Estimates how much frontier compute China holds despite export controls, which affects assumptions about open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → model capability and where future models can come from.
Key takeaways · AI-distilled
Documented cases from indictments and investigative reporting point to nearly 300,000 H100-equivalents diverted by end-2025, about a quarter of the compute China got legally or made domestically. Epoch expects the true total is higher since much smuggling goes undetected.
Epoch's 660,000 median is roughly 3% of the global compute stockpile, comparable to xAI's holdings at the time. At the 1.6 million upper bound, most of China's AI compute at end-2025 would have been smuggled.
Each tightening of US controls raised the incentive to smuggle: the October 2023 expansion, then an April 2025 license requirement on Nvidia H20 chips built to meet earlier rules, after which the Financial Times reported over $1 billion in chips smuggled from April to July.
Server and cloud resellers are a major channel: the DOJ accused Supermicro employees of diverting $2.5 billion in Nvidia chips to a pass-through entity that redirected servers to China.
Epoch names its largest uncertainties as the scale of undetected smuggling and how many allegedly diverted or missing chips actually reached China.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
