BI Report: DeepSeek's Rise Narrows US-China AI Gap to Just 3%
nashnova research
Bloomberg Intelligence reports that after DeepSeek released its V4.1 Flash model, the benchmark gap between top US and Chinese AI models shrank from 15% at the start of the year to roughly 3% — yet China's AI industry may not turn a profit until 2030, exposing a core tension between technical catch-up and commercial viability.
What does a 3% gap actually mean?
BI senior analyst Robert Lea tracks three inflection points: the US-China benchmark gap stood at roughly 15% in January, narrowed to about 9% by May, and fell to approximately 3% after DeepSeek's V4.1 Flash launched in September.
V4.1 Flash ranked sixth on LiveBench, scoring 81.1 versus Anthropic's top mark of 83.4. Lea considers this "comparable performance."
This means → China's strongest model has moved from "clearly behind" into a "comparable" band — the US lead is now thin enough that a single model iteration could erase it.
How did China close the gap so fast?
Lea attributes the progress to two drivers: deepening AI research expertise among Chinese teams and improving ability to optimise models for domestic hardware.
In plain terms = Chinese labs sharpened their algorithms while learning to squeeze more performance out of restricted chips — two tracks running in parallel.
This reflects growing doubt about the effectiveness of US chip-export controls — restrictions designed to slow China's AI development may instead have forced stronger hardware adaptation.
Only 3 of the top 15 — what does that reveal?
Despite DeepSeek's top-six finish, Chinese models hold just 3 of 15 spots on the LiveBench leaderboard.
This means → China's breakthroughs are concentrated in a handful of elite teams; the overall bench depth lags far behind the US — closing in at one point is not the same as catching up across the board.
Lea writes that China's progress "further calls into question the sustainability of US technological dominance in AI."
The technology is catching up — so why is nobody making money?
More than 1,100 large language models have flooded the Chinese market, and intense price wars make it nearly impossible for any single player to build a moat.
On monetisation, ByteDance's Doubao leads in AI-app revenue, while DeepSeek and Tencent chatbots remain free.
In plain terms = everyone is burning cash to grab users, and no one dares raise prices first — the first to charge is the first to lose traffic, so the entire field stays unprofitable.
Profitability by 2030 — is that timeline realistic?
Lea estimates China's AI industry may not break even until 2030, contingent on "competitive pressures easing, industry consolidation, and more rational pricing."
Chinese AI models also face tightening US regulatory scrutiny and even potential bans, partly over allegations of model distillation — a technique that compresses a large model's capabilities into a smaller one.
This means → rapid technical convergence paired with a long road to profitability forms the central contradiction of China's AI sector — the technology can close the gap, but whether the industry can sustain itself financially may take another five years to answer.
市场有风险，内容仅供研究参考，不构成投资建议。