Chinese AI Models Trail U.S. Rivals by 3% on Benchmarks, but U.S. Startups Have Raised Roughly 10 Times More Venture Money

The performance gap between American and Chinese AI models has shrunk to a record low. The funding gap has not.

Top Chinese models lag their U.S. rivals by just 3% on benchmark scores after DeepSeek released V4.1 Flash in September, Bloomberg Intelligence senior analyst Robert Lea wrote in an Oct. 5 report. That is down from about 9% in May and 15% earlier in the year.

The benchmark picture

DeepSeek's V4.1 Flash ranked sixth globally on LiveBench in September. It is the highest-ranked Chinese model since DeepSeek's R1 reasoning model in 2025.

Its latest LiveBench score was 81.1, against 83.4 for Anthropic's best. Lea called that "comparable performance" to leading systems from Anthropic and OpenAI.

The lead is not gone. Only three of the top 15 models on LiveBench are Chinese. Lea also cautioned that rankings shift and that monetizing a model is hard regardless of where it sits on a leaderboard.

Fortune's analysis puts the lag differently, in time rather than score: analysts estimate the best Chinese models are now about four months behind the newest releases from OpenAI and Anthropic, versus seven months at the start of the year.

Moonshot AI's Kimi K3 is the headline open-weight example. It uses a mixture-of-experts design with 2.8 trillion parameters and a one-million-token context window. Its performance stayed below the strongest proprietary systems while exceeding other open and proprietary models in testing.

Usage is shifting too

Chinese models went from 1.2% of token traffic in 2024 to more than half the total by the summer of 2026, per the figures Fortune cites. Token traffic is one common way to measure AI usage. It measures volume, not revenue.

Lea said the improvement spells further market share gains for Chinese contenders. He attributed the progress to deepening expertise and researchers' ability to optimize models for domestic hardware.

What it means for export controls

Lea said China's progress raises questions about the usefulness of U.S. export restrictions on chips such as Nvidia's, which were meant to slow Chinese AI and keep Huawei from building alternatives.

That is an analyst's reading, not a finding by any government. The data shows Chinese labs closing the gap under those restrictions. It does not show how much further ahead they would be without them.

The money gap

The numbers on capital are lopsided. Between 2023 and 2026, venture funding into U.S. AI companies topped $380 billion, while Chinese startups received barely a tenth of that, according to Boston Consulting Group. One tally puts the Chinese figure at about $39 billion.

Venture investment in China totaled $20 billion in the first quarter of 2026, against $267 billion in the U.S. Newly registered Chinese VC funds did hit 154 billion yuan ($22.8 billion) in assets in the first five months of the year, already more than all of last year, but that is far below what American venture capital deploys.

Infrastructure tells the same story. U.S. technology companies spent more than $400 billion on capital expenditures in 2025, versus $63 billion in China, and U.S. spending is projected to top $800 billion in 2026. American AI computing capacity is estimated at eight times China's, measured in H100-equivalent GPUs.

OpenAI announced $110 billion in new investment on Feb. 27 at a $730 billion pre-money valuation, with $30 billion each from SoftBank and Nvidia and $50 billion from Amazon.

Chinese startups look to Hong Kong

With state guidance funds favoring later-stage companies and early-stage venture capital only now recovering from a three-year fundraising drought, Hong Kong's stock market has become a financing route. Zhipu AI raised $558 million in a Hong Kong listing on Jan. 8. MiniMax Group's Hong Kong IPO in January raised $620 million according to Fortune, while another outlet reported $711 million on Jan. 9; the sources disagree on the figure.

AI companies listed in Hong Kong raised $4.9 billion in December 2025 and January 2026, and about 20 more were in the pipeline as of Feb. 3. "New AI companies are going public in Hong Kong, [and] investors can now, for the first time, look beyond AI proxy stocks to direct investment opportunities," said Johnson Chui, HKEX's head of global issuer services.

Those proceeds are small next to the largest U.S. private rounds. Public listings also expose young companies to market pressure while their spending stays high.

Rising costs inside China

Fortune points to inflation within China's AI economy. Memory maker CXMT has raised prices for months and held firm even when Huawei, one of its largest customers, demanded relief. Postings for AI-related jobs rose roughly twelvefold year-on-year in early 2026, and engineers specializing in large language models command some of the highest pay in the country.

State banks are directed to prioritize tech lending, but they are also absorbing rising non-performing loans elsewhere, which could tighten overall credit. Chinese enterprise software firms mostly sell domestically, which caps revenue.

Lea said price wars and fierce competition are squeezing profits, with sustainable profitability expected only after consolidation by 2030.

There is a counterargument inside the same reporting. Fortune notes the next generation of AI ventures may not need vast capital if they build applications on top of existing models.

America's open-weight response

On Monday, Oct. 5, Nvidia-backed Reflection AI announced Beam, its first open-weight model. It has 501 billion parameters but activates only 23 billion per query. The company says it beats U.S. open models including Thinking Machines' Inkling and Nvidia's Nemotron 3 Ultra on agentic coding and reasoning benchmarks.

Reflection concedes Kimi K3 remains "ahead on raw capability." It claims an edge on efficiency, saying Beam matches GLM-5.2 on advanced reasoning with three to four times less compute. Those are the company's own claims. Beam is in final evaluations with early access limited to select parties.

Beam's weights are due later this month. Independent benchmarks of that release will show whether America's open-weight answer holds up against Kimi K3 and the rest of the Chinese field.