USA vs. China – All about hardware

The US hardware advantage dwarfs all else.

• Another study from Bloomberg confirms that China is continuing to close the gap when it comes to performance, but all of this depends on open-source models using non-Chinese silicon, exposing a dependency that could be exploited with the stroke of a pen.
• Bloomberg Intelligence (see here) has published the results of a study that it has conducted that concludes that Chinese model performance is now just 3% behind its US rivals following the release of DeepSeek V4.1 Flash.
• Similar research conducted in Q1 26 and May 26 put the gap at 15% and 9%, respectively, implying that there have been significant advances from China relative to their US counterparts.
• This analysis is conducted on a wide range of model use cases such as “reasoning”, coding, math, data analysis, and agentic coding.
• This is also highly significant, as historically, China has relied on distillation to improve its performance, but since this is being actively blocked, China can no longer rely on this.
• Hence, it has clearly developed its own methodologies and techniques to build and train models, making it a formidable adversary.
• RFM Research (see here) broadly agrees with the improvement in performance and ranks China as ahead of the US in a few areas such as its dominance of open source, token usage, power availability and its citizens’ willingness to adopt AI.
• This all adds up to an impression that China will soon take over the lead in AI, which is used as an argument for the ineffectiveness of export controls.
• However, the evidence points to the opposite as China remains almost completely dependent on foreign silicon to train and run inference on its models, leaving it wide open to action from the US Department of Commerce.
• For example, Bloomberg’s own reporting suggests that DeepSeek is still training on Nvidia, as Huawei’s software stack and interconnect are still not mature enough to complete a full training run.
• When DeepSeek released a version of its kernel for Ascend, it made no mention of a training run using V4, implying that this had not taken place.
• Furthermore, a lot of questions have surrounded a data centre in Inner Mongolia which has recently purchased 160,000 Ascend 950DT accelerators.
• The US has alleged that this data centre is stocked with illicit imports of Nvidia Blackwell systems, and while there is no immediate way to tell whether the systems are illicit, I am pretty sure that the Nvidia systems are there.
• The purchase of 160,000 Ascend processors is part of a scale up of this data centre to 1GW, which together will consume around 300MW of power and produce around 312 ExaFLOPS of compute.
• This will require 3,125 racks of space in the data centre, but if the data centre operator were to use Nvidia Rubin, it would only need to buy 88 cabinets of Vera Rubin, which would consume 11MW of power to produce the same 312 ExaFLOPS of compute.
• Hence, for a data centre operator, choosing Huawei looks like financial suicide unless one has been able to get massive state subsidisation and probably access to free electricity.
• This is why Chinese AI operators are increasingly looking to source their compute from overseas, as by importing the compute, they can still use Western silicon.
• Alibaba negotiating for a data centre in Spain is the latest example.
• Hence, China can close the gap on the USA in terms of AI performance as long as it is trained and run on non-Chinese hardware, creating an uncomfortable dependency.
• Furthermore, the performance and economic gap between Western silicon and Chinese silicon for AI data centres is getting wider with every generation creating the possibility that China becomes more dependent on Western silicon to keep up not less as is widely mooted.
• RFM Research has long concluded that China will increasingly become dependent on importing compute from data centres overseas running Western silicon, which is at risk from US export regulations which could ban the imports at any moment.
• So far, it has chosen not to take this route, but the possibility remains for as long as the dependency continues to exist.
• This is why I believe that the US and the West still has a large advantage over China, which is not going to change any time soon.