Bloomberg Intelligence reports US lead in AI benchmarks narrows to 3% after DeepSeek release

Bloomberg Intelligence says China’s top AI models have cut the United States’ lead in benchmark scores to just 3%, the narrowest gap on record, after DeepSeek released its V4.1 Flash model in September. The shrinking margin has already become political ammunition, with President Donald Trump citing the competition with China to justify faster AI development and data center buildouts despite rising local opposition.

• DeepSeek’s V4.1 Flash ranked sixth globally on LiveBench with a score of 81.1, against 83.4 for Anthropic’s top model.
• Bloomberg Intelligence measured the US-China benchmark gap at 15% earlier this year and about 9% in May before the latest drop.
• An NBC News poll found 69% of adults oppose AI data centers being built near their communities.

• 3% US lead over China in AI benchmark scores, lowest on record
• 81.1 DeepSeek V4.1 Flash’s LiveBench score versus Anthropic’s 83.4
• 69% US adults opposing nearby AI data centers, per NBC News poll
• 120 days deadline for Jay Clayton’s White House AI task force report

The United States’ lead over China in artificial intelligence benchmark scores has narrowed to just 3%, the slimmest gap on record, according to BeInCrypto. The finding comes from Bloomberg Intelligence, which tracks model performance across independent benchmarks.

DeepSeek released its V4.1 Flash model in September, pushing the gap to its lowest level yet. Trump invoked that same narrowing margin in September 2024 when he rejected calls to slow AI development, arguing China was the only party pleased by the pushback.

DeepSeek’s V4.1 Flash closes to 2.3 points of Anthropic on LiveBench

Bloomberg Intelligence had put the US lead at 15% earlier this year and near 9% in May, before DeepSeek’s release cut it further. V4.1 Flash ranked sixth globally on LiveBench, an independent leaderboard that scores models on logic puzzles, coding tasks and reasoning questions. It scored 81.1, within 2.3 points of the 83.4 posted by Anthropic’s best-performing model.

Even so, Chinese models remain a minority at the top of the leaderboard. Only three of LiveBench’s top 15 models are Chinese-built, with the rest coming from US labs.

Bloomberg Intelligence credited the gains to deeper technical expertise at Chinese labs and to models engineered specifically for domestic hardware, a response to years of Nvidia export restrictions. The narrowing gap raises questions about whether those controls are working as intended. Anthropic argued earlier this year that tighter restrictions could preserve a 12-to-24 month US lead through 2028, but BI senior analyst Robert Lea cautioned that benchmark rankings shift quickly and that Chinese labs may struggle to convert high scores into commercial revenue.

Trump says China is ‘the only one’ happy about the AI pushback

The shrinking gap lands amid a domestic debate over how aggressively US labs should race ahead. Anthropic CEO Dario Amodei has urged pacing the frontier rather than racing unchecked, a position echoed by OpenAI CEO Sam Altman and Tesla’s Elon Musk.

Opposition to the industry’s physical footprint is building too. A poll by NBC News found 69% of adults, more than two in three surveyed, oppose AI data centers being built in their communities.

Trump has answered that backlash by framing AI as a zero-sum contest with China. In a post on Truth Social, he wrote:

There is a SICK conspiracy going on against AI and Data Centers, and the only one that is happy about it is China. WHOEVER WINS AI, WINS!

Donald Trump, President of the United States

He has also warned that communities blocking data center construction will end up backwards and poor.

Jay Clayton’s AI task force owes a report within 120 days

Trump has converted that rhetoric into policy. On September 19, he announced the creation of an AI Force and said he would not slow the industry’s pace of development.

A White House task force led by Jay Clayton, the administration’s AI czar, is due to report back within 120 days on AI’s risks and opportunities, along with the federal government’s responsibilities toward the technology. Trump separately signed a voluntary accord with six tech leaders that leaves safety practices to the companies themselves, and an executive order renaming artificial intelligence as “superintelligence.”

The BlockWest read. We read this as a signal for chipmakers and data center financiers more than for the AI labs themselves. If export curbs are not preserving the intended US advantage, the competitiveness case for restricting Nvidia sales abroad gets harder to defend, even as it stays politically useful for justifying faster permitting and lighter oversight at home.

Jay Clayton’s task force is due to deliver its findings within 120 days of the September 19 announcement, putting its report around mid-January 2027. Whether that review addresses the effectiveness of Nvidia’s export controls, or simply endorses the administration’s faster-is-

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.