@samuelszuchan: Chinese AI models are about seven months behind the best American ones, even though China has roughly 5% of the AI chips the US does. The seven-month figure comes from Epoch AI, which has tracked the gap since 2023. The 5% is a rough estimate from Anand Kannappan, who runs the evaluation company Patronus AI. If the US has 20 times the chips, you would expect its models to be more than seven months ahead. The lead stays that small because training a model has a finish line. You borrow enough chips, run the training, and you are done. Serving millions of users every day never finishes, and Chinese labs do not have the chips for that. Moonshot (a Chinese AI lab) launched its flagship model in July, suspended new subscriptions after three days because it ran out of computing power, and eight days later published the model for free so anyone could download and run it themselves. DeepSeek (a Chinese lab spun out of a Hangzhou hedge fund) ran its entire global service at a peak of 278 nodes, about 2,000 chips. MacroPolo's Global AI Talent Tracker (which surveys the major machine-learning conferences) finds that 38% of the world's top-tier AI researchers did their undergraduate degrees in China against 24% in America, and that 72% of the China-educated ones now work at American institutions. China trains more top AI researchers than any other country, most of them leave to work in America, and Chinese models still trail by only seven months. The chips China is short on are all printed by one kind of machine: an EUV lithography scanner. One Dutch company builds it, at about $380 million and 165 tons apiece. Under US pressure, the Dutch license to send one to China lapsed in mid-2019, and none has been issued since. ASML tracks all 314 operational machines on earth for connectivity, so it knows where every one of them is.

Sam Szuchan

Open In TikTok:

Region: US

Thursday 03 September 2026 14:00:00 GMT