China AI – Capacity Problem

Almost as big as the technology problem. 

• There is no doubt that demand for AI in China is strong, but China’s inability to produce enough compute capacity remains a large problem and one that is unlikely to be solved anytime soon.
• China has a big advantage in terms of abundant electricity and a population that is keen to adopt AI, but its inability to produce enough of the semiconductors and components needed is going to remain a major bottleneck for some time to come.
• At the SEMICON China 2026 conference, executives were fairly upfront about the outlook, with ACM Research stating that further developments depended on next-generation semiconductor tools that have yet to be developed.
• Lee Haimling of Chongqing Xinlan Microelectronics admitted that while China remains competitive in consumer chips (thanks to TSMC), it remains 5 to 10 years behind in automotive and data centre semiconductors.
• The press is also full of stories of how the Chinese supply chain can’t keep up with demand, which is causing a lot of investor interest in the Chinese supply chain.
• The fact that capacity constraints are appearing throughout the supply chain was also discussed in the halls and coffee shops of the trade show, but the scale of the problem has not really been looked at closely.
• There are two issues:
  • First, Chinese capex: where Chinese companies are spending a tiny fraction of their US counterparts on data centre capacity.
  • For example, around $750bn is expected to be spent on data centre capacity in the USA in 2026, while China will spend around $72bn or 90% less.
  • With 1.2bn users all keen to try the latest and greatest AI service, this does not seem to be like anything like enough in order to roll out AI services at scale.
  • Second, chip performance: where the best homemade chips on offer produce much less compute in totality, meaning that the buyer needs to purchase many more of them to get the same performance as they would from Nvidia, AMD and so on.
  • Huawei’s data for the Ascend 950PR gives performance of 2.0 PetaFLOPS/chip FP4 peak performance while the Blackwell B300 will produce 30 PetaFLOPS/chip at FP4.
  • This means that a data centre operator will need to install 15x more Huawei chips to get the same level of compute.
  • This is where the problem begins, as Huawei can’t make 15x more chips at the moment, and so its ability to produce compute is constrained both by its capacity to make chips and the disadvantage that they have in peak performance per chip
  • This is why, RFM forecasts that in a 3-year face off with Nvidia, Huawei’s share of total compute production between 2026-2029 will be 3% compared to Nvidia on 97%.
• The net result is that China is going to be far more compute-constrained going forward than the USA even with its energy problem.
• The example of it taking DeepSeek a year to produce a single new model, compared to the frontier labs, which produced 7 each in the same time, makes this point quite clearly.
• This is why I think that China is going to be increasingly dependent on importing compute from overseas to roll out AI services at scale in its home market.
• This means running its services in data centres in neighbouring countries, which may lead to some degradation in responsiveness, but it is much better than nothing at all.
• However, it creates a target for the US Department of Commerce when it updates its export rules in QA3 or Q4 this year, and if this practice were blocked, I think it would cause real consternation in China.
• This is why, despite the energy scarcity and the protests, the USA still has the advantage in the AI race that is currently being run.