China’s top AI models now trail America’s best by just 3% on benchmark scores — a record-low gap — according to a new Bloomberg Intelligence report. The gap sat at roughly 9% in May and around 15% earlier in the year. The narrowing followed the September release of DeepSeek’s V4.1 Flash model.
Why it matters: a 3% benchmark gap is small enough to call the pricing of American frontier models into question. If a Chinese lab can match US capabilities at a fraction of the training and inference cost, the trillion-dollar bet on American compute dominance gets a lot harder to justify to shareholders.
What the report says
Bloomberg Intelligence senior analyst Robert Lea writes that Chinese labs, led by DeepSeek, have closed the performance gap faster than expected. DeepSeek’s V4.1 Flash ranked sixth globally on LiveBench in September with 81.1 points — close to the top score of 83.4, held by an Anthropic model — making it the highest-ranked Chinese model since DeepSeek’s R1 breakthrough last year. Only three of LiveBench’s top 15 models come from China, but Lea says the trajectory points to further market share gains for Chinese contenders, not fewer.
The driver, according to Lea: deepening AI expertise and researchers’ growing ability to optimize models for domestic hardware — including Huawei’s Ascend chips, which some analysts now view as roughly comparable to Nvidia’s H200 in China.
Why this rattles Washington
The gap narrowing “casts further doubt on the long-term sustainability of US technological supremacy in AI,” Lea writes. It also raises uncomfortable questions about the effectiveness of US chip export restrictions, which were designed to keep Chinese labs from catching up.
This is not the first scare. In January 2025, the release of DeepSeek’s R1 reasoning model triggered a record $589 billion one-day wipeout in Nvidia’s market value — the biggest single-day loss in US stock-market history — as investors questioned why US hyperscalers needed to spend hundreds of billions on chips. The stock recovered quickly, and the panic was widely treated as an overreaction. Now, the overreaction is starting to look like a pattern.
The caveats
Lea also notes that China’s AI industry still faces commercialization pressures: low-margin token supply, fierce price wars, and competition among more than 1,100 large language models could keep the sector unprofitable until 2030. ByteDance’s Doubao leads Chinese labs in AI application monetization, while DeepSeek’s and Tencent’s chatbots remain free to use.
Benchmarks are also an imperfect proxy for real-world capability. A 3% average gap across benchmark suites does not mean Chinese models match US models on every task — but as a trend line, it is the most important number in the AI race right now.
Frequently asked questions
What does the 3% figure mean?
Bloomberg Intelligence measures the average performance gap between the top Chinese and top US AI models on benchmark scores. That gap is now about 3%, down from roughly 9% in May and 15% earlier in the year.
Which Chinese model drove the narrowing?
DeepSeek’s V4.1 Flash, released in September. It ranked sixth globally on LiveBench with 81.1 points, near the top score of 83.4 from an Anthropic model.
Does this mean US export controls failed?
Bloomberg Intelligence says China’s progress raises questions about the effectiveness of US chip export restrictions. Chinese labs have partly adapted by optimizing for domestic hardware like Huawei’s Ascend chips. Whether that fully offsets the restrictions is still debated.
Are Chinese models cheaper to run?
Historically, yes — DeepSeek’s models have been dramatically cheaper to train and serve than US frontier models, though per-token pricing varies by provider and deployment. Bloomberg Intelligence notes the sector faces fierce price wars that may keep it unprofitable until 2030.
Sources: Bloomberg, Startup Fortune, InShorts