USA vs. China – No Contest pt. II

Leading and winning are two very different things.

• Jensen is right in that US models are no longer ahead of Chinese counterparts, but with such uncompetitive hardware, China is very unlikely to win the race, no matter how many subsidies China gives to its AI companies.  
• Jensen Huang has made his clearest statement yet about the state of the race, warning that China is going to win due to its lower energy costs and looser regulations, but the actions of the Chinese state do not support this view.
• There is no doubt that the Chinese models are extremely competitive when it comes to raw performance, which combined with the fact that there is a high concentration of developers in China, puts to bed any idea that China is behind.
• There is also fairly widespread adoption of Chinese open-source models outside of China, which is one of the reasons why Meta Platforms is having difficulty, as its position as the king of open source has been usurped.
• If we take this and add much cheaper electricity and lower regulation, it is easy to conclude that China will win the race, as Jensen has stated.
• However, it is one thing to produce cutting-edge AI, but it is quite another to be able to produce it cost-effectively in a way that it can run at scale and create a thriving business.
• This, in my opinion, is where China’s proposition falls to pieces because when it comes to efficiency, China is miles behind, and I suspect that it is not going to catch up in the foreseeable future.
• This is because the Cyberspace Administration of China (CAC) is increasingly enforcing China’s policy of self-reliance when it comes to critical technology.
• Consequently, AI companies in China have been “encouraged” to cease testing foreign-made silicon chips and to only use homegrown variants.
• China has recently sweetened the deal by offering cheap electricity to its data centres in yet another sign that Chinese silicon is far from competitive.
• If it were just as competitive, then Chinese companies would choose local variants naturally and not have to be “encouraged” or be offered subsidies to do so.
• There are a few areas where China is behind, some of which can be fixed and some of which can’t.
  • First, efficiency: which will not be fixed anytime soon and is related to the fact that China does not have access to leading-edge semiconductor manufacturing.
  • This means that homegrown chips are made using the 7nm multi-patterning technique using older equipment.
  • Yields are lower and the chips are larger, which means that these chips are much more expensive to produce than those made at the current leading edge.
  • The impact of this is more profound than I had expected, as Huawei’s numbers indicate that its AI SuperPod, which will ship in H2 2026, is 32x less efficient than the equivalent from Nvidia (see here).
  • This means that on the global stage, Huawei is very uncompetitive, and no rational buyer will select it over Nvidia without some other mitigating circumstance.
  • This is not going to change any time soon, as there is no discussion over whether advanced semiconductor manufacturing equipment should be sold to China.
  • Second, reliability: where Huawei’s Ascend chips still have a much lower mean time to failure than Nvidia or AMD.
  • When AI chips are networked together and work as one, this rapidly becomes a problem as a failure disturbs the entire system.
  • One of the reasons why China’s capacity utilisation of its AI chips is lower than elsewhere is due to this problem, which increases both the time and cost it takes to train and inference AI models.
  • This is something that can be fixed, and given the ingenuity that Chinese companies often display, I suspect it will be greatly improved pretty soon.
  • Third, software: where Chinese AI development software is far less mature than Nvidia’s CUDA.
  • This means that it is harder to work with and still has multiple problems that mean that it is more difficult and more expensive to create and run AI models.
  • Again, this is something that I think will be fixed, and it is already improving as the Huawei software is currently described as “usable”.
• The net result is that there are lots of places where Chinese AI development can be improved, but, for the foreseeable future, it is going to be much less efficient than Western versions, and the gap will grow with every generation.
• Hence, while it is clear that China can produce AI that performs, it cannot do so economically, which, at the end of the day, is likely to be the deciding factor.
• Historically, Chinese technology has been just as good and cheaper to purchase, which is one reason why Huawei saw such success with its 4G and 5G basestation offerings across the world.
• However, now that Chinese technology based on Chinese silicon will be just as good but more expensive, the Western options will be cheaper to purchase.
• I do not think the Chinese economy is strong enough to enable the mass subsidisation that it would take to make Chinese AI cheaper, and so, I think it is likely that it will remain uncompetitive when compared to offerings from the West.
• This is not going to change unless there is a change of heart by the Chinese state to allow foreign silicon to be used once again, and, at the moment, this looks unlikely as China does not want to again be dependent on the technology stack of a competitive power.
• The net result is that while the Chinese can compete on performance, they can’t compete on economics, and it is the economics that will decide which version non-affiliated countries adopt most of the time.
• This will be crucial to the ideological struggle that is being played out between China and the West, and at the moment, I continue to think it is the West that has the advantage.