On a Monday morning in April 2026, Stanford published a 423-page PDF with the number, which is the type of number that alters the tone of a congressional hearing. Claude Opus 4.6 from Anthropic has 1,503 points on the Arena Leaderboard. The Dola-Seed-2.0 Preview from ByteDance is currently at 1,464. The top AI models in China and the United States differ by 39 Elo points. 2.7 percent is the percentage. To retain a 2.7 percent performance edge, the US invested around $285.9 billion in private AI in 2025, which is more than 23 times China’s $12.4 billion. According to Stanford’s 2026 AI Index, the gap is “effectively closed.”
In January 2025, a Hangzhou-based business named DeepSeek published a model known as R1, fourteen months before to the publication of that PDF. For many years, it was believed in Silicon Valley and Washington policy circles that China was at least two to three years behind the United States in AI capabilities, and that U.S. export restrictions limiting access to cutting-edge Nvidia chips would perpetually widen that gap. In a single weekend, DeepSeek’s publication disproved both presumptions.
R1 used older, constrained technology at a stated training cost that was significantly less than what comparable U.S. laboratories were spending, and it momentarily outperformed the top U.S. model on MMLU, one of the industry’s usual benchmarks for comparing capabilities across models. The market’s reaction in Washington was evident almost immediately; in the days that followed the announcement, Nvidia’s stock fell precipitously as analysts revised their estimates of the amount of processing power truly needed to develop frontier AI.
Important Information
| Field | Details |
|---|---|
| Stanford 2026 AI Index | Released April 2026 — finds US-China AI performance gap has “effectively closed”; US lead over China on the Arena Leaderboard as of March 2026: 2.7% (39 Elo points) |
| Current Top Models | U.S.: Anthropic Claude Opus 4.6 — 1,503 Arena points; China: ByteDance Dola-Seed-2.0 Preview — 1,464 Arena points |
| DeepSeek-R1 Moment | January 2025 — DeepSeek-R1 from a Hangzhou lab briefly matched the top U.S. model outright on MMLU; released open-source at a fraction of OpenAI’s training cost |
| DeepSeek V3.2 (Dec 2025) | Achieved gold-medal level at International Mathematical Olympiad (35/42 points); 10th place at International Olympiad in Informatics; 2nd place at ICPC World Finals; released free under Apache 2.0 |
| MMLU Benchmark Gap | Collapsed from 17.5% (U.S. lead) in 2023 to 0.3% in 2024; US and Chinese models have traded the lead multiple times since early 2025 |
| US vs China Investment (2025) | U.S. private AI investment: $285.9 billion; China: $12.4 billion private — the US outspent China by roughly 23 to 1 to achieve a 2.7% performance lead |
| China’s Open-Source Strategy | Qwen, DeepSeek, GLM series distributed as open-weight models; by August 2025, Chinese open-source models held nearly 30% of U.S. enterprise market share (up from 1.2% in late 2024) |
| China’s Robot Lead | Stanford: China leads U.S. in industrial robot installations — the “bodies” side of the AI race; Chinese AI robotics deployment accelerating through 2025–2026 |
| The Efficiency Insight | U.S. chip export controls forced Chinese engineers to optimize for efficiency rather than raw compute — DeepSeek’s training cost was a fraction of comparable U.S. models; constraint bred a generation of efficiency-first AI engineers |
As recently as 2023, the MMLU disparity was 17.5 percentage points in favor of the United States; by 2024, it had shrunk to 0.3 percentage points. By December 2025, DeepSeek had released V3.2, which ranked second at the ICPC World Finals, won gold at the International Mathematical Olympiad, and had a 96 percent pass rate on AIME, well ahead of the similar U.S. models of the time.
All of it was made available to anyone who wished to download and run it as open-source software under Apache 2.0. The coordinated releases from DeepSeek and three other Chinese labs in January 2025 were described by the Center for Strategic and International Studies as a “potential shift in the global AI landscape” and cautiously noted that the simultaneous timing “likely” involved some degree of coordination by the Chinese government.
For those studying AI geopolitics, the mechanism was what made the efficiency tale truly compelling. It seems that U.S. chip export regulations, which were intended to hinder China’s capacity to train sophisticated AI models by limiting access to Nvidia’s H100 and A100 processors, had an unexpected effect. Chinese engineers improved their algorithmic optimization and hardware efficiency because they couldn’t use the same amount of computing power as American labs.
Because its developers had spent years studying how to extract performance that other laboratories just paid for, DeepSeek’s model was able to reach performance comparable to GPT-4 utilizing older, lower-tier hardware. The limitation produced a generation of researchers who prioritize efficiency. As compute costs increase globally and power consumption becomes a real barrier to AI development worldwide, this mentality may prove to be more significant than the raw computing advantage currently enjoyed by U.S. labs.
The simple “U.S. wins, China loses” story is complicated by the Stanford Index in a few additional noteworthy ways. With 40 noteworthy models in 2024 compared to China’s 15, the United States continues to create more superior AI models and higher-impact patents. The paper notes that Chinese models are occasionally tailored for the benchmarks that are used to evaluate them, which is an important disclaimer. Before government guiding funds started making up for it, China’s private AI investment fell from $16 billion in 2018 to $5 billion in recent years.
The open-source narrative also has an extremely peculiar conclusion: in April 2026, reports surfaced that Zhipu and Alibaba’s Qwen team had both switched to closed-source commercial models, coinciding with the Stanford Index’s confirmation of how well the open-weight strategy had closed the performance gap. At this point, the labs that employed open-source distribution to catch up to the frontier seem to have concluded that the sharing period was finished.
It’s difficult to ignore the fact that the topic of discussion in Washington has changed from “whether China is catching up” to something more difficult to articulate: a competition where the most competitive Chinese models are operating on servers in US data centers, where efficiency may be just as important as scale, and where spending more does not necessarily translate into greater advantage.
