The Performance Gap That Wasn't: Chinese AI Reaches Competitive Parity

Frequently asked

What does it mean for AI chip export controls if the capability gap has already closed?

Export controls were designed to prevent China from reaching frontier capability by restricting compute access. With the performance gap now at 2.7 percentage points according to Stanford HAI, the gap the controls were built to preserve is effectively gone. The policy tool remains in place but its strategic justification — maintaining a meaningful capability lead — no longer applies. Policymakers defending the current regime are now defending access restrictions whose intended outcome has already been negated by the labs they were targeting.

Why did Chinese AI models close the gap so fast despite the US outspending China by 23 to 1?

The investment ratio assumed that capital translated directly into capability, and that restricting compute would compound a structural disadvantage. Instead, Chinese labs extracted more capability per unit of compute through architectural efficiency and focused engineering — releasing coding tools in a two-week window last April that matched the best Western models available. The gap closed because the ceiling on what restricted compute could produce was higher than US policy assumed, not because the restrictions failed as enforcement mechanisms.

What should a US AI developer or enterprise expect now that Chinese models dominate global API traffic?

Chinese models already account for 61% of token consumption among the top ten models on OpenRouter. Developers choosing model providers are already operating in a market where Chinese models are the plurality choice, not an emerging alternative. Enterprises that built procurement decisions around an assumed US performance premium now need to re-evaluate that premium against actual benchmark parity — the 2.7-point gap is too small to justify treating American models as categorically superior for most production use cases.

More on this wire