There is something about the American liberal (classical version!) that grates me when it comes to China.
Its not as honest as the more traditional 'right-wing' china-hawk, whose obvious and admitted reason for disliking China and / or pushing for confrontation is grounded in profoundly nationalist thinking. Say what you want but that's an honest take and one that's easy to respect.
This particular brand of liberal though (lived in china, likes chinese culture, 'disappointed' that china hasnt become a liberal version of the US, etc etc) is just irksome. They want a china to be a larger version of japan - culturally cute and attractive and a good source of la mian and poems and ink-paintings and all that other harmless gunk. But crucially it needs to be geopolitically neutered...just to be safe!
Its such a dishonest view, wrapped in endless contradictions. And the insecurity really comes out at times....'WE have escalation dominance! Not you! WEEEE do!!! We're gonna prooove it next tiiimmmeee!! Go get em Donald! You show em Scott!'
Genuinely quite sad to read. Especially because dependencies dont even begin and end with chips (good luck btw getting TSMC to cancel work for every one of its chinese customers without any blowback). Medicines, crucial industrial inputs, etc etc. It will get very very nasty for everybody involved if this truly moves into the upper ends of the escalation ladder (regardless of who has an extra rung here or there).
Perhaps should be called the “Talk Down to China Podcast” @tkz1 spot on! Whenever someone is so insecure that they need to continually tell you they are the best in the world, instead of letting the facts speak for themselves, it tells you all you need to know about their position… It’s very ironic how they claim America’s superiority where as far as I can tell, without the very non American TSMC and their Taiwanese engineering prowess and the nonAmerican ASML they would hardly have any chips at all. Well, there is that American chip powerhouse Intel… not too mention that much of the critical hardware to buildout all the AI infrastructure is coming from China as they are unable to produce it in the USA. So yes, a little humility and self reflection could come in handy! Given the trajectory of US stem education and the reduction in foreign talent, it’s hard to imagine the US being competitive globally 20 years from now. The US is similar to Great Britain in 1905, their best days were far behind them but arrogance was at its peak!!
Watching you little china loving twats pretend it is somehow immoral for the United States to chose not to be an active supporter of China’s geopolitical ambitions is comical.
Go fuck yourself you retard. Or if you’re a Chinese propagandist - go fuck yourself you stupid bitch.
We have a security imperative to get slavery out of our supply chain. Communist China is the most expensive option. We just need to stop capture deletes from making us foot the bill.
This is one of the most thorough hardware net assessments I've seen on the China-US AI semiconductor competition. The math here is compelling - if Huawei needs to make 200 million chips by 2027 to match Nvidia's aggregate compute, that would require essentially replicating TSMC's entire production capacity devoted solely to AI chips. That's not happening with current fab constraints and yield issues at 7nm. What really stands out is how the quality gap is widening, not narrowing. The 26x performance differnce projected for 2027 is staggering. This contradicts the common narrative that China always catches up through scale and iteration. The key difference you highlight - that this is the most complex supply chain in human history requiring EUV and advanced packaging tools they can't access - makes indigenization fundamentally harder than solar or EVs. I'm curious about the HBM constraint you mention. If CXMT can't scale HBM3 production, that becomes the binding constraint even if they solve logic die fab capacity. The policy implications are clear: maintain the tooling restrictions and don't ease up on controls just because headlines suggest breakthroughs. Great analysis.
The quantitative framework you laid out is incredibly clarifying - when you run the actual math on wafer production needs versus TSMC's total capacity, the scale of China's challenge becomes stark. The 200M chips needed by 2027 to match Nvidia would require essentially replicating an entire TSMC across all its production devoted solely to Ascends. What I find most compeling about this analysis is the insight that multiple exponentials are working against China simultaneously - compute demand growth, quality improvements from US side, and their constrained production scaling. The HBM bottleneck you mentioned is underappreciated - even if they could scale logic die production, the memory bandwidth constraint alone would be crippling. This is fundamentally different from solar or EVs where China could brute-force scale to dominance.
This quantitative analysis is exactly what's needed to cut through the noise. The math on China needing 200 million chips by 2027 to equal Nvidia is startling - that's essentially building multiple TSMCs from scratch while stuck at 7nm. What stands out is how the compounding effects work against China: they're not just behind on quality (26x performance gap projected), they also face exponential demand growth that outpaces even generous production scaling assumtions. The HBM constraint adds another critical bottleneck that often gets overlooked in breathless media coverage about Huawei breakthroughs.
This is an exceptionally thorough quantative breakdown of the production capacity gap. The math on the 200 million Ascends needed by 2027 to equal Nvidia's output really drives home how challenging China's position is. What particularly stood out is the point about competing with 'the flagship of global capitalism' - this isn't like solar or EVs where the West ceded interest. The entire weight of global capital is behind TSMC and Nvidia, creating compounding advantages on both quality and quantity dimensions. The HBM constraint you identified is also critical and often overlooked in media coverage. Without domestic HBM3 production at scale, even if SMIC can fab more 7nm logic dies, those chips will be severely bottlenecked. Great conversation.
The production math here is really compelling and often missing from the breathless headlines about Chinese chip breakthroughs. The distinction between designing a chip and manufacturig it at scale is crucial. What strikes me most is the exponential demand curve for AI compute outpacing China's ability to scale domestic production even if they solve some of the technical hurdles. The 50% rule change is interesting policy architecture - it's much harder to play shell games when subsidiaries are automatically covered. The HBM constraint is something that deserves more attention in these discussions since memory bandwidth bottlenecks can negate other advances.
Fantastic deep dive into the semiconductor dynamics. The quantitative analysis comparing Nvidia's 7-8M chips in 2027 versus Huawei's projected 600K chips really puts the scale gap into perspective. What I found particularly compelling is the discussion about HBM as the critical bottleneck - this is often overlooked in mainstream coverage but represents a fundamental constraint that can't be engineered around quickly. The historical parallel to Soviet supercomputer restrictions is apt, though as you note, we're now dealing with commercially available technology rather than specialized lab equipment. The 50% rule change from BIS is also significnt - closing subsidiary loopholes will make enforcement substantially more robust. Great work bringing actual math to a debate dominated by headlines.
The quantitative breakdowns here are exceptionally helpful for cutting through the media hype. The math showing Huawei would need roughly 200M chips by 2027 to match Nvidia's aggregate compute (versus their projected 600k-700k production) really crystallizes how severe the gap is. What strikes me most is the dynamic you highlighted about multiple exponentials working against China simultaneously - they need to close gaps in chip design quality, scale production quantity, AND do all this while demand itself is growing exponentialy. That's basically impossible without access to the upstream tooling ecosystem.
This is an incredibly thorough analysis that cuts through a lot of the noise. The distinction you make between quality and quantity gaps is crucial - it's not just that Huawei is making inferior chips, but that they would need to ramp production by 60-70x over two years just to keep pace with the current gap, let alone close it. What strikes me most is the point about indegenization. China has succeeded at this repeatedly across industries, but semiconductors are fundementally different because the tools themselves are the constraint. You can't reverse-engineer an EUV machine easily when it represents the pinnacle of human manufacturing capability. The export controls are working precisely because they target the choke point where China's traditional strengths don't translate.
This is an excellent deep dive into the hardware net assessment. The point about supercomputers now being 'available in a box off the shelf from Nvidia' really captures the unique challenge. The math Chris lays out is compelling - even if Huawei doubles production to 600k chips, that's still dwarfed by Nvidia's 5-7 million, and the quality gap is increasing by 6-7x over two years. The 50% ownership rule change is also a significnt enforcement improvement that should make it harder to play shell company games. Really appreciate the quantitative rigor here!
Really helpful breakdown of why the gap isn0t just about TFLOPs. Two things that seem underpriced: (1) HBM supply chain lock-in0you can0t brute force bandwidth without stacked memory and advanced packaging yield, and China0s domestic HBM is years behind; (2) software+systems maturity0CUDA, compilers, libraries, and networking stacks that squeeze utilization out of clusters. Even if Ascend hits H100-ish peak specs on paper, sustaining performance at scale needs the whole stack tuned. Add node stagnation (~7nm) limiting density and power efficieny gains, and the math compounds against catch-up. Curious how much CloudMatrix can offset interconnect/topology limits before power and capex walls bite.
Chris's analysis is probably the most neutral and clear-eyed on the semiconductor race that I've read. We would be foolish to underestimate our lead over China in compute, but just like the situation in China where people dare not tell bad news to the party leadership, I wonder if the current structure and style of the current US administration might make it similarly difficult for the truth to get to the ultimate decisionmakers. Hence this administration's vacillation on export control policy.
One implication from Chris's analysis though is that it seems China may have only a narrow window to stop the US from gaining an unassailable lead in compute and thus an unassailable military advantage. We know the CCP can do so by taking control or destroying the world's most advanced fabs that's only 100 miles away from China's coast. I am not sure if either the US or Chinese leadership have realized that the military balance of power may have already shifted decisively to China, given that the Chinese have a huge advantage in industrial capacity, have a significant technological advantage in battery technology and electronics that allows them to produce far more drones and missiles, and a powerful tool for kneecapping the US defense industry in the short term with rare earth controls. However this narrow window of opportunity for them could close quickly once the US's compute advantage kicks in in a few years and AI is fully integrated into the US military. If the Chinese leadership realizes that time is not on their side, I wonder how they would react. As Paul Atreides said in Dune, the power to destroy a thing is the absolute control over it.
All this to say, if we keep climbing the escalation ladder, at least during this narrow window, I think we may not like what we find when we reach the top. Even if the Trump administration is making the wrong policies now, if it could lure the CCP leadership into a false sense of security until the military balance of power shifts, those policies may end up becoming the correct ones for the long term.
Nice pod. I’m a little conflicted, I had other people in the semi supply chain say that Huaewei was actually closer to 1M dies this year. Still this 5x doesn’t close the gap, but it is obvious that have huge error bars.
There is something about the American liberal (classical version!) that grates me when it comes to China.
Its not as honest as the more traditional 'right-wing' china-hawk, whose obvious and admitted reason for disliking China and / or pushing for confrontation is grounded in profoundly nationalist thinking. Say what you want but that's an honest take and one that's easy to respect.
This particular brand of liberal though (lived in china, likes chinese culture, 'disappointed' that china hasnt become a liberal version of the US, etc etc) is just irksome. They want a china to be a larger version of japan - culturally cute and attractive and a good source of la mian and poems and ink-paintings and all that other harmless gunk. But crucially it needs to be geopolitically neutered...just to be safe!
Its such a dishonest view, wrapped in endless contradictions. And the insecurity really comes out at times....'WE have escalation dominance! Not you! WEEEE do!!! We're gonna prooove it next tiiimmmeee!! Go get em Donald! You show em Scott!'
Genuinely quite sad to read. Especially because dependencies dont even begin and end with chips (good luck btw getting TSMC to cancel work for every one of its chinese customers without any blowback). Medicines, crucial industrial inputs, etc etc. It will get very very nasty for everybody involved if this truly moves into the upper ends of the escalation ladder (regardless of who has an extra rung here or there).
“Genuinely quite sad!”
What a stupid whiny bitch you are. You’re deranged if you’re an American. You’re a run of the mill cunt if you’re Chinese.
China is apparently entitled to everything they’d like from the United States? How dare America do anything that doesn’t support China’s interests!
Perhaps should be called the “Talk Down to China Podcast” @tkz1 spot on! Whenever someone is so insecure that they need to continually tell you they are the best in the world, instead of letting the facts speak for themselves, it tells you all you need to know about their position… It’s very ironic how they claim America’s superiority where as far as I can tell, without the very non American TSMC and their Taiwanese engineering prowess and the nonAmerican ASML they would hardly have any chips at all. Well, there is that American chip powerhouse Intel… not too mention that much of the critical hardware to buildout all the AI infrastructure is coming from China as they are unable to produce it in the USA. So yes, a little humility and self reflection could come in handy! Given the trajectory of US stem education and the reduction in foreign talent, it’s hard to imagine the US being competitive globally 20 years from now. The US is similar to Great Britain in 1905, their best days were far behind them but arrogance was at its peak!!
Watching you little china loving twats pretend it is somehow immoral for the United States to chose not to be an active supporter of China’s geopolitical ambitions is comical.
Go fuck yourself you retard. Or if you’re a Chinese propagandist - go fuck yourself you stupid bitch.
We have a security imperative to get slavery out of our supply chain. Communist China is the most expensive option. We just need to stop capture deletes from making us foot the bill.
This is one of the most thorough hardware net assessments I've seen on the China-US AI semiconductor competition. The math here is compelling - if Huawei needs to make 200 million chips by 2027 to match Nvidia's aggregate compute, that would require essentially replicating TSMC's entire production capacity devoted solely to AI chips. That's not happening with current fab constraints and yield issues at 7nm. What really stands out is how the quality gap is widening, not narrowing. The 26x performance differnce projected for 2027 is staggering. This contradicts the common narrative that China always catches up through scale and iteration. The key difference you highlight - that this is the most complex supply chain in human history requiring EUV and advanced packaging tools they can't access - makes indigenization fundamentally harder than solar or EVs. I'm curious about the HBM constraint you mention. If CXMT can't scale HBM3 production, that becomes the binding constraint even if they solve logic die fab capacity. The policy implications are clear: maintain the tooling restrictions and don't ease up on controls just because headlines suggest breakthroughs. Great analysis.
The quantitative framework you laid out is incredibly clarifying - when you run the actual math on wafer production needs versus TSMC's total capacity, the scale of China's challenge becomes stark. The 200M chips needed by 2027 to match Nvidia would require essentially replicating an entire TSMC across all its production devoted solely to Ascends. What I find most compeling about this analysis is the insight that multiple exponentials are working against China simultaneously - compute demand growth, quality improvements from US side, and their constrained production scaling. The HBM bottleneck you mentioned is underappreciated - even if they could scale logic die production, the memory bandwidth constraint alone would be crippling. This is fundamentally different from solar or EVs where China could brute-force scale to dominance.
This quantitative analysis is exactly what's needed to cut through the noise. The math on China needing 200 million chips by 2027 to equal Nvidia is startling - that's essentially building multiple TSMCs from scratch while stuck at 7nm. What stands out is how the compounding effects work against China: they're not just behind on quality (26x performance gap projected), they also face exponential demand growth that outpaces even generous production scaling assumtions. The HBM constraint adds another critical bottleneck that often gets overlooked in breathless media coverage about Huawei breakthroughs.
This is an exceptionally thorough quantative breakdown of the production capacity gap. The math on the 200 million Ascends needed by 2027 to equal Nvidia's output really drives home how challenging China's position is. What particularly stood out is the point about competing with 'the flagship of global capitalism' - this isn't like solar or EVs where the West ceded interest. The entire weight of global capital is behind TSMC and Nvidia, creating compounding advantages on both quality and quantity dimensions. The HBM constraint you identified is also critical and often overlooked in media coverage. Without domestic HBM3 production at scale, even if SMIC can fab more 7nm logic dies, those chips will be severely bottlenecked. Great conversation.
The production math here is really compelling and often missing from the breathless headlines about Chinese chip breakthroughs. The distinction between designing a chip and manufacturig it at scale is crucial. What strikes me most is the exponential demand curve for AI compute outpacing China's ability to scale domestic production even if they solve some of the technical hurdles. The 50% rule change is interesting policy architecture - it's much harder to play shell games when subsidiaries are automatically covered. The HBM constraint is something that deserves more attention in these discussions since memory bandwidth bottlenecks can negate other advances.
Fantastic deep dive into the semiconductor dynamics. The quantitative analysis comparing Nvidia's 7-8M chips in 2027 versus Huawei's projected 600K chips really puts the scale gap into perspective. What I found particularly compelling is the discussion about HBM as the critical bottleneck - this is often overlooked in mainstream coverage but represents a fundamental constraint that can't be engineered around quickly. The historical parallel to Soviet supercomputer restrictions is apt, though as you note, we're now dealing with commercially available technology rather than specialized lab equipment. The 50% rule change from BIS is also significnt - closing subsidiary loopholes will make enforcement substantially more robust. Great work bringing actual math to a debate dominated by headlines.
The quantitative breakdowns here are exceptionally helpful for cutting through the media hype. The math showing Huawei would need roughly 200M chips by 2027 to match Nvidia's aggregate compute (versus their projected 600k-700k production) really crystallizes how severe the gap is. What strikes me most is the dynamic you highlighted about multiple exponentials working against China simultaneously - they need to close gaps in chip design quality, scale production quantity, AND do all this while demand itself is growing exponentialy. That's basically impossible without access to the upstream tooling ecosystem.
This is an incredibly thorough analysis that cuts through a lot of the noise. The distinction you make between quality and quantity gaps is crucial - it's not just that Huawei is making inferior chips, but that they would need to ramp production by 60-70x over two years just to keep pace with the current gap, let alone close it. What strikes me most is the point about indegenization. China has succeeded at this repeatedly across industries, but semiconductors are fundementally different because the tools themselves are the constraint. You can't reverse-engineer an EUV machine easily when it represents the pinnacle of human manufacturing capability. The export controls are working precisely because they target the choke point where China's traditional strengths don't translate.
This is an excellent deep dive into the hardware net assessment. The point about supercomputers now being 'available in a box off the shelf from Nvidia' really captures the unique challenge. The math Chris lays out is compelling - even if Huawei doubles production to 600k chips, that's still dwarfed by Nvidia's 5-7 million, and the quality gap is increasing by 6-7x over two years. The 50% ownership rule change is also a significnt enforcement improvement that should make it harder to play shell company games. Really appreciate the quantitative rigor here!
Really helpful breakdown of why the gap isn0t just about TFLOPs. Two things that seem underpriced: (1) HBM supply chain lock-in0you can0t brute force bandwidth without stacked memory and advanced packaging yield, and China0s domestic HBM is years behind; (2) software+systems maturity0CUDA, compilers, libraries, and networking stacks that squeeze utilization out of clusters. Even if Ascend hits H100-ish peak specs on paper, sustaining performance at scale needs the whole stack tuned. Add node stagnation (~7nm) limiting density and power efficieny gains, and the math compounds against catch-up. Curious how much CloudMatrix can offset interconnect/topology limits before power and capex walls bite.
Tech things aside, any bully needs to be put to rest.
After separating the calm goers in the herd!
Chris's analysis is probably the most neutral and clear-eyed on the semiconductor race that I've read. We would be foolish to underestimate our lead over China in compute, but just like the situation in China where people dare not tell bad news to the party leadership, I wonder if the current structure and style of the current US administration might make it similarly difficult for the truth to get to the ultimate decisionmakers. Hence this administration's vacillation on export control policy.
One implication from Chris's analysis though is that it seems China may have only a narrow window to stop the US from gaining an unassailable lead in compute and thus an unassailable military advantage. We know the CCP can do so by taking control or destroying the world's most advanced fabs that's only 100 miles away from China's coast. I am not sure if either the US or Chinese leadership have realized that the military balance of power may have already shifted decisively to China, given that the Chinese have a huge advantage in industrial capacity, have a significant technological advantage in battery technology and electronics that allows them to produce far more drones and missiles, and a powerful tool for kneecapping the US defense industry in the short term with rare earth controls. However this narrow window of opportunity for them could close quickly once the US's compute advantage kicks in in a few years and AI is fully integrated into the US military. If the Chinese leadership realizes that time is not on their side, I wonder how they would react. As Paul Atreides said in Dune, the power to destroy a thing is the absolute control over it.
All this to say, if we keep climbing the escalation ladder, at least during this narrow window, I think we may not like what we find when we reach the top. Even if the Trump administration is making the wrong policies now, if it could lure the CCP leadership into a false sense of security until the military balance of power shifts, those policies may end up becoming the correct ones for the long term.
Nice pod. I’m a little conflicted, I had other people in the semi supply chain say that Huaewei was actually closer to 1M dies this year. Still this 5x doesn’t close the gap, but it is obvious that have huge error bars.
stay tuned for the tool where you can plug in different assumptions!
tldr is that if huawei isn't sandbagging its roadmap theyre ngmi