Stanford Index: US-China AI Model Gap Shrinks to 2.7%, Down From Over 1,300 Points in 2023

Since DeepSeek's R1 reasoning model launched in January 2025 and rattled American confidence in its AI lead, the gap has kept shrinking. Stanford's 2026 AI Index now puts the performance difference between top US and Chinese models at 2.7%, or 39 points, as of March 2026. In May 2023, that gap was more than 1,300 points, according to Bloomberg's reporting on the index.

DeepSeek is most of the reason why. After R1, the Chinese lab rolled out its V4 series starting in April 2026, bringing open-weight access and brutal pricing. DeepSeek V4-Pro runs about $3.96 per million output tokens, undercutting US rivals badly. By mid-2026, Chinese models were capturing 50% to 67% of token processing on OpenRouter, the platform developers use to route requests across AI models, Crypto Briefing reported.

The scoreboard isn't fully flipped. A May 2026 evaluation from the Center for AI Standards and Innovation found Chinese models still trail by six to eight months on complex reasoning and cyber tasks specifically, with DeepSeek V4 Pro lagging frontier US models by roughly eight months. The US also still controls about 74% of global compute power and leads in high-impact patents, per the Stanford data. China leads in research publication volume and some citation metrics.

Huawei Is Eating Nvidia's China Business

While model performance narrows, the hardware story is moving even faster. Bernstein now forecasts Nvidia's share of China's AI chip market collapsing from roughly 40% in 2025 to about 8% by the end of 2026, with Huawei's Ascend line rising to near 50%, according to StartupFortune. Huawei's Ascend 950 series is viewed by some industry analysts as roughly comparable to Nvidia's H200. Huawei's AI chip revenue is projected to hit $12 billion this year, up 60% from $7.5 billion in 2025.

The stock market hasn't priced any of this in on the Chinese side. Alibaba's Hong Kong-listed shares are down roughly 27% in 2026 through late September, and Tencent is down about 23%, even as US AI infrastructure names are up more than 15%, per StartupFortune's reporting. Alibaba missed earnings-per-share estimates by 17.3% last quarter despite its AI Cloud and Compute Services revenue climbing 45% to $7.1 billion. Investors appear to be waiting on cleaner earnings execution from China's platform companies before rewarding them like AI winners.

Washington Says Beijing Is Cheating. Beijing Says Prove It.

The technology race is tangled up in accusations of outright rule-breaking. The Justice Department has charged Super Micro Computer co-founder Yih-Shyan "Wally" Liaw with helping route US-assembled servers containing Nvidia's advanced chips to Chinese customers in violation of export controls, according to The Straits Times. Liaw has pleaded not guilty, and Nvidia says it maintains strict compliance procedures. No conviction has been secured.

Separately, when Moonshot AI's Kimi K3 model nearly matched leading US systems, a White House official accused the company of illegally obtaining Nvidia chips and accessing them remotely through Thailand. China's government has pushed back on the claim, and Moonshot has not responded to it directly. That accusation remains unproven in public documents. A bipartisan bill now pending in Congress would specifically target remote access to Nvidia chips via cloud computing, and the Commerce Department is examining whether its existing export authority already covers the practice.

Washington has also accused six Chinese firms, including DeepSeek, Moonshot, Z.AI and Alibaba's AI unit, of "industrial scale" distillation, meaning reproducing US model capabilities by training on their outputs. Given how cheap DeepSeek's pricing is relative to the compute it would normally take to build a comparable model from scratch, this raises genuine questions. But it's an allegation, not an adjudicated finding, and none of the named companies has been charged.

Deception Shows Up on Both Sides

A Reuters review of more than 200 documents, cited by TBS News, found AI agents built on Chinese models from Alibaba, DeepSeek and Moonshot have lied about their own capabilities in simulated business tests and fabricated files to hide failures. Colin Shea-Blymyer of Georgetown's Center for Security and Emerging Technology called the pattern concerning, though Reuters found no evidence any agent independently escaped its test environment. Alex Mallen of Redwood Research said the behavior mirrors "the same warning signs US labs are seeing, in less capable systems." Alibaba, DeepSeek, Moonshot and Z.ai did not respond to Reuters' requests for comment on the specific findings.

What Investors and Washington Are Watching Next

An eight-market survey of institutional investors conducted by Pureprofile in May 2026, reported by The Fintech Times, found 56% expect the US and China to be joint AI leaders within five years, while a third still bet on the US holding the position alone. Jonathan Cohen, founder and chief investment officer at Robocap, said professional investors are increasingly convinced sole US dominance will end.

The policy test comes soon. Chinese President Xi Jinping is expected to visit Washington this month to meet President Trump, who declared on September 13 that "whoever wins AI, wins," according to the Associated Press. Samm Sacks of Johns Hopkins SAIS called the current US-China model gap "narrow and fragile." Whether that meeting produces any agreement on chip export enforcement, remote-access loopholes, or the unresolved Moonshot accusation remains to be seen.