China vs. USA – The Hidden Edge

Economics always wins.

• Chinese AI looks competitive with the West on paper, but when one looks at the real economics, I don’t agree with the view that China is catching up, as from my perspective, it is falling behind.
• The Wall Street Journal has run an analysis (see here) that concludes that China is rapidly eroding the West’s lead, and while its conclusions are basically correct, I think it misses a key nuance that flips this view on its head.
• There has never really been any doubt in my mind that China would produce AI that performs just as well as anyone else, but where I have doubts is whether it can do so economically.
• With the release of DeepSeek and now all of the other Chinese models, the illusion that the Chinese were behind in AI has been well and truly shattered.
• This is evident in the WSJ’s analysis which puts DeepSeek in 3^rd place, just slightly behind Gemini and ChatGPT and Alibaba, Tencent & Minimax around 10^th in the pecking order.
• The top of the table is also very tight with just 7% separating the top 11 models in this analysis.
• This again demonstrates that progress in terms of improving the performance of language models has slowed to a crawl, which is how the late entrants have been able to catch up so quickly.
• However, it is not just performance that is going to make the real difference, but best performance delivered at the lowest price.
• The WSJ identifies that the Chinese have gained some traction in South East Asia, Middle East, Latin America and Africa, largely on the basis that Chinese models like DeepSeek are substantially cheaper to operate and just as good.
• There are two ways in which this has been achieved:
  • First, innovations: where DeepSeek was the first to make use of certain innovations that substantially reduce the compute overhead required for training and inference.
  • However, all of these innovations have been quickly incorporated into the models of others, and now I don’t think that this is a competitive advantage any longer.
  • For example, Llama 4 makes use of the same innovations as DeepSeek and Google claims that it has been cheaper to operate than DeepSeek the whole time.
  • Regardless of the veracity of Google’s claims, I don’t think that the technical edge that DeepSeek had that made it cheaper to run exists any longer.
  • This means that anyone downloading and running DeepSeek or Alibaba Qwen in their data centre will see similar costs to operate as if they used Llama or Mistral.
  • Second, subsidisation: which is where China subsidises the prices that the AI companies charge to customers to access and use LLMs running in the model provider’s data centre.
  • In practice, this means that one’s AI service will be running in mainland China, which will not be a problem for local players, but in this climate, many others will think twice.
  • It is only when one runs in China, that the models are much cheaper and it is not because China has figured out how to make cheap compute, but because prices are being subsidised to get customers on board.
  • Not only are the Chinese competing with others like OpenAI, but also aggressively with each other.
  • This means that as long as one has no issue with running in China, this is by far the most economical place to run an AI algorithm today.
  • However, this can’t last, and at some point, prices are going to have to rise to a point at which the Chinese datacentres can make money.
• This is the point at which the real difference between the Western offerings and the Chinese offerings will become clear.  
• Chinese data centres are going to be built using Chinese chips, which are likely to be made at 7nm, which will be at least 2 generations behind Western competitors.
• This means that Chinese AI will generate significantly fewer tokens per dollar invested, which in turn will eventually force the Chinese to charge more per dollar than Western companies.
• This is the key nuance that I think the WSJ article misses, which is not a big surprise, as I get the impression that almost all of the rest of the commentariat also make this omission.
• In the long run, it means that when it comes to a choice for a non-affiliated country, the Western version will be cheaper than the Chinese one if one is also going to use Chinese hardware and the same if one is going to run it on Western hardware.
• Given that using Chinese standards comes with strings attached to Beijing and the Western versions much less so, this will make the Western version more attractive to a non-affiliated country.
• We have already seen some evidence of this as Saudi Arabia, UAE and potentially Qatar are already pivoting away from China, and others may follow.
• The net result is that while the Chinese can compete on performance, they can’t compete on economics, and it is the economics that will decide which version non-affiliated countries adopt most of the time.
• This is why I think the Chinese are behind, and there is very little scope to catch up, given that it has no access to cutting-edge hardware.
• This will be crucial to the ideological struggle that is being played out between China and the West, and at the moment, I think it is the West that has the advantage.