U.s.: China AI Watch: The Gap Just Collapsed

“The race is moving from who creates the frontier to who industrializes it first.”

Bloomberg Intelligence now estimates that the best Chinese models trail leading U.S. models by only 3% on benchmark performance, down from roughly 9% in May and 15% earlier in 2026. The frontier has not changed hands—but the rate of convergence has accelerated sharply.

DeepSeek's September V4.1 Flash scored 81.1 on LiveBench versus 83.4 for Anthropic's best model, making it the strongest Chinese showing since DeepSeek R1. Only three of the top 15 models are Chinese, so the United States still leads the frontier. What is changing is the economics around it.

1. 01 The second-order effect outweighs the benchmark

   If a 15% performance gap can compress to 3% within months while Chinese models remain materially cheaper, the premium attached to frontier intelligence becomes harder to defend outside the most demanding workloads.

   That shifts value away from model scarcity and toward distribution, inference cost, proprietary data, accelerated-computing efficiency and industrial deployment. signal: Value moves from model scarcity to the production and diffusion of intelligence.

2. 02 Abundance, not dominance, reshapes the economics

   China does not need to dominate the absolute frontier to reshape AI economics. It only needs to make near-frontier intelligence abundant enough to accelerate adoption across manufacturing, logistics, energy, robotics and mobility. signal: Near-frontier intelligence made abundant is the strategic lever.

3. 03 The U.S. counter-signal

   Nvidia-backed Reflection is preparing a powerful U.S. open-weight model explicitly aimed at matching leading Chinese open models and enabling enterprises to build localized, lower-cost "AI factories" on Nvidia infrastructure. Axios notes that the most powerful open-weight models are currently overwhelmingly Chinese.

   The U.S. response appears to be moving beyond closed frontier leadership toward defending the open, lower-cost inference layer where China has been gaining strategic ground. signal: America is moving to defend the open, lower-cost inference layer.

Key takeaways

the strategic inflection

• The frontier has not changed hands; the economics around it are changing faster than expected.
• The U.S. still builds the strongest intelligence—superior models, capital, chips and hyperscale infrastructure.
• China is getting increasingly close and is trying to make that intelligence cheaper, more open and more pervasive.
• The AI Supercycle stays bullish for infrastructure but more discriminating for standalone model margins.