Stanford's big AI report: The most important takeaways

Every year, Stanford University releases what has become the closest thing the AI industry has to an official scorecard. Now in its ninth edition and running 423 pages, the AI Index tracks nearly everything: how many models were released and by whom, how much money flowed into the industry, how AI is reshaping labor markets, what it's doing to the power grid, and how the public feels about all of it. The report is widely cited by policymakers, journalists, and executives — and supported by partners including Google and OpenAI while being partly written by people who work at these and other AI companies.

With that in mind, here are a few findings worth pulling out.

China is catching up fast

The U.S.-China AI model performance gap has effectively closed. As of March 2026, Anthropic's top model leads the best Chinese competitor by just 2.7 percentage points, a margin that has flip-flopped repeatedly since DeepSeek's R1 briefly matched American models in February 2025.

The U.S. still produces more top-tier models — 50 notable releases in 2025 compared to China's 30 — and commands a massive private investment lead, $285.9 billion to China's $12.4 billion. But the report notes that figure significantly understates China's total spending, since government guidance funds have channeled an estimated $184 billion into Chinese AI firms since 2000. China also now leads the world in AI publications, citation share, patent grants, and industrial robot installations.

Some American AI companies have their own theory for why the gap is closing: they say Chinese labs have been stealing it. OpenAI, Anthropic, and Google have begun sharing intelligence on what they call adversarial distillation — training models on a competitor's outputs to replicate their capabilities at a fraction of the cost. They claim DeepSeek and others have done this without authorization, though they have yet to release evidence showing how much of China's recent progress is actually attributable to distillation rather than independent development.

One area where the U.S. lead is unambiguous is data centers

The country hosts 5,427 of them, compared to 449 in China and around 525 each in Germany and the United Kingdom. Total AI data center power capacity reached 29.6 gigawatts by the end of 2025, roughly equivalent to New York state at peak demand.

That scale comes with a cost. Training a single model, Grok 4, produced an estimated 72,816 tons of CO2 equivalent, more carbon than roughly 1,000 average cars emit over their entire lifetimes. Running models creates its own footprint. Annual water use for GPT-4o inference alone could exceed the drinking water needs of 12 million people, according to the report's estimates.