China vs. USA – The Problem of Scale

China cannot build enough compute

• China is making big plans to build AI infrastructure, but the scale and the performance of the technology it will build are not nearly high enough to compete without importing compute from overseas.
• China’s National Development and Reform Commission is drawing up plans to build a network of connected data centres across China that will be almost entirely sourced from local vendors.
• The plan is to spend CNY10tn ($1.5tn) by the end of 2030 (of which CNY7tn will have been spent by the end of 2026) on what China refers to as “6 networks” (see here)
• These include power grid, water network, AI compute, logistics, communications and underground pipes.
• These are all deeply interrelated, but it is clear that the bulk of the remaining CNY3tn ($447bn) will be spent on AI.
• This equates to roughly $150bn per year until 2030, which, when one takes into account what its arch-rival is doing and the performance it can achieve, this is not nearly enough.
  • First capex: where assuming that all of the remaining investments are spent on AI compute, China remains very far behind.
  • The USA is expected to spend around $750bn building data centres this year, and if Nvidia and the memory companies are to be believed, this will still be higher in 2027.
  • Alibaba, the Chinese cloud leader, spent $18.3bn last fiscal year (March 2026), meaning that with large increases, the private sector in China might spend $100bn in 2026.
  • Altogether, the state sector and the private sector might spend $250bn in 2026, which is still 3x lower than the USA.
  • Given where Huawei and SMIC are in their build-out of chip manufacturing capacity, I have real doubts whether they will be able to meet this demand, but for argument’s sake, I will assume that they can.
  • Second, performance: which is where the real differential is to be found.
  • Using Huawei’s publicly stated data for the performance of its Ascend 950 chip, one standard cabinet will produce 0.1 ExaFLOPs while consuming 100kW of power, while the Ascend 960 will produce 0.273 ExaFLOPs per rack while consuming 120kW of power.
  • Compare this to Nvidia’s Vera Rubin, which will produce 3.6 ExaFLOPs per rack while consuming 125kW and the issue immediately becomes very clear.
  • Even assuming that Huawei’s chips cost ¼ of Nvidia’s and the other infrastructure is much cheaper in China, the difference is so large that the Huawei system is far more expensive to build based on an equivalent amount of compute.
  • RFM has estimated that to build 1GW equivalent of Vera Rubin, which costs around $60bn, a data centre operator using Huawei would need to spend $120bn, which would consume around 9GW of power.
  • Power is much less of a problem in China as it already has triple the generation capacity of the USA and is building much more at a much faster rate.
  • However, performance per dollar invested is a problem as China does not have an endless source of cash to draw on.
• Putting all of this together implies that by spending a third of what the US is spending and getting half the amount of compute per dollar invested, leads me to conclude that China is adding compute at a rate that is around a sixth of what the USA is.
• This means that rolling out AI services to 1.4bn users is going to be a real problem and could result in China falling behind unless it can get compute from elsewhere.
• It is through importing compute from Western-powered data centres overseas that I expect China to plug the gap, but this is not without risk.
• 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 the 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.