China vs USA – Rounding Error

Huawei will produce a tiny fraction of global compute.

• Huawei will increase AI chip revenues by 60% this year, but the tokens per dollar that its new chip produces are so low that the total amount of compute it will produce will be a tiny fraction of what China needs.
• The FT (see here) is reporting that Huawei’s AI chip revenues will increase by 60% YoY to $12.5bn in 2026 as a result of launching its latest chip, the Ascend950PR.
• The Ascend 950PR is unique in that it is less performant than its predecessor (Ascend 910C), which is something that I have never seen before in 30 years of following the semiconductor sector.
• According to Huawei’s data, which it disclosed at launch, the Ascend 950PR will produce about 2.0 PetaFLOPs (FP4) per chip, whereas its predecessor produces 2.4.
• I suspect that this is due to the 950PR being produced on SMIC’s 7nm multi-patterning process, while the 910C was made by TSMC procured through shell companies.
• The current mainstream product from Nvidia (B300) will produce 30 PetaFLOPs (FP4) per chip, while Vera Rubin will produce 50.
• That means to produce the same amount of compute a data centre will need to buy 15 Ascend 950PRs to match 1 B300 chip and 25 chips to match 1 Vera Rubin chip.
• If I assume that a Huawei chip is ¼ the price of an Nvidia chip, then I am still spending between 3.8x and 6.3x the amount of money to buy the equivalent compute from Huawei than I would from Nvidia, even including its large gross margins.
• Furthermore, it will take up a lot more space, and because it is a 7nm chip, it will consume more power for every token that it produces.
• It is due to these factors that RFM Research has concluded that Huawei’s current line-up is uneconomic to run and that no rational investor would finance a Huawei data centre.
• The other issue is that $12.5bn is dwarfed by the $300bn+ that Nvidia is expected to sell in data centre chips this year, and when I adjust that for the difference in performance, it becomes even smaller.
• $12.5bn spent with Huawei will generate the same amount of compute as $3.3 bn – $2.0bn spent with Nvidia, meaning that in compute production terms, Huawei will produce around 1% the compute that Nvidia will.
• Huawei is China’s market leader, and this demonstrates that the lack of compute is going to be an increasingly difficult problem to solve going forward.
• When I ask any Chinese AI company what its biggest constraint is, the answer is always a lack of compute.
• This is why I think that China will remain dependent on importing compute from overseas data centres, which do not have import restrictions, but this leaves it open to further action from the US Department of Commerce.
• I expect that the Department of Commerce will target this practice of compute import the next time it updates its rules in Q3 or Q4 of this year.
• The net result is that while Chinese companies will buy Huawei chips for deployment in China, it is unlikely that anyone else will meaning that China will not gain market share overseas.
• Hence, any revenues that Nvidia, AMD, et al lose in China, they are more than likely to make it up in overseas markets.
• This is why I continue to think that for now, the US has the advantage.