Not only is Jensen an american with an american company, he also is also almost single handedly keeping the taiwanese economy alive and critical on an international level. It's ridiculous to claim he's some Chinese shill. It's far more likely Google and Broadcom switch to Intel than Jensen from TSMC.
DeepSeek V4 specs leaked this morning — Yifan Zhang (Princeton/YFZ.ai, 11.9K followers) posted the architecture: 1.6T params (MoE, 6 active in 384 experts), DSA2 attention (NSA+DSA, head-dim 512), 1M context, text-only. Chinese ML sources are confirming it's being trained entirely without CUDA.
If the architecture holds, it validates your "chip exports as red herring" argument from the other direction — China's top lab isn't trying to get more H100s, they're building the 1.6T model around Huawei compute by design.
is it unreasonable now to think someone is talking their book? i admire Jensen immensely but you can't have it both ways on compute doesn't matter and also it's the most important thing for American companies (see the Cursor-xAI deal)
Some of your WarTalk guests' prescriptions for how to respond to Mythos don't make any sense to me. Why are hard-wired networks safer against something that is really good at finding security bugs? Those sound like responses to cryptography being broken (e.g. what might happen with quantum computers to some cryptography algorithms). Yes air-gapped networks make sense.
I lived in Kyrgyzstan and one thing that shocked me was how all the water regulators/ dykes were hand cranked. I finally asked my host brother one day why they didn’t just automate it and he laughed at me. He then said it was classic that the American wanted everything to be a computer. They didn’t connect it so that we (the Americans) couldn’t hack it.
I think the deeper issue here is that equipment controls are harder to treat as a fixed lever than chip sales. They are not a simple on/off switch, but a dynamic industrial contest that evolves across firms, layers, and time. That is why Huawei’s Ascend progress matters so much. It is not just another domestic AI chip. It is helping generate demand across a much broader domestic stack, from design and manufacturing to packaging, servers, software, and deployment. So I agree upstream constraints matter a lot, but I am less convinced they can function as a durable strategy of absolute denial. More likely, they slow China down, raise costs, and buy time, rather than permanently lock the system in place.
Better said than me yelling in the comments like a old man yelling at clouds. Honestly I think this is one of the most underrated debates in our modern history/life's we are not taking seriously enough. Great write up!
Not only is Jensen an american with an american company, he also is also almost single handedly keeping the taiwanese economy alive and critical on an international level. It's ridiculous to claim he's some Chinese shill. It's far more likely Google and Broadcom switch to Intel than Jensen from TSMC.
DeepSeek V4 specs leaked this morning — Yifan Zhang (Princeton/YFZ.ai, 11.9K followers) posted the architecture: 1.6T params (MoE, 6 active in 384 experts), DSA2 attention (NSA+DSA, head-dim 512), 1M context, text-only. Chinese ML sources are confirming it's being trained entirely without CUDA.
If the architecture holds, it validates your "chip exports as red herring" argument from the other direction — China's top lab isn't trying to get more H100s, they're building the 1.6T model around Huawei compute by design.
Not mentioning the propaganda of if you buy enough Nvidia chips you can build your own 'Sovereign AI" they tell countries.
is it unreasonable now to think someone is talking their book? i admire Jensen immensely but you can't have it both ways on compute doesn't matter and also it's the most important thing for American companies (see the Cursor-xAI deal)
Some of your WarTalk guests' prescriptions for how to respond to Mythos don't make any sense to me. Why are hard-wired networks safer against something that is really good at finding security bugs? Those sound like responses to cryptography being broken (e.g. what might happen with quantum computers to some cryptography algorithms). Yes air-gapped networks make sense.
I lived in Kyrgyzstan and one thing that shocked me was how all the water regulators/ dykes were hand cranked. I finally asked my host brother one day why they didn’t just automate it and he laughed at me. He then said it was classic that the American wanted everything to be a computer. They didn’t connect it so that we (the Americans) couldn’t hack it.
I think the deeper issue here is that equipment controls are harder to treat as a fixed lever than chip sales. They are not a simple on/off switch, but a dynamic industrial contest that evolves across firms, layers, and time. That is why Huawei’s Ascend progress matters so much. It is not just another domestic AI chip. It is helping generate demand across a much broader domestic stack, from design and manufacturing to packaging, servers, software, and deployment. So I agree upstream constraints matter a lot, but I am less convinced they can function as a durable strategy of absolute denial. More likely, they slow China down, raise costs, and buy time, rather than permanently lock the system in place.
Better said than me yelling in the comments like a old man yelling at clouds. Honestly I think this is one of the most underrated debates in our modern history/life's we are not taking seriously enough. Great write up!