DeepSeek Narrows AI Gap With US to Just 3 Percent, Bloomberg Says

Three percent. That's the entire benchmark gap Bloomberg Intelligence now measures between China's top AI models and America's best. The firm says the shrinkage is fast enough to rattle Silicon Valley's trillion-dollar bet on staying ahead.

Top Chinese AI models now trail their US rivals by just 3% on benchmark scores, according to a Bloomberg Intelligence report published October 4 by senior analyst Robert Lea. That gap sat at roughly 9% in May and around 15% earlier in the year. The move came after DeepSeek released its V4.1 Flash model in September, and Bloomberg Intelligence says the trend points to further market share gains for Chinese AI labs, not fewer.

Three percent is not a rounding error. It is the kind of number that, if it holds, makes the argument for paying a premium for American frontier models a lot harder to sustain.

This is not DeepSeek's first scare for Washington. In January 2025, the release of DeepSeek's R1 model helped trigger a record Nvidia rout, as investors suddenly questioned why US hyperscalers needed to spend hundreds of billions of dollars on chips if a Chinese lab could build something comparable for a fraction of the cost. Nvidia erased $589 billion of market value on January 27, 2025, the biggest one-day wipeout in US stock-market history. Its shares fell nearly 17% that Monday after closing at $142.62 the prior Friday. The stock recovered quickly, and Bloomberg Opinion wrote in February 2025 that the DeepSeek panic looked more like an aberration in an otherwise unbroken AI boom than a genuine Sputnik moment.

Now, the aberration looks more like a pattern. Bloomberg Opinion, in a separate piece published the same day as Lea's report, argued that DeepSeek and Huawei are now taking direct aim at Nvidia's competitive moat in China, not just its stock price. That is a different kind of threat. A one-day selloff fades. A closing benchmark gap does not.

Washington's entire strategy for keeping distance between US and Chinese AI rested on cutting off the chips. That strategy has been loosening all year. The Bureau of Industry and Security published a rule in January 2026 that moved export reviews for Nvidia's H200 and AMD's MI325X chips from a presumption of denial to case-by-case approval, so long as applicants could show the shipments wouldn't starve US customers of supply and Chinese buyers met compliance and testing conditions. CNBC reported in August that the US was also racing to close a loophole letting Chinese firms rent Nvidia compute through data centers in Southeast Asia, since the export rules target who owns the physical chip, not who is running workloads on it remotely.

Meanwhile the big four hyperscalers, Microsoft, Amazon, Alphabet and Meta, are on pace to spend more than $600 billion on AI infrastructure in 2026 alone, according to figures reported by Yahoo Finance. That spending has helped keep Nvidia near record highs this year, but investors still ask whether AI revenue will actually scale to match the capital going into the ground. A narrowing capability gap is exactly the kind of data point that sharpens that question. If a Chinese lab working with less compute can get within 3% of the frontier, the case for spending at this scale gets harder to make with a straight face.

None of this means the US lead has vanished. Three percent is still a lead. But the direction of travel, 15% to 9% to 3% inside a single year, is the story, and Bloomberg Intelligence is explicit that it expects Chinese labs to keep gaining share, not plateau. Frankly, the people writing checks for the next generation of US data centers should be paying closer attention to that slope than to the headline number itself.

The Fortune piece on China's AI efficiency gains noted in September that Chinese labs have been closing the gap specifically by being more resourceful with less compute, not by matching US spending dollar for dollar. That is the uncomfortable part for anyone defending a multi-trillion-dollar infrastructure bet: the challenge isn't coming from someone trying to outspend Silicon Valley. It's coming from someone proving they don't have to.

Also read: Micron CEO Warns the Memory Chip Crunch Gets Worse Before It Gets Better • Reddit is cutting off free API and RSS access while keeping Google and OpenAI paying • A Home AI Cluster Build Hit Electrical Fuses Before It Hit GPU Limits

This article is posted in AI News, check it out for more related stories.