Stanford HAI’s 2026 AI Index: Capabilities sprint, US-China race tightens, hardware chokepoints and rising risks

Key findings: faster, stronger, stranger

Stanford HAI’s 2026 AI Index finds that AI capability is accelerating, not plateauing. Industry produced over 90% of notable frontier models in 2025, several models now meet or exceed human baselines on PhD‑level science questions and multimodal reasoning, and a key coding benchmark (SWE‑bench Verified) jumped from about 60% to nearly 100% in a single year. Adoption is staggering: organizational adoption hit 88% and four in five university students now use generative AI. But progress isn’t tidy—Gemini Deep Think earned an IMO gold medal while the top model still reads analog clocks correctly only about 50% of the time. Curious? You should be.

Geopolitics, chips and capital

It has been reported that U.S. and Chinese models exchanged leads multiple times since early 2025, and as of March 2026 an Anthropic model led by roughly 2.7%. The U.S. still claims more top-tier models, higher‑impact patents, vastly more data centers (5,427, more than ten times any other country) and far larger private investment—$285.9 billion in 2025 versus $12.4 billion in China. But China leads in publication volume, citations, patent output and industrial robot installs; South Korea tops patents per capita. Meanwhile, almost every leading AI chip is fabricated by a single company, TSMC, creating a clear supply-chain concentration even as a TSMC U.S. expansion began operations in 2025. That’s a big single point of failure in a globally vital industry.

Adoption, impact and uneven benefits

Generative AI reached roughly 53% population adoption within three years—faster than the PC or the internet—though adoption tracks GDP per capita (Singapore 61%, UAE 54%; the U.S. sits at 28.3%). Stanford estimates generative tools deliver about $172 billion annually in value to U.S. consumers, and median per‑user value tripled between 2025 and 2026. Yet the talent flows that fueled past U.S. leadership are cooling: new AI researchers and developers moving to the U.S. declined 89% since 2017, with an 80% drop in the last year alone. Is the boom turning into a bottleneck? It sure smells like one.

Risks, reporting gaps and the road ahead

The report flags rising harms and gaps in governance: documented AI incidents climbed to 362 from 233 in 2024, reporting on responsibility benchmarks is spotty, and research shows that improving one safety dimension can degrade another—tradeoffs are real. AI agents improved on OSWorld from 12% to about 66% task success, but still fail roughly one in three structured benchmarks. Policymakers and institutions face an urgent balancing act: accelerate benefits while containing harms, build hardware resilience, and make responsible‑AI reporting standard rather than optional. The technology is moving fast. Are our guardrails keeping up? Not yet.

Sources: hai.stanford.edu