Ray Wang has posted a graphic from Epoch AI that says the total computational power troughed below 40% in 2021 and is now 75%. Meanwhile, the percentage in China dropped from over 40% to about 20%.
Meanwhile, the U.S. has massive new AI infrastructure announcments:
In fact, Epoch AI separately on their website notes that the total computational power in supercomputing clusters worldwide doubles every 10 months. For example, Trump can go to Saudi Arabia and offer a massive AI infrastructure investment, but we haven't seen Xi do the same thing, perhaps because of limited manufacturing capacity by SMIC to make Ascend chips. If this is true, the export controls sadly may be working. China is getting starved of raw compute, which is growing exponentially.
A Tencent executive recently said lack of compute was one reason hindering the diffusion of AI in China. This affects China's AI models too, as Deepseek and Qwen have comparable performance to the best models but slower response times per token, according to some metrics I looked at.
China's only options at this point are either ramp up manufacturing somehow, or try to blockade Taiwan to at least slow down the AI infrastructure buildout of the U.S. The latter would be a nuclear option that would burn through a lot of goodwill but might be a desperate last chance effort. Many American chips are already made in Arizona though.