China vs USA – Economic Illusion

GLM-5.2 can only be competitive on non-Chinese silicon.

• China is once again causing industry gyrations with its new model but this model is either running on Chinese hardware and is much more expensive to run, or is running on Western chips, where it could be cheaper than the frontier models.
• Z.ai (formerly Zhipu AI) has released a new model, GLM-5.2, which is a 735bn mixture of experts (MoE) model that is almost entirely open source.
• The model is available everywhere under the MIT licence, which allows the user to do pretty much whatever they want with the model without having to pay Z.ai royalties.
• What is not open is the training data set and the techniques that were employed to train the model to its current level of performance.
• On performance, it does very well, coming within touching distance of frontier models on the core benchmarks, underlining that despite silicon shortages, China is still very capable when it comes to AI development.
• However, it is here that the claims being made start to diverge from reality.
  • First, Training: where it is claimed that the whole model was trained from scratch using Huawei silicon.
  • This is one that I tend to believe as necessity is forcing the Chinese to innovate with less, and the Ascend chips are capable, albeit not competitive.
  • Furthermore, it looks very likely that the GLM-5.2 was trained using the Ascend 910B. which RFM has concluded was built by TSMC rather than Huawei.
  • This was part of the illicit 2-3m wafers that Huawei managed to procure, which have subsequently been exhausted, explaining why the 910C (2026 chip) has a lower specification than the one that came before it.
  • Huawei’s software is a lot better than it used to be, and so there is no reason why the training runs could not have been completed using Ascend.
  • Second, Inference: which is a very different issue.
  • With the use of AI increasing very quickly, the market is already skewing very heavily towards inference, and it is here where the economics of AI will be determined.
  • A lot of fuss is being made about how GLM-5.2 offers almost as good intelligence but for a fraction of the price, but for copies of GLM-5.2 running on Chinese silicon, this is simply not the case.
  • RFM has concluded that Chinese silicon costs 3x as much to roll out the same amount of compute compared to Nvidia, meaning that GLM-5.2 will be 3x more expensive to inference on Huawei than it is if it is run on Western chips.
  • This means that Z.ai will be competitive with OpenAI et al, when it is exported from China and run in non-Chinese data centres.
  • However, GLM-5.2 running on Huawei hardware will be much more expensive, meaning either higher prices or heavy subsidisation.
  • This does assume that the efficiency of frontier models is the same as GLM-5.2, which is clearly not the case today.
  • GLM-5.2 is smaller than the models it competes with and has been designed to be as efficient as possible from the outset, which I think will help it to close the gap.
  • However, I also think that the difference between Western silicon and Huawei is so large that the only way GLM-5.2 can compete is by using Western silicon.
• The advent of GLM-5.2 is good news for the industry as economics are beginning to become important, and this will force Anthropic, OpenAI and so on to look more closely at how they can run inference more efficiently.
• Hence, I think that GLM-5.2 will see a lot of use outside of China as it can be downloaded, run locally and modified for any task that the user wishes, which is something that the frontier models do not support.
• However, it is not the beginning of a very cheap, very powerful model run out of China that will trigger a wave of compute to move to China.
• I think that China has real problems with running AI at scale locally, and so it will continue to import compute from overseas from data centres running on Western silicon, which is a risky proposition.
• The US is well aware of this workaround, and I suspect that when the Department of Commerce updates its rules in Q3 or Q4 of 2026, it will target this practice.
• This will cause real consternation and further constriction of the compute capacity in China, meaning that China will fall further behind.
• Its models are still likely to see a lot of use outside of the USA, but China’s ability to subsidise compute to drive usage is already limited and may become much more so.
• This is why I continue to think that the West and the USA in particular have the advantage in the AI race that is currently being run.