How China Is Closing the AI Gap With the United States

Stanford University's Institute for Human-Centered Artificial Intelligence (HAI) published its 2026 AI Index in April, reporting that the gap between the best American and Chinese AI models had shrunk to 2.7 percentage points. Anthropic's Claude Opus 4.6 topped the Chatbot Arena leaderboard with a score of 1,503 as of March 2026, while ByteDance's Dola-Seed-2.0-Preview followed at 1,464 — a 39-point difference, according to the report.

That margin is a sharp contraction from May 2023, when US models led Chinese rivals by 17.5 to 31.6 percentage points across major benchmarks, Stanford HAI found. The two countries' models have traded the top spot on various leaderboards multiple times since early 2025, including a moment in February 2025 when DeepSeek-R1 briefly matched the leading US system, the report said.

Spending far less, catching up anyway

The narrowing performance gap comes despite a wide funding gulf. US companies drew $285.9 billion in private AI investment in 2025, versus $12.4 billion in China — a ratio of roughly 23 to 1, Stanford HAI reported. The institute cautioned that the Chinese figure likely understates real spending, since Beijing channels resources through state-directed vehicles that private databases do not fully capture; Stanford estimated China has deployed roughly $912 billion through such funds across industries between 2000 and 2023.

Private deal sizes illustrate the contrast: Anthropic alone raised a $30 billion Series G round in February 2026 at a $380 billion valuation, according to reporting cited by implicator.ai, while Chinese AI labs collectively drew about $100 million from Gulf investors around the same period, compared with roughly $15 billion flowing into US labs.

The open-source bet

A central reason for China's rapid catch-up is its open-source distribution strategy. Chinese labs such as DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi, MiniMax and Zhipu have released many models as open-weight downloads that developers worldwide can run and modify for free, according to reporting by Rest of World. US leaders OpenAI and Anthropic have instead kept their most capable systems behind paid APIs and subscriptions, protecting margins but limiting global diffusion, per the same reporting.

The strategy paid off in adoption terms: Chinese open-source models — led by Qwen, MiniMax and DeepSeek — reached roughly one-third of global large-language-model usage, up from almost nothing in late 2024, according to a Fortune analysis. DeepSeek's cost-efficiency push was part of this shift; the company's parent said it could generate one million tokens of output for about 3 yuan (roughly 50 cents), about a twentieth of ChatGPT's contemporaneous cost, Fortune reported.

Stanford HAI's 2026 Index noted a partial reversal of this trend: some Chinese firms are moving toward closed, paid models. Alibaba launched Qwen3.6-Plus and Qwen3.5-Omni as paid cloud services, and Zhipu released its GLM-5-Turbo as closed-source, pricing it at $1.40 per million input tokens and $4.40 per million output tokens — below Anthropic's $5/$25 rates for Claude Opus 4.6, according to implicator.ai's summary of the report.

Chips and workarounds

Nvidia's practical access to China's data-center market has shrunk amid Washington's restrictions on advanced-chip exports. By the end of its fiscal year 2026, the company recorded no revenue from H20 chips — a China-tailored model — in its most recent quarter, and Beijing has separately discouraged domestic firms from using Nvidia products over security concerns, according to Gulf News reporting on the company's chip production plans.

Huawei has used the opening to expand its Ascend chip line. The company plans to produce roughly 600,000 units of its Ascend 910C in 2026 — about double 2025 output — and raise total Ascend die production to as many as 1.6 million units, Gulf News reported, citing production targets. The chips are still made on 7-nanometer process technology through domestic partner SMIC, compared with 4-nanometer nodes used in Nvidia's newest GPUs, and Huawei still relies on foreign high-bandwidth memory components, the same report said.

Where each side actually leads

Stanford HAI's assessment is not that China has overtaken the US outright, but that the two countries lead on different metrics. The US still leads in total investment, production of top-tier frontier models, and computing infrastructure. China leads in AI publication volume, citations, patent filings and industrial robot installations, according to the Index's summary of 2026 findings.

On patents, Chinese entities filed 69.7% of AI patents worldwide, and Chinese researchers produced 23.2% of global AI publications and 20.6% of citations, compared with 12.6% for the US, according to Stanford HAI's 2026 Index as reported by The Next Web. On robotics, China installed about 295,000 industrial robots in the period measured — roughly nine times the United States' 34,200, the same report found.

The Index also flagged a US vulnerability unrelated to chips or funding: immigration of AI scholars into the United States has fallen 89% since 2017, with an 80% drop in the most recent year alone, even though the US still hosts more AI researchers than any other country, according to Stanford HAI's own summary of the report.

Industrial policy versus private-sector bets

The two countries' underlying strategies differ structurally. China's approach embeds AI development inside a state-coordinated framework of industrial policy and government guidance funds, as reflected in the investment figures described above, while the US model relies more on private capital mobilization and corporate competition among labs such as OpenAI, Anthropic, Google and xAI.

Commercial tensions inside the US industry have also become public. OpenAI chief executive Sam Altman and Tesla and xAI chief Elon Musk have traded public jabs — Musk has called Altman "Scam Altman," while Altman has criticized Musk's satellite data-center plans as unworkable, according to Business Insider Japan. Despite that rivalry, major US AI firms — including OpenAI, Anthropic, xAI, Google and Nvidia — agreed to a White House-led framework of voluntary safety standards, the report noted.