past potential rivals. Certainly by the beginning of 1946, when he was 33 years old, there was no question that he was first among equals in Soviet eyes. He wasn’t yet leader of what we’d call North Korea because the North Korean state didn’t exist for a few more years, but he was effectively the leader of North Korea by 1946.

How do we know this? Songs were composed in his honor. Kim Il-sung University was named after him that year. Books began praising “General Kim Il-sung” in terms that went beyond mere appreciation — they portrayed him as a messianic figure. This over-the-top language began very early in Kim Il-sung’s tenure as leader of what would become North Korea.

Kim Il-sung as Messiah

Jordan Schneider: Can we talk about this eureka moment where he hears this peasant getting strung along by a colleague who’s talking about him like he’s a messiah? Then he says, “You know what, it’s fine. If they want to think I can walk on water, let them think I can walk on water.” The quote was, “If people believe in a legend, we must say the legend is true. Why deny these facts and discourage him?” And then he writes, “I made up my mind to live up to the people’s expectations and trust.”

Jonathan Cheng: That is one of my favorite moments in the whole book, and I’m glad you’re highlighting it, Jordan, because that is kind of the key to the whole mythology around Kim Il-sung.

The bit that you quoted there is from the eight-volume memoir, With the Century: Reminiscences, published in his name. Most North Korea scholars for decades have basically dismissed it, because why would you take seriously something that is so full of fictions? He describes all of these battles and all these little skirmishes that he fights against the Japanese, very few of which are corroborated by independent historical documentation.

But there is this moment very early on in this eight-volume memoir where he is overhearing a simple peasant and his belief that this Kim Il-sung figure can teleport, can effectively walk on water, if not literally. He is play-acting as Messiah — that’s the phrase I use in the book. He pulls back the curtain a little bit. There’s a bit of removing the fourth wall here: “Let me show you a little bit about how I decided to do this.”

We probably even need to take this admission, this acknowledgment, with a grain of salt, because there are layers and layers here of meta-reading that need to be done. But the very fact that he would acknowledge at all that he was aware of the self-mythologizing that went into being Kim Il-sung is pretty remarkable. This incident happens during these years that we actually know very little about Kim Il-sung otherwise, before his public appearance in Pyongyang. But he tells us that this is how you build this legend, how you build this mythology.

Alex Boyd: That reminds me a little bit of the early Chinese communist leader Li Lisan, who did the same thing at the Great Strike of Anyuan — one of the first Communist Party achievements in China in the 1920s. It involved Mao, Liu Shaoqi, and Li Lisan.

Li Lisan was like, “Yeah, no, I am invulnerable to bullets. Don’t worry about it.” A lot of the peasants were mixing up Christian ideology — because they’d been exposed to that — and communist ideology. They had a lot of trouble distinguishing the hammer and sickle versus the cross. There were also these secret brotherhood societies, and Liu Shaoqi took a very dim view of that. Obviously, there’s real power in attaching Soviet ideology to more charismatic local traditions and the Christian idea of salvation through a savior.

Jonathan Cheng: There’s syncretism happening at all these levels. All these different ideologies — whether it’s Christianity, socialism, or Marxism-Leninism — blend with local conditions and mores to become something brand new.

What happens in Manchuria isn’t entirely Korean because it’s occurring in this liminal space, this Wild West that Manchuria was at the time. It’s very lawless. We know from our knowledge of Republican China that they didn’t really have much of a grasp on what was happening in northeastern China. During this pretty lawless period, Manchuria was a place without much government.

You have Zhang Zuolin — they called him the “Mukden Tiger” because he was the warlord up in Shenyang — and later his son Zhang Xueliang. All sorts of carpetbaggers, desperados, and opportunists were flowing into this region. The Republican Chinese government had no real control. The warlords were fighting for control. You had the Bolsheviks, the White Russians who had fled from the new USSR. This was like a big vacuum. You had the Japanese making inroads, and all the Koreans who had fled the Korean Peninsula.

Into this crazy cauldron of influences, you have Kim Il-sung coming of age and being exposed to all these different ideologies, while also building up his own mythology. It’s an ideal place to do it because it’s this frontier land where anything goes, and he embraces that ethos. He’s all into it.

Out of this, he’s able to create this legend of Kim Il-sung — because that’s not even his birth name. His birth name is Kim Song-ju (김성주 / 金聖柱), which, as I detail in the book, doesn’t mean what he later claims it to mean. It’s not jin cheng ju (金成柱) in Mandarin, “to become a pillar.” It is jin sheng ju (金聖柱), which means “a holy pillar, a pillar of holiness.” Sheng, as in holy. That was the name he was given because he was raised in this Christian family and his father wanted to give him a Christian name.

Kim Song-ju is his birth name in Korean, but he’s later reborn as this Kim Il-sung figure, adopting the legendary name of this almost Robin Hood-type figure able to vanquish the Japanese and perform all sorts of extraordinary feats out there in the no man’s land of Manchuria.

The Origins of the Cult

Jordan Schneider: How does this cult start? Where do the Christian elements begin to get woven in?

Jonathan Cheng: Kim Il-sung was 33 when he came to power, and 36 by the time the DPRK was founded. From the start, his personality cult took on a life of its own. Stalin helped establish it — he had many “little Stalins” across the Eastern Bloc — but North Korea quickly moved beyond pure Stalinism.

Very early on, diplomats from Socialist Bloc countries posted to Pyongyang began sending cables back home. These translated cables reveal their concerns: they reported that what they were witnessing wasn’t orthodox Marxism-Leninism — it had morphed into something different. Many drew connections to religious worship, describing it as grotesque, unorthodox, and entirely separate from socialist principles.

1956 marked a crucial turning point. At the 20th Congress of the Soviet Communist Party, Nikita Khrushchev delivered his secret speech denouncing Stalin’s cult of personality and calling for collective leadership.

This presented a fork in the road for Kim Il-sung, who had been in power for a decade. He had launched the Korean War and narrowly avoided disaster—without Mao’s intervention,

This presented a fork in the road for Kim Il-sung, who had been in power for a decade. He had launched the Korean War and narrowly avoided disaster—without Mao’s intervention, he would likely have been relegated to the dustbin of history. Yet he turned the war into a personal victory, dumping blame on his lieutenants and emerging not just as first among equals, but approaching levels of deification.

The only real challenge to Kim’s power came in August 1956, right after Khrushchev’s speech. Opponents accused Kim of building a cult of personality and sought to remove him. But Kim successfully outmaneuvered them, fending off the coup attempt.

After that, Kim faced no meaningful challenges until his death in 1994 — nearly four decades of unchallenged rule. Facing no real opposition, he felt no need to enact reforms. Unlike the Soviet Union’s later glasnost and perestroika, Kim maintained total control. Despite the DPRK being smaller than China or the USSR, he proved remarkably savvy at playing Beijing and Moscow against each other to secure his unique brand of totalitarian rule.

The three Kim leaders have been absolute geniuses at playing a relatively weak hand extraordinarily well, giving them the cover to run the country exactly as they want.

I argue in the book that Kim Il-sung recognized the value of the Christianity he was steeped in, borrowing its doctrine and ritual to weave into what became Kim Il-sungism. Experts often refer to Juche (주체) — or “self-reliance” — but North Korea makes clear that Juche is merely a subset of that larger, overarching ideology.

This became a runaway train without brakes. By 1946, when Kim was just 34, his cult of personality had already reached extraordinary heights. There were epic poems, books, and songs praising him, and a university bore his name. Statues began appearing in the late ’40s—now there are tens of thousands.

Kim Il-sung study centers, heavily compared to churches, were built for mandatory weekly gatherings. Religious-like rituals emerged — portraits required daily dusting, and eventually, every North Korean had to wear a pin of Kim Il-sung (and later Kim Jong-il) over their heart.

Eventually, North Korea’s socialist allies declared this was insane. They said it resembled a religion, not communism. I don’t believe this was accidental — it was by design.

Whether this design was conscious or subconscious is unclear, but I’d guess subconscious. Kim likely grew up understanding how faith and power intersect, observing how churches in the “Jerusalem of the East” built political power under Japanese colonialism. He understood that religionizing politics could be highly effective. That’s essentially what he did, and the momentum never stopped.

Jordan Schneider: Caesar wept at 32 because he hadn’t conquered Asia. Should I be weeping at 36 because there are no statues of me? No one’s polishing Jordan Schneider China Talk icons in their living room.

Jonathan Cheng: It’s never too late, Jordan.

Jordan Schneider: What’s wild is that because he’s so young, it can last for so long. You have generations that are born into it completely.

You tell this fascinating story about how in the early years he has to navigate the fact that there are actually Christians — not a few of them, but real communities. It’s like the best civil society you have in the North that he’s kind of dancing around trying to co-opt. He eventually gets some of these preachers to buy into his cult of personality, as you alluded to earlier.

But you go far enough into the future, and he’s dropping his own Ten Commandments. Once you’re into the ’70s and ’80s, it becomes so alienated from the original Father, Son, Holy Ghost that when people leave the country and try to process what happened, they get thrown a Bible and someone tells them, “Just think of Kim Il-sung as the Father and Kim Jong-il as Jesus.” Then it kind of all makes sense.

On this conscious versus unconscious piece — you conclude on this a little bit in the book — to what extent is this just human nature? Some weird twisted Lord of the Flies thing where everything reverts to emperor god worship if you can run a country for that long? Would anyone have ended up there anyway, or are there particular Jesus-y bits that stuck out?

Jonathan Cheng: First, you make a great point about his longevity. Ruling for 49 years after taking power at 33 is an incredibly long time to put a stamp on a country.

Mao ruled from 1949 to 1976, but his reputation took a beating during the Great Leap Forward. Stalin didn’t fully consolidate power until 1945 and died just eight years later. But Kim Il-sung had half a century.

He learned from Khrushchev’s de-Stalinization and later from Deng Xiaoping’s efforts to de-Mao-ify China. Kim didn’t want his legacy dismantled. He wanted a successor who would preserve the system exactly as he built it, so he turned to his son, Kim Jong-il, who in turn passed it to Kim Jong-un.

We’re now 81 years into this dynasty. Almost all 25 million North Koreans were born into this reality, and the number of people who remember life before the Kims is diminishing by the day.

Kim recognized that cultivating absolute devotion required building what is essentially a religion. It was likely piecemeal rather than a master blueprint from day one.

His son Kim Jong-il — who understood the power of iconography and film—played a massive role in elevating the cult. He introduced the mandatory badge pins, proliferated the statues, and created revolutionary operas and movies that took worship of his father to the next level. It really takes a village to build a cult of personality this extraordinary.

Kim Jong-un, the third generation heir, has defied expectations. In politics or business, third generations often get decadent and coast on past laurels. Yet 15 years in, Kim Jong-un has shown he knows exactly what he’s doing.

The Jonestown Connection

Jordan Schneider: You mentioned cult and family — I have to bring this up. Kim Il-sungism was really interested in expanding its reach globally in the 1970s. You could read the New York Times and there’d be a full-page ad extolling the wisdom of Kim Il-sung. People noticed, specifically Jim Jones. I thought that was one of the most fascinating parts of this book. How has the global world understood Kim Il-sungism, especially those who admired it like Jim Jones? Explain who Jim Jones is.

Jonathan Cheng: Jim Jones was actually from Indiana, just like Samuel Moffett, the missionary who brought Presbyterianism to Pyongyang. Jones was a charismatic, Christian-adjacent preacher who founded the People’s Temple in San Francisco. From the beginning, they were dogged by allegations of being a cult.

I should pause to unpack that word. I prefer “cult of personality” because it’s easily definable. “Cult” is a loaded, pejorative term — casually defined, it’s just a religion the speaker deems illegitimate. People join movements; they only call it a cult later if they leave.

Because they felt persecuted, Jones moved his followers to the jungles of Guyana. There, he had total dominion over about 1,000 people. They were cut off from the outside world, lived under heavy surveillance, and used an in-group vocabulary. Jones demanded everyone call him ‘Father’ and imposed strict sexual mores that he exempted himself from.

The Korea connection is fascinating — Jones adopted three Korean children and was a massive fan of Kim Il-sung. He constantly praised Kim during his marathon sermons and even mused about moving Jonestown to North Korea. His leadership team actually met with the North Korean embassy in Guyana to trade tips on running a closed religious society.

When US federal agents investigated the compound, Jones famously ordered his followers to commit revolutionary suicide by drinking poisoned Kool-Aid. Hundreds died. This is one of the more traumatic, dramatic images of the 1970s — to see hundreds and hundreds of Jonestown members, followers of Jim Jones, splayed out on the ground because they’d consumed this poisoned Kool-Aid. They followed him into the eternal Jonestown in the sky, as it were.

My favorite coda — the People’s Temple building in San Francisco was later bought by a Korean Presbyterian church.

There was a genuine meeting of the minds between Jones and Kim Il-sung. Both understood the human need to believe in and follow someone. If you put yourself at the center of that faith, you can make people do almost anything — from committing mass suicide to acting as human detonators on the front lines of the Russia-Ukraine war. It’s less about fighting for a country and more about a religious fervor for their leader.

Alex Boyd: You’ve actually been to church in Pyongyang and sat in a service. What was that like? Did it feel like Jonestown, or did it feel more like a normal religious service?

Jonathan Cheng: It felt a lot like walking into any Presbyterian church in South Korea, which is no accident. It has pews, a pulpit, and a choir in traditional robes.

What it doesn’t have are images of the Kim family on the walls or congregants wearing Kim Il-sung badges over their hearts. It’s a very different sort of space.

I visited in 2017. The sermon wasn’t about following Kim Il-sung — it was about Jesus and God, quoting directly from a standard 66-book Korean Bible. The Bibles and hymnals in the pews were actually donated by South Korean Presbyterians. It’s a very strange place.

Alex Boyd: You mentioned they’re not wearing Kim Jong-un pins. Are they taking them off upon entering the church in the same way that you might take off your shoes before entering a mosque?

Jonathan Cheng: I asked the pastor about the pins, and he simply said, “Well, we’re Christian. We don’t need to wear them.” Because you’re on such a short leash as a journalist in North Korea, I couldn’t follow congregants home to investigate how that worked in practice.

To me, that statement implicitly acknowledges that Christianity and Kim Il-sungism are, to a degree, rival faiths. Kim Il-sung had a genuine fondness for his Christian upbringing and his mother, and I think he created a carve-out for genuine worship in these two Pyongyang churches — provided the congregants are vetted, loyal, and heavily monitored. Many are quite old, remembering a time before Kim, and are seen as relatively harmless.

Towards the end of his life, North Korean sources say Kim’s mother appeared to him in a dream. He asked a lieutenant what he should do, and the advice was to build a church on her birthplace. That’s one of the two Protestant churches there today.

When I visited in 2017 for the Wall Street Journal, the government asked what we wanted to see, and I requested the churches. It didn’t fully answer whether they are ‘real,’ but my best guess is yes — within very prescribed parameters. Not just anyone can waltz in, and there isn’t true religious freedom, but it goes beyond a mere Potemkin village.

The 768-Page Rabbit Hole

Jordan Schneider: How do people take to this idea — from South Korean Christians to your parents?

Jonathan Cheng: It’s mostly been surprise and shock in a positive sense. Scholars like Bruce Cumings and many Korean Americans have told me they had no idea Kim Il-sung was raised in the church.

There have been two main criticisms. First, some Christian readers worried I was suggesting Christianity inevitably leads to a Kim-style cult. I’m not. This is narrative history, not a polemic. Every ideology can be used for good or evil. I’m simply connecting the dots between Kim’s upbringing and the state he built.

The second criticism is that I foregrounded Christianity too much, excluding factors like Confucianism or feudalism. My response is that those other inputs are well-documented, whereas Christianity’s role in shaping Kim Il-sungism has been entirely ignored. Kim’s family identified as Korean nationalists inspired by Jesus Christ, not by shamanism or Japanese emperor worship. To ignore those first twenty years of his life is a massive oversight.

Jordan Schneider: We have to ask one Taiping Rebellion question. What are your thoughts on comparing the two stories?

Jonathan Cheng: The Taiping Rebellion is a fascinating parallel. Hong Xiuquan failed his civil exams, read a Christian missionary tract, and remixed it to create the Taiping Heavenly Kingdom. It governed half of China and led to a disastrous civil war with 20 million dead.

It highlights a clear pattern — Western teachings morphing in unexpected ways when they reach Asia, resulting in massive consequences.

Jordan Schneider: Do you see any Jesus in Xi?

Jonathan Cheng: That’s a provocative question. Do I see messianic claims in Xi? Yes, there are echoes. The rhetoric around the “rejuvenation of the Chinese nation” and needing a savior to lead an embattled nation to the promised land definitely has a messianic flavor. It pales in comparison to Pyongyang, but the elements are there.

Alex Boyd: Xi kept Kim Jong-un at arm’s length for a while, but last year, he came to the military parade. What’s the state of play right now?

Jonathan Cheng: All indications suggest Xi Jinping isn’t terribly fond of the Kim family, but there’s a grudging acceptance. The Kims have been masterful at playing Russia and China against each other. China isn’t thrilled about North Korea’s nuclear arsenal, but in the grand scheme, a nuclear North Korea causes more headaches for the West than for Beijing. So, they tolerate it, and occasionally find it advantageous.

Jordan Schneider: Let’s close on the whole amateur historian arc. You’ve spoken about how this has taken a lot from you spiritually over the past decade. Any reflections for listeners considering biting off some piece of 20th-century history?

Jonathan Cheng: I genuinely didn’t set out to write a 768-page book. I just wanted to read it, but it didn’t exist. I initially thought it would be a 2,000-word weekend piece for the Journal, but it became the ultimate rabbit hole. I wouldn’t do it again unless a topic grabbed me by the lapels the same way. It took over twelve years and more man-hours than anything else I’ve ever attempted.

I studied history in college — this was my path not taken. Because I’m a journalist, my goal was to write for a broad audience rather than produce a dense academic dissertation. I hope it inspires others with better Korean language skills to take the baton and dig even deeper.

Jordan Schneider: Any closing thoughts on China and Korea? It was interesting watching you navigate through the propaganda. What’s the line about “munching rhinoceros sausage”? You’ve spent so much time with North Korean texts. What’s your main takeaway?

Jonathan Cheng: Reading North Korean texts requires a lot of meta-reading to understand what they’re really trying to say, but I actually had great fun doing it.

My ultimate takeaway is that North Korea is less a nation-state and more a religious society. Religion can be harnessed and weaponized for political power. Sam Altman once posted a quote noting that successful people create companies, very successful people create countries, and the most successful create religions. Kim Il-sung understood that he founded both a country and a religion — and it’s the religion that has allowed the state to endure against all odds.

In 1991, with the fall of the Soviet Union, or in 1994 during the famine and Kim’s death, few would have bet on North Korea surviving to 2026. Yet here it is, entering its third or fourth generation, armed with nuclear weapons. It needs to be taken seriously and understood not as an incomprehensible ‘other,’ but as a society whose ideological heritage traces back, surprisingly, to American Presbyterianism.

To discuss why Kim Il-sung kind of thought he was Jesus — or at least wanted to convince everyone else he was — we have the honor of Jonathan joining ChinaTalk today, with Alex Boyd.

Our conversation covers:

• Billy Graham meets Kim Il-sung — how America’s pastor visited Pyongyang in 1992 and bonded with the dictator over their shared Presbyterian childhoods, while touring a birthplace resembling the nativity story.

• From “Jerusalem of the East” to totalitarian theocracy — how 22-year-old missionary Samuel Moffett brought eschatological founder energy to the wickedest city in Korea, and transformed Pyongyang into the most Protestant city in Asia, complete with jazz clubs and a Chevrolet dealership.

• The Manchurian cauldron — how the lawless frontier of 1920s Manchuria blended Christianity, Korean nationalism, and Marxism to forge Kim Il-sung’s messiah complex — and why the 33-year-old decided to let peasants believe he could walk on water.

• Jim Jones, superfan — how the Jonestown cult leader became obsessed with Kim Il-sung, adopted Korean children, and traded tips with North Korean diplomats on running a closed religious society.

• Christianity as political technology — how Kim weaponized Protestant doctrine and ritual to build a state religion that has outlasted the Soviet Union by three decades.

Plus, why North Korea is best understood not as a nation-state but as a religious society, and what Hong Xiuquan and Xi Jinping have to do with any of this.

Learn more about your ad choices. Visit megaphone.fm/adchoices

击鼓传花 (ji gu chuan hua), roughly translated as “beat the drum, pass the flower,” is a Chinese children’s game in which players pass a flower around while a drum sounds. When the drum stops, whoever is left holding the flower loses the round.

The phrase has recently acquired a secondary use in Chinese financial and policy commentary as a shorthand for a speculative chain in which risk is passed between participants. From anonymous investor forums to listed-company risk disclosures, it has been applied to the 2015 A-share bubble, property speculation, local government debt rollovers, P2P lending, and, most recently, the venture capital (VC) market. As one Chinese venture capitalist explained to me,1 the game is China’s AI market, and investors are the players. Each investor races to pass along overvalued AI companies before the “beat” of the capital-raising cycle stops.

Many domestic financial actors caution that this dynamic indicates China’s early-stage AI market is becoming a financial bubble. Valuations have inflated faster than revenues, capital is concentrated in a handful of hot subsectors, and money keeps flowing toward companies with no clear path to profitability. They cite the roughly 80,000 Chinese AI companies that have dissolved or deregistered in the past two years.

While they may be right that these are indicators of a market about to correct itself, it is more likely that they are signs that Chinese industrial policy is at work. Since 2017, when AI was designated a strategic technology, Beijing has taken an active role in promoting the industry’s growth. Over the past eighteen months, the state has prioritized funding early-stage companies, launching dedicated national AI funds and committing hundreds of billions of yuan toward AI ventures. This influx of concentrated capital has contributed to bubble-like dynamics as funding rushes in faster than the industry can absorb it.

Yet the strategy of financing new ventures in strategic sectors is one Beijing has used before. From solar panels to electric vehicles (EVs), the government used state capital to build other priority industries. Whether AI follows a similar trajectory depends on how the funding is deployed, including the mechanisms and stakeholders involved.

The Rise of Policy-Driven Funding

The origins of the government venture funding architecture stem from the history of China’s AI financing ecosystem. Over the past decade, shifting domestic and international policies have shaped the sources of venture capital and the structure through which it is deployed.

Before 2018, foreign venture capital played a major role in Chinese tech financing. Firms like IDG, Goldman Sachs, and Peninsula Capital funded the first wave of internet giants, such as Alibaba, Tencent, and Baidu. These firms brought both large-scale capital and a Western investment philosophy of high-risk, staged investment, which shaped what got funded and how within the tech startup world. However, driven by rising US-China tensions, the share of Chinese VC deal value involving foreign investors fell from 54.8% in 2018 to roughly 20% in 2023. President Biden’s August 2023 executive order (EO) on outbound investment, which prohibited or required notification for certain US investments in Chinese AI, semiconductor, and quantum firms, accelerated this process. Although the EO didn’t take effect until January 2025, many international firms split off their China entities, downsized their domestic presence, or withdrew from China altogether in anticipation of its enforcement. Dollar-denominated funds, which accounted for nearly 14% of China VC fundraising in 2021, accounted for barely 1% by early 2025.

Domestically, private capital actors were in no position to fill the gap. As foreign funding retreated, a wave of new regulations hit the Chinese tech sector, stymying the industry’s main sources of funding and exit avenues. The 2018 asset management rules restricted the underlying wealth management structures that supplied the largest source of domestic private funding. This was exacerbated by Beijing’s 2021 tech crackdown, beginning with regulators fining Alibaba a record 18.2 billion yuan ($2.8 billion) for anticompetitive behavior. The campaign subsequently swept up Tencent, Meituan, and many of the other Chinese consumer internet champions. Beijing then launched a cybersecurity probe into ride-hailing giant DiDi as it pursued a New York IPO, leading regulators to pull Didi apps from app stores and forcing the company to delist. On the US side, the Holding Foreign Companies Accountable Act threatened to delist Chinese companies due to disputes over auditing standards. This made US listings, once the clear objective of most startups, no longer a viable option. By 2024, venture capital fundraising sank to its lowest level in almost a decade.

With the retreat of private funding, state capital came to dominate early-stage funding. The same year private VC was at its lowest, state-owned capital accounted for 82% of all new LP contributions. Since then, state funding has continued to pour into the venture capital ecosystem, especially within AI. As AI has grown in importance, the government has expanded its breadth of funding to various stages and types of AI companies, with state investors leading some of the sector’s largest financing rounds.

Government Funding’s Role in the AI Industry

As Beijing increases its investment in the AI industry, government priorities and funding mechanisms are playing a large role in shaping the industry. The set of stakeholders involved in the government financing apparatus, the terms attached to the funding, and how this money is directed and allocated now disproportionately determine what gets built.

Currently, much of this capital has been deployed via government guidance funds (GGFs), vehicles set up to channel state capital in line with industrial goals, including promoting “strategic emerging technologies” (战略性新兴产业 zhan lue xing xin xing chan ye), advancing industrial upgrading, and reducing reliance on foreign suppliers in critical sectors. GGFs have played a crucial role in incubating “leapfrog technologies” such as EVs and renewables, 5G and telecom, and now, semiconductors and AI. Over the past two decades, GGFs have amassed 7.7 trillion yuan ($1.1 trillion) in committed capital,2 roughly twenty times what the US government’s public-private investment program, the Small Business Investment Company (SBIC), held at its record 2025 high, and enough to fund DARPA at its current budget for two centuries.

Between 2000 and 2023, nearly a quarter of GGF capital flowed to AI-related firms. Within the AI supply chain, the National Integrated Circuit Industry Investment Fund (the “Big Fund”)3 has been instrumental in shaping the semiconductor industry. Its first 138.7 billion yuan ($21.8 billion) anchored national champions along the chip-making supply chain from downstream chipmakers like foundry leader SMIC and memory maker YMTC to upstream equipment and materials suppliers like NAURA and Piotech. The fund was renewed several times, first in 2019 with 204 billion yuan ($29 billion) after China’s AlphaGo moment, and again in 2024 with 344 billion yuan ($47.5 billion) after ChatGPT’s launch to further increase AI chip and chipmaking equipment production.4

Chinese government spending was further accelerated by US attempts to curb the industry. In January 2025, days after Washington tightened semiconductor export controls, the Ministry of Industry and Information Technology (MIIT) and the Ministry of Finance (MOF) established a dedicated 60 billion yuan ($8.2 billion) National AI Industry Investment Fund.5 Unlike the Big Fund, this fund explicitly targeted early-stage ventures across the AI supply chain from upstream inputs, such as compute and data, to downstream applications, especially within embodied intelligence and autonomous vehicles.

Less than a year later, the National Development and Reform Commission (NDRC) and MOF launched another 100 billion yuan ($14 billion) national venture capital guidance fund. This fund set up three regional sub-funds (Beijing-Tianjin-Hebei, Yangtze River Delta, and the Greater Bay Area) as innovation centers registered at roughly 30 to 47 billion yuan ($4.4 billion to $6.9 billion) each. Since their launch, the sub-funds have signed over 25 deals across integrated circuits, quantum, biopharma, AI, and other priority sectors.

GGFs have been criticized for creating the bubble dynamics experienced by industry and being “poorly conceived and implemented.” In some ways, this is true. Yet the dynamics that look like a financial bubble at a surface level are the same mechanisms that allow Beijing to achieve its industrial policy goals.

Not-so-patient Capital

Early-stage companies, especially in capital-intensive fields like semiconductors, robotics, and foundation models, often require years of research and capital before they turn a profit. GGFs are meant to supply this source of “patient capital”6 within hard technology sectors (硬科技 ying ke ji). While these funds nominally have 15 to 20 year time horizons, in practice, their effective horizons are usually much shorter.

This timeline compression is driven by the process through which GGFs are often deployed. National GGF managers rarely invest directly. Instead, they often delegate decision-making downward7 to various regional levels and/or private VCs to bring in industry expertise and “amplify the leveraging effect” of government capital. Central capital is thus passed through a chain of investors before reaching companies. The 2025 fund, for example, operates through a three-tier structure where the national guidance fund company seeds regional funds, which in turn seed sub-funds typically run by private VCs.

As capital flows through the system, each actor attaches their own, often short-term, objectives. Local officials, among the most active stakeholders in this process, operate on cadre rotation cycles (干部交流 gan bu jiao liu), in which they switch departments and geographic posts every three to five years. They therefore optimize their funding decisions to meet their promotion targets within that time frame, prioritizing investments that can demonstrate measurable progress by the time they are up for review. One popular condition among officials is a geographic reinvestment requirement (返投 fan tou), which mandates that companies bring capital back to the local economy in some form. To hit these targets, companies often trade off investing in longer-term ambitions for resourcing tangible short-term projects.

Private VCs also have their own short-term incentives. Chinese funds have shorter lifespans than their Western counterparts, and as the IPO pipeline has narrowed, investors are turning to instruments that guarantee near-term returns. The dominant instrument of this kind is the redemption clause, a contractual right that requires the company to buy back an investor’s stake plus accrued interest if the firm misses an IPO deadline, valuation target, or revenue metric. This clause is written into more than 80% of Chinese venture and private-equity deals, and roughly two-thirds of the time, that liability extends to the founder personally. As funds struggle to return capital to their own LPs, they have begun enforcing these clauses aggressively, pursuing founders’ personal assets and placing those who cannot pay on a national debtor blacklist. A Supreme People’s Court reference guidance further compressed the window, holding that buyback rights must be exercised within six months of being triggered. This incentivizes fund managers to call in capital before rivals do, passing on the time pressure to portfolio companies.

The result is a capital chain that is “patient” at the national level and impatient at the layers actually making investment decisions. By the time capital finally reaches companies, the 20-year horizon at the top has often compressed into a demand for returns within just three to five. This short-term optimization creates a sense of urgency to quickly sell and exit. While this roughly follows the pattern of a speculative market, its underlying driver is the structure of the flow of government funds rather than investor sentiment.

Institutionalized Herding: Concentration of Capital in “Hot Sectors”

When Beijing signals8 a priority sector, provincial and municipal governments realign their guidance fund mandates to demonstrate political alignment, private VCs shift their portfolio allocations, and founders flock to that sector. GGFs compound this effect by drawing in additional “social” capital (社会资本 she hui zi ben) to their own investment rounds. The 2025 national venture capital guidance fund, for example, is meant to mobilize a trillion yuan ($140 billion) in regional and private capital into state-designated priority sectors over its lifetime. The effect of this capital concentration is that both regional governments and private sector actors are piling into the same state-signaled priorities.

One way this realignment is playing out is through geographic convergence. Regional governments have historically specialized in distinct industrial sectors: Shenzhen in electronics and hardware, Hangzhou in digital platforms, Hefei in display panels and EVs. AI investments nominally reflect this specialization, but capital is increasingly pushing localities into the same subsectors. The “Eastern Data, Western Computing” (东数西算 dong shu xi suan) plan, for example, was intended to route compute-heavy workloads to cheaper-power western provinces. However, between 2023 and 2024, more than 500 data center projects were proposed, many of them not in the West.

At the model layer, the Hundred Model War (百模大战 bai mo da zhan) is another example of capital concentration. After ChatGPT’s launch triggered a global LLM development race, new Chinese companies appeared almost overnight. By mid-2023, China had at least 130 large language models, roughly 40% of the global total. Established tech companies, such as Baidu, Alibaba, and Tencent, were joined by ByteDance, Zhipu, Moonshot, Baichuan, MiniMax, and dozens of smaller entrants who released their own, nearly identical LLMs.

Within the application layer, humanoid robotics has become the system’s newest concentrated bet. In November 2023, MIIT published national guidelines designating humanoid robots as a potentially disruptive technology. Soon after, provincial and municipal governments issued their own humanoid robotics action plans and stood up dedicated local funds. In the first nine months of 2025, Chinese embodied intelligence startups drew 610 funding rounds worth roughly 50 billion yuan ($7 billion). Within two years, more than 150 manufacturers had entered the market.

Unitree Robotics is a clear example of this funding concentration and how valuations can quickly skyrocket. Founded in 2016, the Hangzhou-based firm raised a modest seed round and remained relatively unknown for several years. However, following the 2023 MIIT guidelines, the company quickly raised nearly 1 billion yuan ($147 million) from Shenzhen Capital Group and the state-backed China Internet Investment Fund, alongside private investors such as Meituan. By its Series C, its cap table9 had grown to include the Beijing Robotics Industry Investment Fund, the Shanghai Science and Technology Innovation Fund, CITIC, Tencent, Alibaba, Ant Group, and Geely Capital. After a dancing-robot debut at the 2025 CCTV Spring Festival Gala, secondary trades reportedly valued it above 15 billion yuan, and it has since filed for a STAR Market IPO targeting a 42 billion yuan ($6.2 billion) valuation.

The immense concentration of capital and the inflated valuations are largely driven by investors’ taking “safe bets” and following government signals. While many have and will continue to lose their investment, the funding has counterfactually propelled various AI subsectors forward and produced emerging national champions.

The Export Advantage of Overcapacity

As funding concentrates in a few sectors, companies producing highly similar products begin competing for resources, customers, and political relevance. Many foundation model companies launched just a few years ago are accepting down rounds, pivoting to other business models, or shutting down entirely as margins compress and returns shrink. This pattern of increasing competition over shrinking returns, known as involution (内卷 nei juan),10 has already played out in electric vehicles, solar/polysilicon, food delivery, and now in AI. Baidu CEO Robin Li called the Hundred Model War “a significant waste of resources, particularly computing power,” and President Xi Jinping called for a halt to “disorderly” price competition. Since the boom in 2023, many data center facilities have remained idle or operated at just 20% to 30% capacity, and more than 100 projects have since been scrapped.

In response, the Politburo has launched a formal campaign against involution, but so far the attempt has not amounted to any meaningful results. The NDRC has publicly directed localities to act “according to local conditions” and avoid duplicative investment, and has tightened approval requirements for new data center projects. Various initiatives such as industry self-discipline pledges, drafted amendments to the Pricing Law barring below-cost sales, and a handful of state-brokered consolidations look like concrete steps to addressing price and profit issues. However, many industry observers are skeptical the state will put real force behind curbing overcapacity. This is because while overcapacity destroys investor returns, it also pushes the technology into the economy. By driving down prices, forcing rapid product iteration, and building dense supplier ecosystems, the competition that hurts individual companies is, in aggregate, useful for promoting Beijing’s goals of AI adoption.

Moreover, a saturated domestic market pushes companies abroad to find new customers where they can compete on cost. China’s EV industry is a clear precedent. Between 2015 and 2019, 500 EV ventures emerged, fueled by state subsidies, tax breaks, and easy access to local production licenses. When subsidies began phasing out and competition intensified in 2020, roughly 400 of those companies closed.11 The domestic price wars, however, meant the surviving companies had cost structures that made them incredibly competitive overseas. In 2025, BYD overtook Tesla as the world’s largest seller of battery-electric vehicles (BEVs), selling roughly 38% more BEVs globally.12

We’re beginning to see this pattern repeat in AI. The Hundred Model War cut model API costs by more than 90% and helped drive model usage on China’s public clouds to 536.7 trillion tokens in the first half of 2025, nearly five times the volume of all of 2024. Companies such as DeepSeek, Alibaba (Qwen), Moonshot AI (Kimi), Zhipu (GLM/Z.ai), MiniMax, ByteDance (Doubao), Tencent (Hunyuan), and StepFun are all now finding customers abroad. In February this year, Chinese models processed more tokens than American ones on OpenRouter for the first time. The following month, Chinese open-weight models also overtook US models on Hugging Face, accounting for roughly 41% of downloads.

Within physical AI applications, China’s manufacturing base compounds the advantage. A host of AI-enabled products have been piloted, adopted, and scaled domestically before exporting abroad. The Guangzhou-based firm WeRide now holds robotaxi permits in eight countries and operates more than 100 robotaxis in the Middle East. Gausium has deployed autonomous cleaning machines across European supermarkets, and Supcon is running AI inspection projects with Saudi Aramco.

While the level of AI exports has yet to reach the scale of EVs, if Beijing’s approach to the AI industry continues to follow that of EVs, we could see the same cycle of impatient capital causing convergence and overcapacity, ultimately leading to an export-dominant industry model.

So What Happens Next?

These dynamics explain why foreign commentary on Chinese AI tends to settle into one of two stories. In the first, China is a well-oiled industrial machine. Beijing can channel unlimited resources to priority sectors, mobilize entire industries around a common goal, and force technological advancement almost overnight. In the second, Chinese central planning produces a system that burns through enormous amounts of state funding, incentivizes companies and investors to crowd into the same sectors, and rewards political alignment over actual market value. In reality, it is both at once. The system does waste capital, and at the level of individual companies, it does look more like a bubble than a sustainable industry. It also rapidly creates world-leading industries and enables companies to compete in the global market.

China’s industrial growth model attempts to pair the selection pressures of an “effective market“ with the speed of a “capable government.”13 As a mechanism for generating returns, the system is largely inefficient. As a mechanism for industrializing a technology, it is, so far, remarkably effective. Various measures, including the U.S. outbound rule,14 may slow this process at the margin, but are unlikely to change the overall outcome. This is because they are targeting funding supply, which is not the binding constraint. Instead, if governments around the world want to materially shape the growth of Chinese AI companies, they should focus on the demand side of the equation. Domestic overcapacity means companies need customers abroad to buy their products and services. As a first step, countries can limit access to their own markets if they hope to prevent a repeat of the EV precedent.

Beijing will keep the 击鼓传花 game going as long as it continues to create a growing domestic AI industry with internationally competitive national champions. It’s up to the rest of the world to decide if these AI champions will also dominate the global market in the near future.

To discuss, we have ANU Professor Fengming Lu, who recently wrote a fantastic paper in The China Journal, titled The Rise of China’s Electric Vehicle Industry: Strategic Alliances Between Local Governments and Private Capital, alongside Irene Zhang, contributor and analyst at China Talk. We also get into the lessons this story holds for other strategic industries like robotics and AI — plus a little detour into the Chinese podcasting ecosystem.

Our conversation covers:

• China’s automotive transformation from 130 small carmakers in the 1990s to the world’s most competitive EV industry.

• The rise of Chery, Geely, and other Chinese auto champions — how local connections, foreign technology, and creative financing helped them break into an industry dominated by state-owned enterprises.

• The downside of success — how China’s EV boom created an overcapacity problem.

• The role of guanxi and elite politics in deciding who gets to build cars.

• When Chinese EVs go global — record exports, overseas factories, trade tensions, and whether the same local-government playbook will shape the next wave of industries, from robotics to AI.

Listen now on your favorite podcast app.

Local Beginnings of the Chinese Auto Industry

Irene Zhang: I had an amazing time reading your paper, and I found it incredibly interesting to learn the long history of how the global wave of Chinese EVs we’re seeing today actually came about. I’d like to start by hearing the story from your side. How did you and your co-author come up with the idea for this paper — how did you get interested in EVs as a research topic and decide to write it?

Fengming Lu: It’s connected to my own research agenda. I started doing research on EVs a few years ago, roughly during the final years of COVID. After that, I was mostly doing fieldwork, and I began writing papers and working on book projects last year.

The original idea was to write some of the first papers for my EV research project. My co-author and I knew each other very well, and we’d come across several questions from senior scholars who study Chinese government and the Chinese economy. One of the first puzzles they raised was this: over the last ten or twenty years, especially after 2012, you see the rise of SOEs — state-owned enterprises — in many corners of the Chinese economy. But the automotive sector, or the EV sector, is an exception. It’s probably not the only one. If you look at the tech sectors in China, most are dominated by private firms rather than SOEs. That’s a pretty interesting puzzle for anyone interested in China.

My co-author and I thought it would be a good idea to start with a paper about this. He added some other cases from his own expertise, like pharmaceuticals and steel manufacturing, which also fit the pattern: the private sector, building on collaboration with Chinese local governments, excels at the expense of traditional, more established SOEs.

Jordan Schneider: This was the most fascinating arc in your paper. On one hand we have Xi Jinping, the biggest SOE cheerleader we’ve had in a few premiers. Yet perhaps the most globally impressive, successful industrial leapfrogging has taken place in an industry where the SOEs were basically stuck, doing nothing.

And then two different development pathways were tried — partnering with foreign automakers, and building it themselves — and they were fine, but nothing all that impressive. Then all of a sudden, as you explore in this paper, you get this really interesting mix of private capital, entrepreneurs, local governments, and regional incentives, all dancing around whatever space the central government and the big SOE carmakers left open, creating the incredible flourishing we’ve seen over the past ten years.

Let’s do a bit of a timeline of the Chinese car industry. Why don’t you take us back to the ’90s? Where can we start to see the seeds of the EV boom in that period?

Fengming Lu: Absolutely. And thank you, Jordan and Irene, because you really read my paper rather than just the abstract.

Back in the ’90s, the Chinese automotive industry was in a very different shape from what it is now. If you’re talking about passenger vehicles — cars, SUVs, and so on — the market was only about half a million sold every year. It was much smaller, comparable to Poland back then, or Vietnam today.

For the Chinese government, the main concern was a legacy of the Maoist era and the economic decentralization of the 1980s: the automotive industry was super fragmented. As far as I remember, every province had its own carmaker except Tibet. All twenty-nine or thirty others had their own. By the mid-1990s, someone did the count and China had about 130 carmakers across the country — roughly the same number of carmakers you’d find at the peak of the EV boom, even though back then their combined output was tiny by comparison.

Jordan Schneider: What I thought was fascinating, Fengming, is that it actually made sense to have a local manufacturer in whatever corner of your province. You had these Maoist controls on who got which automobile, so if you wanted any cars in your region, you kind of had to make them yourself. On its face it seems absurd to have a company that makes 500 cars a year, but it may actually be economically productive if the difference is between having 500 cars or trucks versus having none at all.

Fengming Lu: To some extent, it’s similar to the British automotive sector, with all those countless brands and carmakers. What worried the Chinese government was a shock that began in the late 1970s and early 1980s, when they started negotiating with European carmakers, especially Volkswagen. Volkswagen — and many others — basically told them that to make a carmaker efficient, you need to manufacture at least 150,000 cars a year. That’s the minimum economy of scale.

The Chinese government, the regulators, and the policymakers kept this magic number in mind: 150,000 cars a year. So they did some simple math, especially in the 1980s. The market was below one million, which meant it could only support four to six large-scale carmakers. Very simple math.

That’s why, starting in the late 1980s, the Chinese government came up with a concept that’s very famous in the automotive sector: the “three bigs and three smalls” (三大三小). Back in the 1990s, those were the legitimate carmakers you could see on Chinese roads — though of course not the only ones. FAW (一汽) had a joint venture with Volkswagen. Dongfeng (东风), still known as the Second Automotive Works, had one with Citroën. Shanghai already had its joint venture with Volkswagen as well. There were smaller ones too, like Beijing Jeep (北京吉普) and Guangzhou Peugeot (广州标致). And Tianjin Xiali (天津夏利), which had a license from Daihatsu to make smaller Japanese cars.

That wasn’t the only list, either. Later, some carmakers with defense or arms-manufacturing backgrounds — all state-owned, of course — somehow sneaked into the game too, like Chang’an . But it was still a very restrictive list. And elite politics played a role: many of those carmakers pulled various strings to get onto the list.

What you see in the 1990s is a set of approved carmakers in China, but plenty of other provinces and cities still wanted to get into the game. Back then, China hadn’t joined the WTO, so tariffs were very high. In the early 1990s, tariffs on cars ran around 150 to 200 percent, and later came down to somewhere between 80 and 120 percent. So for many local governments and firms, making cars was extremely profitable. An easier route, of course, was to smuggle cars — especially for places like Guangdong, Fujian, and, very interestingly, Shandong, Liaoning, and Jilin, because they were close to South Korea.

But some other cities tried to figure out their own ways into automotive manufacturing. Several of the giant carmakers in China today trace their roots to the 1990s, when they started from very humble beginnings and were later either created by local governments or collaborated with them — companies like Geely (吉利), which today owns Volvo Cars and runs its own brands like Geely and Zeekr (极氪). And Chery (奇瑞). Chery is a very interesting company. They’re rising so quickly now in the European, British, and Australian markets, yet people outside China know so little about them — even people outside Anhui Province know so little.

Jordan Schneider: Great, let’s start with Chery, then. Why don’t you tell the origin story of how it grew starting in the ’90s?

Fengming Lu: It’s such an interesting company. Originally, it was a state-owned enterprise — and it stayed one through the ’90s and most of the 2000s — but it wasn’t owned by any large city, large province, or the central government. It was the state-owned enterprise of a city called Wuhu (芜湖). Very few people have heard of Wuhu outside China, and even inside China not many have. It sits in the middle of Anhui Province, about 500 to 600 kilometers (~400 miles) from Shanghai.

In the early 1990s, the city and the province weren’t in great shape. Anhui had been lagging behind its neighbors — Jiangsu, Zhejiang, and of course Shanghai — since the start of the reform era, in many ways. So they’d been looking for new industries to boost their economy.

I went there for in-depth fieldwork, and combined with other materials I collected, I can reconstruct the story. In the early 1990s, Wuhu was searching for a new industry. They tried many things, and then they noticed a township-and-village enterprise in a village near the city. If I remember correctly, it had started collaborating with an arms maker in Henan Province — they’d somehow made a connection over there, found some engineers, and started building cars based on kit parts from South Korea. They were manufacturing on a very small scale: about 3,000, 4,000, 5,000 cars a year.

Jordan Schneider: Before we go on, can you do two sentences on TVEs? If people have heard of them, they mostly think of them as the thing that got agriculture going in the ’70s. How do township-and-village enterprises end up turning into small carmakers and manufacturers?

Fengming Lu: In the 1980s you saw a lot of small township-and-village enterprises flourishing across China. But starting in the late ’80s and early ’90s, some of them were hit very hard, because their products weren’t technologically advanced or sophisticated. There was also a harsh crackdown on TVEs by the central government after 1989. So a lot of those enterprises were looking for new directions.

This particular TVE found a new way: making cars. It was immensely successful. Their cars weren’t of very good quality, as far as I can tell, but they sold for around US$10,000 — roughly 70,000 or 80,000 Chinese yuan — much cheaper than what the “three bigs and three smalls” state-owned carmakers offered. They made quite a lot of money and became the first company in Wuhu to exceed 100 million yuan in revenue.

So Wuhu’s officials started asking why this company made so much money, and the answer turned out to be cars. The leadership of Wuhu discussed it with the leadership of Anhui, and they decided to give car-making a try. One provincial official put it roughly this way: maybe the party and the government can do something in car-making.

They first pulled strings with one of China’s largest state-owned carmakers, FAW — First Automotive Works. FAW’s general manager at the time was a man named Geng Zhaojie (耿昭杰), still a very respected figure in the Chinese automotive industry for many things he did. He happened to be from the area near Wuhu. The leader of Wuhu befriended him through the National People’s Congress, China’s rubber-stamp legislature. It may not be very meaningful as a legislature, but it’s quite meaningful as a social club. They also befriended another entrepreneur nearby.

So they thought about establishing a carmaker in Wuhu, but it didn’t end well. Government officials in Wuhu told me FAW was so arrogant toward them because FAW saw itself as the grand “eldest son” of the Chinese automotive industry. In administrative rank, FAW was more senior than the city of Wuhu, and it had been the cradle of many of China’s political leaders. So FAW felt it didn’t need to listen to Wuhu. All FAW ended up doing was making peripheral components — chassis for trucks and so on — in Wuhu.

The most telling story officials told me was that the managers FAW sent would smash into the conference room in the middle of Wuhu’s own meetings, shouting, “Is your party secretary here? I want to talk to him.” It was very rude, and eventually Wuhu couldn’t stand it.

At that point, they had a windfall. They sold a cement works to a nearby company and somehow got 200 million Chinese yuan — a large amount back then. That cement works later became Hailuo Cement (海螺水泥); if you follow the Chinese stock market, it was once one of the largest cement companies in China and a Fortune 500 company. They used that money — plus bank loans arranged through the provincial government — when they heard that some engine-manufacturing assembly lines were on offer in Britain. So they organized a trip to Wales, where Ford of Britain was selling an engine-manufacturing line. What they’d learned from FAW was that without your own engine, you can’t make your own cars. Later, they recruited engineers from FAW and elsewhere, and started their own carmaker: Chery.

How they picked their first product is also interesting. It was reverse-engineered from the SEAT Toledo — essentially the Spanish version of the Volkswagen Jetta. The Jetta was also sold in the US in the ’90s, but in China it was one of the very few affordable family cars. Chery decided to go head-on with the Jetta, partly because they had engineers from FAW who knew how to build that kind of car.

Later, in the 2000s and 2010s, Chery gradually became more of a private firm — through a management buyout, and more recently an IPO in Hong Kong. The governments of Anhui and Wuhu still hold about 30 percent of the stock, but they’re no longer the majority shareholder.

That’s the story of Chery, and it’s a great example of how local governments were essentially blocked from making cars. Wuhu had no central-government license to build cars, and for the first few years, they could only sell within Anhui Province. They knew the Anhui market alone could never sustain a carmaker, so they tried several workarounds. One thing I remember — I was a kid then, maybe ten years old, and a car fan — was that they sold cars in Chengdu, an emerging market for car buyers, with Anhui license plates. If you wanted one of the first batch of Chery cars in Chengdu, you had to register it with an Anhui plate. The first people to complain were the Chengdu police, who couldn’t stand it and told buyers they had to register their cars locally.

Manufacturing from Scratch

Irene Zhang: If you were Wuhu, you had a TVE that manufactured vehicles. But for some of the other cities you discuss in the paper, I’m curious how they even got the technology to make cars in the first place. How do you make a car if you’ve never made one, as a city trying to build this kind of relationship?

Fengming Lu: In the ’90s it was actually very hard, because for many of them, all they could start with was reverse engineering — basically copying existing models. Some could afford to buy a bit of technology from carmakers: licenses to build engines, or some blueprints. But few could start from scratch. Even the Chinese joint ventures were still learning how to manufacture cars properly throughout the 1990s. So most of the firms that began as these strategic alliances between local governments and private capital were building relatively low-quality cars, much like the humble beginnings of Japanese and Korean carmakers.

From the 2000s onward, things changed quite a bit. One of the most interesting changes was recruitment. Many firms had started by recruiting talent from within Chinese carmakers and joint ventures — excellent manufacturing engineers who didn’t necessarily know how to develop or engineer cars from scratch. So they began recruiting from overseas: Detroit, Japan, and Europe. Many of those recruits, of course, were ethnically Chinese.

Another interesting change was the capital market. I’m working on a book project, trying to explain why the Chinese EV and automotive sectors have surprised so many people. Ten years ago, they were seen as a joke by many observers, but now they’re one of China’s key pillar industries. I’d say there are three pillars. One is the central government’s industrial policy, which we’ve already discussed a lot. The second is local government, which is central to my paper. And the third is the capital market.

Let me tell you something interesting. Can anyone guess which company was the first from China — even from Greater China, including Hong Kong and Taiwan — to IPO on the New York Stock Exchange or NASDAQ?

Jordan Schneider: Oh man, that’s tough. Can you give us a year?

Fengming Lu: It was 1992. Even before TSMC.

Irene Was it an energy company?

Fengming Lu: Not an energy company — the state-owned enterprises went to overseas markets much later. It was actually a car company. More precisely, a minivan maker, called Brilliance Auto (华晨汽车).

A lot of things happened to that company later, but they were the first to go overseas to sell their stock. They raised somewhere between fifty and sixty — actually about seventy-something — million US dollars on the New York Stock Exchange. Their tagline was that they were the first company seeking an IPO from a communist state, which was sexy enough in a lot of ways. They raised around US$70 million, five years before TSMC. That was huge.

For most other Chinese carmakers back then, all of them essentially state-owned, the only way to raise tens of millions of dollars — or even US$100 million — was bank loans from the state-owned banking system, or in some cases from Western banks on very restrictive terms. The Chinese carmakers found this new way. And later, of course, the Chinese internet companies — Sina, Sohu, and then Alibaba and others — really maximized this model.

Jordan Schneider: That’s basically the story of the ’90s. If you wanted to do something more creative than the SOEs, you needed some other type of guanxi to get around the regulatory restrictions and local selling barriers. Maybe you got drunk with someone at an NPC session and they hooked you up; maybe a mayor happened to be really into cars. It was incredibly difficult, because you had the centralized banking system as well as the large SOE carmakers, who had no interest in your small town and its big dreams.

But coming into the 2000s and 2010s, you start to have this other pathway — raising foreign capital, or pooling money from lots of local banks. We can turn now to Geely as the exemplar of just how creative you can be, going around official Beijing priorities to build what’s now a world-beating firm.

Fengming Lu: Absolutely. Geely has been moving aggressively in the West recently. I think both of you are based in San Francisco, right? You can see some of the new Waymo vehicles — the van-type ones. Those were made by Geely.

It’s a fascinating story. We just talked about Brilliance Auto, which seems totally unrelated, but here’s the connection: around 2003 or 2004, Geely recruited Brilliance Auto’s CFO to be its own CFO. He brought all the know-how on managing the stock market and raising capital into Geely. From that point on, Geely made a lot of smart moves. They were among the first firms to IPO in Hong Kong.

The highlight for Geely was leapfrogging a huge barrier in automotive manufacturing, engineering, and technology by acquiring Volvo Cars in 2010 — Volvo Cars specifically, not the Volvo Group — from Ford. It’s a great illustration of how Geely combined the capital market and local governments to do something big.

I interviewed some of the people who participated in that deal; it stretched from 2007 to 2010, taking Geely almost three years. One of my favorite details — which I also posted on Twitter — involves the lead investment banker, a woman named Jennifer Yu from Rothschild & Co., the European investment bank. Around 2008, she was pitching her colleagues on committing to help Geely. She told them that although Geely was a mouse trying to swallow an elephant — the elephant being Volvo Cars — it was a deal worth doing. At the time, Geely’s revenue was only about 10 billion Chinese yuan, yet acquiring Volvo would cost over US$1 billion. They simply didn’t have that much cash in the bank; they had to find a new way.

Jennifer’s pitch was that, although it was a mouse trying to swallow an elephant, it was backed by a gigantic dragon. She was probably referring to China’s foreign reserves. But that’s not how things work in China — you can’t just go to Beijing and say, “I’m a company trying to acquire an established foreign carmaker. Can you give me a billion or two from your foreign-reserve account?” You need the backing of the banking system, a kind of endorsement, to pull off the deal.

So Geely’s idea was to raise some money from the Hong Kong stock market — but that wasn’t enough — and then lobby for support from various Chinese municipal governments. Starting in 2008: Ford was initially unwilling to sell Volvo Cars, but after the financial crisis they said that if Geely could take good care of Volvo and offer a reasonable price, they could do the deal, on the condition that Geely raise the money itself. So Geely’s most important job throughout 2009 was getting roughly US$2 billion together.

One of the investment bankers on the deal told me Geely talked with as many as 60 cities in China. China has about 300 prefecture-level cities, so that’s a fifth of them. What Geely showed municipal officials was the next-generation Volvo lineup — the XC60s, XC90s, and so on that you see today. The pitch was: “I can let you manufacture some of these models, but in exchange, five fingers.” Five fingers meant 500 million Chinese yuan — about US$70 or 80 million back then. That was the ticket to get on board. With credit from your local banks, or your investment as equity into the acquisition vehicle, “if we assemble four or five cities into this game, Geely can make the system work.”

Initially, the biggest supporter — something I didn’t include in the paper, but probably will in a future book — was Beijing. Surprisingly, Beijing wanted Volvo. Geely even showed Ford’s CEO a letter from one of Beijing’s municipal leaders to demonstrate that it had Beijing’s backing. But in late 2009, Beijing withdrew, because it had its own SOE, BAIC (北汽). BAIC told Beijing’s leaders not to back a competitor within the city — to support BAIC as the sole municipal champion and not stir up trouble by finding a rival.

So Geely — and especially its founder and CEO, Li Shufu (李书福) — was furious and frantic. How could he raise all that money before the deadline? Ford had set it at, if I remember correctly, December 20th — the last business day before Christmas. Could he get the money from Chinese cities before then? Long story short, Geely lobbied many cities and eventually found support from several that wanted an automotive industry. Chengdu, for instance, had been struggling to land an anchor carmaker, so it was very pleased to see Geely and put in 500 million yuan as a bank loan. Some other cities injected funds as equity into a special-purpose vehicle for the acquisition — very interestingly, Daqing, famous as an oil city. Daqing had a lot of cash from drilling but knew it would run out of oil soon, so it wanted a Volvo factory. Shanghai was involved as well.

Jordan Schneider: Let’s recap. In the ’90s, if you were Shanghai you got a joint venture because Jiang Zemin and Zhu Rongji liked and trusted Shanghai. If you were Wuhu, you had this weird connection — Jiang Zemin was sort of from there too — so they let it go. But by the late 2000s and early 2010s, you could be more entrepreneurial.

I’m curious where the central government’s head was on all this. Who knows how much money was changing hands behind the scenes to get things approved — but how happy were the central planners with this entrepreneurial activity, or were they still annoyed?

Fengming Lu: Several things were going on. With Geely’s acquisition of Volvo, the central government’s initial instinct was still to favor the SOEs. What made the deal work was that Geely was very crafty about it: it didn’t move until it had unwritten approval from the NDRC, the National Development and Reform Commission.

If you look back at the news from late 2007 and 2008, the earliest candidates to buy Volvo were companies like Chang’an, a major SOE that also had a joint venture with Ford — a natural partner. But all those state-owned enterprises eventually said no, claiming they didn’t have enough money or that buying Volvo wasn’t a priority. Once the NDRC made sure none of the SOEs were buying Volvo, they told Geely it could go.

There were other Chinese private automakers trying similar games. One of the craziest — a company from Sichuan — tried to buy Hummer, the maker of the US Army’s Humvee. They didn’t get approval, but they announced the deal anyway. The regulators in Beijing reportedly learned about it from the newspaper and TV — “We didn’t hear about this” — and killed it. So there’s nuance here.

The main point is that there was still a bias toward SOEs among Chinese regulators back then. But toward the end they were also happy to see a private carmaker acquire a foreign company if it could boost China’s technology. The catch was, “We’re not providing any financial support. You need to sort that out on your own.” For Geely, that meant dealing with all those local governments first. Even after the NDRC’s approval, the policy banks — like China’s Export-Import Bank — said no. It wasn’t until Geely had secured all the funding that China Development Bank decided it could offer some loans for later construction and plans. They needed to see that Geely had secured everything first.

The Birth of the Chinese EV Ecosystem

Jordan Schneider: So we’re up to the mid-2010s, when electric vehicles really start being part of this story. There’s a big debate — in the literature and among policymakers — about who gets credit. Is it the entrepreneurs, the local governments, or the central government? Why don’t you narrate the birth of the Chinese EV ecosystem and tell us who you think deserves the plaudits for it being born the way it was?

Fengming Lu: My paper tells one aspect of the story — an important part — but I’m writing a book to tell the fuller version. All of them deserve some credit.

We’ve already heard a lot of praise for the central government’s industrial policy: that China had a master plan to build EVs starting in the early 2000s. Parts of that are true. China’s support for EVs has been remarkably consistent over the last 20 years, which makes a real difference compared to all the back-and-forth in the US recently. But it’s not the only part of the story. A decade ago, many people weren’t optimistic about the sector, even though the state was already giving big subsidies. What’s been missing from much of today’s discussion is the role of local governments — and not just private entrepreneurs, but the capital market too.

The most important case is BYD. BYD started as a humble battery maker. Initially, they weren’t even making batteries for cell phones — they made batteries for toys, electric tools, and emergency lights in the 1990s, and later for cell phones. They entered automotive manufacturing in the early 2000s, largely because they’d raised a lot of money on the Hong Kong stock market and figured car-making was lucrative. One big difference was that they were committed to EVs because they could make batteries.

The early years were rough. BYD started making plug-in hybrids around 2008, but the major breakthrough came in the early 2010s. It was made possible in part by central-government industrial policy, but that wasn’t the only factor.

Here’s the story. A lot of discussion of the Chinese EV sector focuses on the early pilot projects. The idea was that EVs would be too expensive for private buyers when batteries were so costly, so governments should buy them first, as buses, taxis, or cleaning vehicles. Public-use fleets, even in small numbers like a few thousand a year, could create a market from scratch.

That’s why China ran two demonstration programs — during the Beijing Olympics and the Shanghai Expo — to showcase its willingness to adopt EVs and other clean vehicles. Around 2011, China also launched the “Ten Cities, Thousand Vehicles” project, initially among ten cities and later among thirty.

The name tells the story. The Chinese government said that for cities willing to adopt more EVs, the goal was to have at least 1,000 electric vehicles running as buses, taxis, and so on within two or three years. It’s a very Chinese way of designing policy: the central government dangles bait for local governments, then assesses performance after two or three years.

BYD realized that many local governments wanted into these projects — eventually 30 cities joined the pilot, far more than ten. The selected cities wanted to maximize their EV usage as a kind of political performance. So BYD chose early local-government partners carefully. The most interesting cases are its hometown, Shenzhen, and the city of Changsha. Changsha had just sold a coachbuilder to BYD around 2008–2009, and BYD worked closely with both cities on the Ten Cities, Thousand Vehicles project.

If you visited Shenzhen around 2010 or 2012, you’d find that overnight, all the taxis had been replaced by BYD EVs. To be frank, they weren’t of great quality — many people told me so — but the Shenzhen government insisted on using them. Many other cities copied this once BYD agreed to invest or build factories there: Changsha, Taiyuan around 2015 or 2016, and others. Overnight, all the taxis became BYD EVs.

BYD also created strong incentives for local governments to succeed in these programs. Ilaria Mazzocco of CSIS has discussed some of this, but there are more fascinating stories. BYD set up its own financing companies to help taxi drivers and bus companies buy EVs. And although Shenzhen didn’t host the Olympics or the Expo, it was holding a much smaller event — the Universiade, essentially the Olympics for university students. The Shenzhen government took it very seriously and decided to copy the Beijing and Shanghai model with its own EV demonstration program. They asked BYD to deliver EVs, and Shenzhen eventually put 2,011 EVs into service for the event — BYD contributed a quarter of them.

It was a humble start, and BYD copied the model in many other cities. Throughout the 2010s they invested in city after city. Today, they have major assembly plants and other facilities across more than a dozen Chinese cities. Their strategy was interesting: they tended to choose second- and third-tier cities, especially in central or western China, where labor was much cheaper.

They also had this exchange to offer: “We’ll replace all your taxis and buses with our EVs.” To smooth collaboration further, BYD has a dedicated rail-transport division. They’d tell local governments, “If you want more from us, we can build a light-rail project for you.” It’s a mass-transit rail system the local boss could point to as a big achievement.

That was one secret behind BYD’s quick expansion after the late 2010s. Actually, “succeeded” is too strong for that period — they sustained themselves through the 2010s. Back then, BYD was much smaller, with the nickname “B-thirty-thousand” (B三万 B sānwàn), because they could only sell about 30,000 vehicles a month. Now they sell around 300,000 a month — about ten times as many. They survived those ten years largely by selling EV buses and taxis to cities, in China and abroad. The key link was that by helping local governments hit their policy targets, BYD simultaneously expanded its manufacturing and market share.

Irene Zhang: This conversation just unlocked a deeply buried memory. I lived in Shenzhen for a few years as a kid in the 2000s, and I remember going back around 2017 after a long absence. I got out of the train station with my family, walked toward the taxi lane, and every single cab was the exact same make. I didn’t know what it was back then — I was just a kid thinking, “Wow, they exchanged all the cabs.”

Fengming Lu: Yes — there’s a very good reason for that.

The Overcapacity Problem

Irene Zhang: This is a good segue toward the present: this entire machine of an industry going abroad and reshaping the global automotive market. Do you want to talk about the capacity-expansion side and how Chinese carmakers ended up shaping the industry writ large?

Jordan Schneider: And also, to what extent is the overcapacity something the central government is happy or upset about?

Fengming Lu: Overcapacity has been chronic in the Chinese economy, especially in manufacturing. One key driver is this combination of private capital and local governments. It happened before in steel and other sectors — and now in autos.

What really happened in the 2010s is widely seen as a turning point. In 2014 and 2015 — the peak of “mass entrepreneurship and innovation” (大众创业、万众创新) — the Chinese government lifted the restriction on private capital, and even private startups, building cars. That’s why China has not only the companies you see today, like NIO (蔚来), XPeng (小鹏), and Li Auto (理想) — what people call the “new forces” (新势力) — but also a wave of now-bankrupt firms. I did the count for another paper I’m writing: there were 60 or 70 EV startups in China after 2015. That’s crazy. In the US you have Tesla, Rivian, Lucid, maybe Fisker — five or six. In China, 60.

Many were purely private and weren’t very successful because they ran out of cash early. Some tried to form joint ventures with local governments and still ran out of cash. The ones that survived could raise capital not only from their founders and local governments, but also from corporate venture capital. Look at the more successful EV startups, and you’ll see many backed by Chinese internet and tech giants like Alibaba, Meituan, and Tencent, as well as corporate VCs from emerging battery makers like CATL (宁德时代), the largest battery maker. With 50 or 60 startups competing, that created enormous waste and overcapacity.

Around 2020, China tried to absorb the overcapacity. The NDRC halted approvals for new EV manufacturing capacity around 2020–2021. But then came the fascinating Hefei model (合肥模式). A few months ago, there was a New York Times article about it.

Here’s what happened. NIO had some market success in 2018 and 2019, among the first Chinese startups to release EVs competing with Tesla. But from late 2019, they began running out of cash, partly because they were building so much — including many battery-swapping stations — and their NASDAQ stock price plummeted during COVID and for other reasons. So from late 2019 they started raising capital from local governments. Some of NIO’s early colleagues — colleagues of Li Bin (李斌), known in English as William Li — told me NIO was initially very reluctant, because local-government capital comes with conditions: invest a certain amount in my locality, and so on. But by late 2019 they had no choice.

Their initial supporter was, again, Beijing — but at the last minute, Beijing canceled the deal, again because of BAIC, its own SOE, which didn’t want a new competitor. NIO was desperate by early 2020. The savior turned out to be Hefei, which already hosted NIO’s factory. Hefei injected about 7 billion Chinese yuan — roughly US$1 billion — as equity into NIO. Initially, it was a win-win: NIO got the cash it needed to survive, and Hefei sold all that equity on the stock market about a year later for a three- or four-fold return — pretty remarkable for such a large investment.

After that, the Hefei model was celebrated across China as a great example of local governments acting like venture capitalists — even though they aren’t — to save a startup and boost the local economy. Ironically, when the central government tried to minimize overcapacity, the Hefei model fueled more of it. From around 2020, many cities that had missed out on EV startups rushed in, looking for new startups to back with equity. In 2020, Guangdong Province and the city of Guangzhou injected about 2 billion yuan into XPeng, adding ammunition to one of those startups. GAC (广汽) created its own startup, Aion (埃安) — you see those cars everywhere as Uber, Didi, and taxi vehicles — and it got massive equity investment from Guangdong Province and Guangzhou. Ningbo, which had missed the earlier startups, decided to collaborate with its home carmaker Geely and injected about 10 billion yuan — a massive amount — into Geely’s own startup, Zeekr.

They all hoped to replicate Hefei’s success: back a startup, then sell the stock for a two-, three-, or four-fold return. But it didn’t pan out, because the valuations of both EVs and Chinese stocks plummeted in the US after 2021. By then, though, they’d already fueled the overcapacity.

Irene Zhang: This is a good place to dig into the overcapacity. When did people in China start hearing the alarm bells, and how does it spill over?

Fengming Lu: Starting around 2023 and 2024 there were quite a lot of alarms. People are talking about it this year, especially, because you’ve seen a 20 percent year-on-year drop in China’s new-car sales, which is pretty alarming. But early signs showed up in 2023 and 2024.

The central government essentially stopped approving new automotive manufacturing plants — even EV plants — around 2022–2023. There were also new restrictions, from 2024, on local governments setting up their own private-equity investment funds. During my fieldwork in 2023, I saw many counties establishing their own investment funds, looking for tech firms — not necessarily EV firms — to invest in their localities in exchange for new business and investment.

Overcapacity has been a big issue, and people have been thinking about further consolidation among EV makers, which has proven hard. Some companies went bankrupt — most prominently HiPhi (高合) in 2023, and more recently Neta (哪吒), whose parent company is Hozon (合众新能源), last year. They were failures among the mass of Hefei-model imitators. Neta had received massive equity investment from Yichun, a city in Jiangxi Province — not a particularly wealthy place. Yichun had a lot of lithium reserves and feared missing the last-chance saloon, so it wanted its own EV maker. When Neta went bankrupt, it meant at least several billion yuan in losses for the local government’s public finances. Dealing with all this overcapacity — gradually absorbing the losses and the bankrupt carmakers — has been very challenging.

Another challenge that people outside China tend to overlook is what to do with the established SOEs. Many haven’t been very successful in EV manufacturing or engineering. Can Beijing — the central government — simply tell them to merge with one another? Since early last year, there have not just been rumours but concrete moves to have Changan, which is more successful than other SOEs, merge with Dongfeng. It took almost 12 months, and nothing really happened, because neither side’s home city — Chongqing and Wuhan, respectively — was willing to make concessions. That’s now been shelved.

Irene Zhang: Earlier, we touched on your previous work on elite politics — you have some really interesting papers about the black box at the top. How has it been doing this more down-to-earth work on EVs compared with that?

Fengming Lu: There are still several links between them. When you talk about elite politics in China, almost all senior officials have rich experience in provinces and cities. Many have connections, but they also need to show deliverables to their bosses — to prove they can get something done. For many of them, having a major automaker in their city or province is a big achievement they can sell to superiors.

Take Xi Jinping’s time in Fujian Province. Before the 1990s, there was almost no automotive manufacturing in Fujian, so one of his highlights was establishing a company called Southeast Motor (东南汽车) with a creative solution. In the 1990s there was a set quota for joint ventures — no new ones beyond the three bigs and three smalls. So Fujian’s provincial automotive company found its joint-venture partner not abroad but in Taiwan, from a company that was simultaneously manufacturing Mitsubishi cars there. They told the regulators in Beijing that it wasn’t a joint venture with a foreign carmaker but with a Chinese carmaker in Taiwan — the partner was even called China Motor Corporation (中华汽车). So it was effectively a cross-provincial joint venture: “Taiwan is one of our provinces, right? You should approve it.” That was one of the highlights of Xi’s record.

Or take Li Qiang, the current premier. His biggest achievement in Shanghai was simply Tesla — building that Gigafactory in roughly 14 months. So there’s a fascinating link between elite politics, political selection, and the auto sector: in many ways, your political performance is automotive manufacturing.

The second thing is that you always find these fascinating little stories where you need your personal connections to make things happen. Let me tell you a couple from the auto sector.

We talked about the three bigs and three smalls. The quota was supposed to be only two or three state-approved makers of passenger vehicles, because the main concern back then was that China was so poor in the 1980s — few people could afford cars, so the government shouldn’t approve too many makers. But several SOEs tried to squeeze in. FAW and the Second Automotive Works, Dongfeng, had essentially been truck makers. FAW made the Hongqi (红旗 Hóngqí), but only in tiny numbers — limousines for a handful of leaders. So they wanted approval to make passenger vehicles. The Second Automotive Works — later Dongfeng — squeezed in with the help of a good friend named Bo Xicheng (薄熙成). From the name you can already tell who he is: the son of Bo Yibo and the younger brother of Bo Xilai. Bo Yibo was extremely powerful in the 1980s, so that connection was crucial.

Then, regarding why the government in 2014–2015 opened the gate to new EV startups — having previously thought EV-making should be a game only for established automakers and especially SOEs — one big reason was a gentleman named Chen Qingtai (陈清泰), very influential among think tanks and policymakers. He wrote many policy reports arguing that China should allow more Tesla-like firms to emerge in EV manufacturing. He believed companies like Tesla represent a new, revolutionary path: you raise money from the capital market, especially venture capitalists, who do a much better job of incentivizing innovation, and you offer employees stock options to spur innovative ideas. China should copy the Tesla model.

Two things about Chen Qingtai are striking. First, he was the general manager of Dongfeng — the very company that sneaked into the joint-venture game with the help of Bo Yibo’s son. Second, he had special connections from studying automotive engineering at Tsinghua University in the 1960s, where he was a member of the university’s dance troupe. There, his “little brother” — by courtesy — was Hu Jintao. Chen was a close friend of Hu’s, and in fact sponsored Hu Jintao’s Communist Party membership. With that special connection, his policy proposals carried a lot of weight in the early 2010s. He probably wasn’t the only person who made it possible, but it was an important move that shaped the Chinese EV sector into what it is today.

Jordan Schneider: Who’d have known? This comes back to the theme. There’s an aspect of the story rich in entrepreneurial dynamism, but it’s overlaid on a system not entirely set up for it. To compete in the game of EVs and have any chance, you need to be an incredible entrepreneur — an amazing company, the best cars, product–market fit — and you need the guanxi: a friend who’s a premier or a provincial governor.

It’s fascinating that there were just enough people with enough guanxi to spin this up, so that you still get creative destruction despite the obstacles. You don’t necessarily have that in other markets, where you have to run through all these weird regulatory loopholes set up to prevent you from even starting.

Fengming Lu: In the US, you don’t really need to deal with the government until your market cap reaches, say, 10 billion or even 100 billion dollars. In China, from the beginning — if you’re a small business it’s probably fine, but if you’re starting a decent-sized business — you need to know how to deal with the government first. Maybe starting with counties or townships, then cities and provinces, and later the central government. You learn that very early on. If Elon Musk were starting his business in China, he’d probably be learning it from day one, rather than only this year.

Jordan Schneider: Totally. It’s an actual core competency you have to develop, and it doesn’t happen by accident. These CEOs are very strategic about who they want on their cap table, how much government money to take, what type to take and not take, and which politicians to bet on advancing in the future.

It’s a fascinating game. You wouldn’t necessarily think — okay, Xiaomi can design a phone, so maybe they can design an EV too; there are core competencies that align. But being good at the local, provincial, and central-government game doesn’t seem like a natural fit with being good at manufacturing cars. The fact of the matter is, you don’t get a shot at manufacturing cars unless you can figure out the political dynamics first.

Fengming Lu: Absolutely. It’s acquired to some extent, but you also need some gift or talent for it.

Share

Irene Zhang: I’m specifically interested in whether you think the international market is the solution to all this overcapacity, or whether you have reservations about that thesis.

Fengming Lu: That goes beyond my current papers — maybe even beyond my book — and political scientists usually don’t make predictions. But I’m willing to share some observations.

Starting from my paper, what’s interesting is that Chinese carmakers are trying to copy their experience of dealing with local and central governments abroad. Take BYD’s rail-transport division: BYD is now building rail transport in Brazil — I think they’ve already delivered a project there. They see it as a way to build their public image, not only with governments but with ordinary people, signaling that they’re a good corporate citizen. Many other carmakers wouldn’t do that — General Motors and Ford never built rail transport in Brazil.

But the strategy can be risky, because many of the countries they’re investing in still hold elections. In Hungary, BYD seemed to be betting on Viktor Orbán staying in power for another term, but now they’re hitting some pushback. With a new prime minister in Hungary, their earlier plans for an automotive plant — and their hope of using Hungary as a kind of outpost in the European Union — look a bit shaky.

The second thing — and part of why people are interested in my research — is that this year’s surge in Chinese car exports is massive. By many projections, if the pace continues, China’s car exports could reach about 12 million by the end of this year. For perspective, the peak of Japanese car exports in the 1980s — around 1986, during the trade disputes between the US, Japan, and Europe — was about 6 to 7 million cars. So China is exporting almost twice as many cars as Japan did in the 1980s, when that triggered all those trade disputes.

That’s why you’re seeing new policy moves, especially from the European Union, which still makes a lot of cars — Germany, France, and so on. They’re trying to fix the loopholes; there are proposals to tariff Chinese plug-in hybrids, where before the high tariffs applied only to Chinese battery-electric vehicles. There’s also talk of tariffs similar to those imposed by the Trump administration. Over the long term, this massive export probably isn’t sustainable — you can’t imagine China exporting 15 million cars to the rest of the world next year; that far exceeds everyone’s expectations.

So China’s carmakers will need to think seriously about shifting toward more localized manufacturing — as American, European, Japanese, and Korean carmakers already have. Some are already building plants in Europe, South America, and Southeast Asia, and increasingly in Europe they’re acquiring idle manufacturing capacity, because European carmakers are also suffering from overcapacity — partly from the influx of Chinese cars (EVs and gasoline cars alike), and partly because during COVID many of them, especially in Europe, overestimated market growth. People were waiting months for new cars in Europe and the US, so automakers expanded capacity that’s now a huge burden.

It’s not really a problem for the US at the moment, but in Europe younger people are less willing to buy cars or even get driver’s licenses. Over time, the European market may suffer not just stagnation but decline. In that sense, Chinese automakers need new ways to localize manufacturing — and eventually parts supply, design, and engineering too. Many have already made some moves.

Jordan Schneider: How do you see this evolving? The story of the past 30 years is more about companies and provinces driving it than Beijing. Now that it’s a big source of international tension — a flagship industry — how do you see the relationship between the central government, the industry, and these companies changing in the coming years?

Fengming Lu: At least for the auto sector, it will still largely rest on partnerships between cities and firms. The central government already sees the EV sector as a relative success — still troubled by overcapacity, but clearly one of China’s tech successes. And over Chinese history, the central government has never really had experience directly managing the automotive industry.

There were earlier efforts. In the 1980s, the government considered creating a national-level conglomerate to manage all the automotive plants — they set up something called the China National Automotive Industry Corporation (中汽公司). But it was very short-lived, lasting only a few years, because no local government was willing to give up its carmakers. The central government has been fairly successful over the last 10 to 15 years at merging other SOEs, but in the auto sector it’s been much more challenging.

It’s a little beyond my research, but many people are interested in how this model might evolve in other tech sectors — robotics, semiconductors, AI. I’m not a semiconductor expert, but before 2017–2018 that sector followed a path very similar to autos and EVs: most new startups were products of collaboration between local governments and private firms or private capital. But because of the trade war, Huawei, and other events from 2017–2018, semiconductors suddenly became a chokepoint issue, and the central government stepped in with central-level funds.

Robotics is closer to EVs, I’d say, because it isn’t yet seen as strategically important as semiconductors — not as much an infrastructure-type thing. For AI, let’s keep an eye on it. There’s a lot of local-government involvement. DeepSeek, for instance, was essentially initiated by a hedge fund in Hangzhou and was, to some extent, protected by Hangzhou.

Jordan Schneider: That point was really interesting. Over the 2000s and 2010s, the central government wasn’t particularly excited about a hedge-fund industry existing — you had all these statements about financial speculation not creating value. But Hangzhou said, “No, this will create some rich people, we’ll get tax benefits, whatever — you can have your hedge-fund village.” Very Tang Dynasty: you hang out here by the nice streams, make some money, and we’ll just tax you. Then you’re lucky enough to have a billionaire who thinks it’s cool to train models, and DeepSeek comes out of that. It’s totally not centrally planned, but because you had regional variation that let the industry flourish, you ended up contributing to national goals in a roundabout way.

Fengming Lu: That’s true. A recent development with DeepSeek: they just did a kind of pre-IPO round and raised a lot of money. If you look at the investors, they’re dramatically different from the earlier ones. You still see plenty of private investors, VC or PE funds, and corporate VCs — but now you also see investors with a state background getting in, and some with a central-government background. That isn’t happening massively in the EV sector, because EVs aren’t quite as infrastructural, but it may happen with some of China’s key AI companies. Robotics, we’ll see. Over the next few years, let’s keep an eye on it.

Jordan Schneider: One of my favorite things to do is watch Bilibili reviews of Chinese EVs. The long list of random features these companies throw in to differentiate themselves is a font of endless fascination. Fengming, do you have any favorites?

Fengming Lu: The far-sensing is pretty innovative — I’ve seen it on some of the smart-brand cars, the brand co-engineered by Mercedes-Benz and Geely. As a traditional car guy, what I appreciate most about these cars is something simpler. Every Chinese EV I’ve driven has a genuinely good built-in navigation system. With most ICE vehicles I’ve owned, the built-in navigation exists, but you never use it — you just use your phone, because phones are far better. But Chinese EVs have very good built-in navigation in their infotainment systems, which has solved a lot of hassles for me.

Another thing I like is that in many other markets you have to pay a lot — buy the fully loaded version — to get features like ventilated or even massage seats. In China, they’re offered on very affordable models. That’s partly thanks to this evolution, and partly because the rise of the EV startups brought many Chinese parts suppliers into the supply chain. They used to be peripheral, but they became capable of building good-quality parts — seats, interiors, and more advanced ones like automotive semiconductors, braking systems, and drivetrains.

This also means a lot is changing for automotive engineers in China. I regularly do Chinese-language podcasts with engineers based in or dealing with China. One of their biggest complaints is having to meet unreasonable demands from project managers — “Add this and that within two or three weeks” — which is often impossible.

A big issue for Chinese carmakers now is how they treat their parts suppliers. Many companies have become notorious for delayed payments — some take a year to pay suppliers, essentially borrowing at their suppliers’ expense. That’s not sustainable. If the Chinese EV industry wants to prosper over the next few years, there needs to be a better arrangement between parts suppliers and the OEMs — the original equipment manufacturers.

Irene Zhang: This is a very basic point, but I can’t get over how quiet they are. I’m reporting live from Guangzhou, which is still kind of a motorbike city, and the traffic in the old parts is still chaotic. I’ve literally become bad at crossing the road because of how quiet EVs are. When you cross, you expect to hear cars approaching, but they just sneak up on you. It’s amazing.

Fengming Lu: Yes — so all those urban-noise problems got solved in a very unexpected way.

Irene Zhang: And the emissions problems too.

Jordan Schneider: You mentioned Chinese podcasts. On your homepage, it says you’ve been listening to podcasts since you had to download them to your iPod — a memory I fondly recall. Talk a little about your perspective on the Chinese podcast ecosystem. How have you seen it evolve over time?

Fengming Lu: It’s amazing. The first podcast I listened to was called Antiwave (反波), back around 2004. I listened to it not even on an iPod but on my own personal computer. There was very early podcasting in China then. You probably know him — let me find his English name —

Jordan Schneider: Goldkorn.

Fengming Lu: Jeremy Goldkorn, yes. He had his own podcast in the 2000s, and I was one of those early listeners. Later, with the arrival of iPods, I bought several iPod Classics, because they were large enough to hold all those podcasts. From around 2008 to 2010, people were downloading from iTunes — and I still remember how terrible iTunes was for downloading podcasts, full of bugs.

Most early Chinese-language podcasts were just for entertainment and fun. But starting in the 2010s, people began to realize that many US podcasts — especially the Silicon Valley type — had real, concrete knowledge in them. So Chinese shows emerged that discussed books and specific research, especially in the humanities and social sciences, from the mid-2010s — podcasts like Hu Zuo Hu You (忽左忽右). I was an early listener to those too.

From around 2020 you got a much more diverse set of podcasts. The one I take part in was created in 2021–2022. It’s called The Isle of Auto (孤岛车谈), started by a group of automotive engineers based in the UK — hence the name. It’s a Chinese-language podcast about the auto sector, and I’ve made many friends in the industry through it. It’s a relatively small show by listenership.

There are far more popular shows, many clearly trying to mimic successful American podcasts like Money Talks or The Money Show. Some have half a million, one million, or two million listeners — a lot.

There’s a big difference in how Chinese-language podcasts are consumed. In the US, most people listen through a podcast app or Spotify, with little interaction. But in China, since the 2010s, platforms like Lizhi (荔枝), and later Ximalaya (喜马拉雅), and more recently Xiaoyuzhou (小宇宙) — XYZ — built features that Chinese users appreciate. You can comment under a specific episode, like specific episodes, and discuss with people in the comment section, creating much more interaction. A lot of US podcasters probably wouldn’t appreciate those features, but they’re very successful in China.

One big problem for Chinese podcasts is that it’s very hard to get people to pay. In the US, you can run advertisements — including programmatic, customized ads, like on YouTube. But none of the Chinese platforms are really willing to try those monetization methods.

Jordan Schneider: Do you have any idea why?

Fengming Lu: I’ve discussed it with many friends. At the end of the day, most Chinese podcast listeners still prefer free episodes. A few are willing to pay for specific episodes — maybe a few thousand people — but that doesn’t make much money.

And it seems that Chinese listeners are more willing to accept what they call “soft ads” (软广) — promotional content built into the episode — and less willing to accept ad-network placements. Similar things are happening with Chinese video sites like Bilibili: whenever Bilibili tries anything with an ad network, it causes a lot of resistance from users. So even though Bilibili has so many users, it isn’t making much money, because it can’t simply use an ad network.

Jordan Schneider: It’s interesting — you see far more creator-read ads, or a food creator visiting a city because the city paid them to. There’s a lot more sponsored content than straight YouTube-style ads that have nothing to do with the show. Honestly, I’m jealous of Chinese content creators. When I post something on Spotify, I get maybe two comments, but you see these long discussions where people argue about the points the hosts make on the Chinese platforms. I hope folks learn from that.

To discuss, we have ANU Professor Fengming Lu, who recently wrote a fantastic paper in The China Journal, titled The Rise of China’s Electric Vehicle Industry: Strategic Alliances Between Local Governments and Private Capital, alongside Irene Zhang, contributor and analyst at China Talk. We also get into the lessons this story holds for other strategic industries like robotics and AI — plus a little detour into the Chinese podcasting ecosystem.

Our conversation covers:

• China’s automotive transformation from 130 small carmakers in the 1990s to the world’s most competitive EV industry.

• The rise of Chery, Geely, and other Chinese auto champions — how local connections, foreign technology, and creative financing helped them break into an industry dominated by state-owned enterprises.

• The downside of success — how China’s EV boom created an overcapacity problem.

• The role of guanxi and elite politics in deciding who gets to build cars.

• When Chinese EVs go global — record exports, overseas factories, trade tensions, and whether the same local-government playbook will shape the next wave of industries, from robotics to AI.

Learn more about your ad choices. Visit megaphone.fm/adchoices