USA vs. China – The DeepSeek Effect

V4 shows the impact of US export restrictions.

• DeepSeek has launched V4, but the time that it took to release it, its performance, and the extreme limitations being placed on being able to access it, are a sign that the US export restrictions are putting the brakes on both the speed and scale of China’s rise as a global leader in AI.
• DeepSeek V4 is large (1.6tn parameters) and boasts a 1m token context window as well as pretty good performance and high compute efficiency.
• It is also pretty efficient as only 49bn of its parameters are active at any one time, making it much more efficient (albeit slower) when it comes to inference.
• However, what it does not do is close the gap on the frontier models from US companies, meaning that the closest China has come to challenging the US thus far is R3, which was launched in January 2025.
• There are also plenty of signs that US export restrictions in terms of AI silicon and chip-making equipment are finally having a meaningful effect.
  • First, time: where it has taken DeepSeek about 15 months to release a new version compared to OpenAI, Google and Anthropic, which have launched at least 7 each.
  • This is not because DeepSeek researchers are less productive or are of lower quality than anyone else, but because of the problems the company has been having with compute.
  • R3 was mostly trained on Nvidia, whereas DeepSeek has been under pressure from the Chinese state to use domestic silicon going forward.
  • Huawei’s silicon is described as “usable”, but the development platform that goes with it is not mature, and there have been lots of problems in getting even a single training run to completion.
  • This is one reason why it has taken so long for DeepSeek to be in a position to release its new model
  • Despite the availability of the new silicon, I suspect that a lot of the training of V4 was run on Nvidia silicon, either smuggled into the country or on datacentres in overseas locations.
  • With restrictions likely to be tightened further, it may take DeepSeek another year or more before we see V5.
  • Second, performance: where V4 is a large step forward from R3, but looks to be further behind the frontier labs than R3 was at the time of its launch.
  • R3 was within spitting distance of frontier labs when it launched and now based on the benchmarks, it looks like it is 3 to 6 months behind.
  • This is a result of the time element discussed above which in turn is caused by China’s extremely limited access to cutting edge silicon.
  • This implies that the US has a faster cadence of advancement, meaning that unless something changes, V5 will be even further behind when it launches.
  • One can also see this difference in pricing where DeepSeek API calls are priced at 1/10^th of API calls to Claude Opus 4.7, which is Anthropic’s latest and greatest model that is generally available.
  • It is 1/8^th of the price for GPT5.5 and ¼ the price of Opus4.6, again clearly pointing to large price premia being paid the closer one gets to the cutting edge.
  • Third, compute constraints: which I continue to argue lies at the heart of all of China’s problems, because when one looks at other factors such as talent or open source, China is extremely competitive, if not ahead.
  • There are signs of this everywhere, and this remains in my opinion the only factor that is holding China back from catching up with the USA.
  • DeepSeek even mentions in its materials that service capacity for V4 Pro is extremely limited, which supports my contention and RFM Research’s conclusion that China is going to have real problems launching AI services at scale.
  • This is due to China being stuck with an inefficient 7nm multi-patterning process, from which it cannot evolve, which results in its domestic AI silicon falling further and further behind the cutting edge.
• The net result is that V4 is an indication that the US strategy to slow China down is working, but there are still loopholes.
• The main loophole is the ability for Chinese companies to rent data centre capacity in countries that are not restricted from being advanced AI silicon and running their models there.
• The fact that most Chinese models are open source greatly helps here, as there are no restrictions on the export or inference of these models outside of China.
• RFM Research has concluded that the importation of compute is the only way that China will be able to launch performant AI services at scale because it cannot produce enough compute domestically.
• This means that should the USA close the loophole and close it efficiently, then China will be further hindered, and its pace of development further slowed down.
• Hence, this is what I expect from the Department of Commerce the next time it updates its rules, which I am expecting in Q3 or Q4 of this year.
• If successful, this would be a hammer blow to China’s AI ambitions and would have the impact of closing a major loophole in the strategy of containing China’s rise as a technological and geopolitical power.
• This is why I continue to think that for now, the US has the advantage.