What the 2026 Stanford AI Index Reveals About Global AI Power Shifts

AI Capability Acceleration

In 2025 more than 90% of the world’s leading large‑language models were released, and their performance on doctoral‑level scientific tasks, multimodal reasoning, and competition‑grade mathematics matched or exceeded human baselines. On the SWE‑bench Verified coding benchmark, the average pass rate rose from 60% to nearly 100% within a single year. Enterprise adoption is high: 88% of organizations have integrated generative AI into routine workflows, and four out of five university students regularly use these tools for coursework.

US‑China Model Gap Closing

Since early 2025 the leadership on top‑tier models has alternated between the United States and China. In February 2025 DeepSeek‑R1 briefly matched the performance of the US‑ranked model, and by March 2026 Anthropic’s best model led by only 2.7%. The US still dominates in the sheer number of models produced and in high‑impact AI patents, while China leads in academic paper volume, citation counts, total patents, and industrial robot installations. South Korea has the highest per‑capita AI patent count.

Compute Concentration

Global AI compute infrastructure is highly uneven. The United States operates 5,427 data centers—over ten times the count of the second‑largest country—making it the world’s largest consumer of AI‑related electricity. The hardware supply chain is fragile: virtually every leading AI accelerator is fabricated by TSMC. To mitigate this concentration risk, TSMC’s US production line began operating in 2025.

Jagged Frontier of Model Ability

Researchers describe a “jagged frontier”: models excel like a brilliant scholar on some tasks but behave like a naïve child on others. For example, Gemini Deep Think won a gold medal at the International Mathematical Olympiad, yet its accuracy reading analog watches is only 50.1%. In the OSWorld real‑OS benchmark, task‑success rates climbed from 12% to roughly 66%, but structured test suites still see failures in about one‑third of attempts.

Responsible AI Progress Lags

Higher capability does not guarantee safety. Companies publish impressive benchmark scores, but responsible‑AI metrics remain fragmented and inconsistent. Documented safety incidents rose from 233 in 2024 to 362 in 2025. Improving one safety dimension (e.g., tightening security constraints) often degrades another (e.g., output accuracy), creating a technical trade‑off.

US Talent Drain

Private AI investment in the US reached $285.9 billion in 2025—23 × the $12.4 billion reported for China—but the flow of overseas researchers to the US fell 89% since 2017, with an 80% drop in the most recent year. Despite 1,953 newly funded AI startups in 2025, attracting top talent is becoming increasingly difficult.

Generative AI Adoption Records

Generative AI tools achieved 53% global population penetration within three years, outpacing the diffusion rates of personal computers and the internet. Adoption correlates strongly with GDP: Singapore (61%) and the UAE (54%) lead, while the US ranks 24th at 28.3%. By early 2026, these tools are estimated to generate $1.72 trillion annually for US consumers, with median per‑user value tripling from 2025 to 2026.

Education System Lagging

Over 80% of US high‑school and college students already use AI assistants for coursework, yet only 50% of K‑12 institutions have formal AI‑use policies, and merely 6% of teachers consider those policies clear.

Policy Direction & Open‑Source Momentum

Governments now treat AI as a sovereign asset, accelerating national strategies and super‑computing investments. While large‑model development remains US‑China‑centric, open‑source contributions from the rest of the world are surging—GitHub activity from non‑US/China regions has surpassed Europe and is rapidly approaching US levels, fostering multilingual models and diverse benchmarks.

Expert‑Public Perception Gap

73% of industry experts are optimistic about AI’s impact on work, compared with only 23% of the general public. Trust in government regulation is lowest in the US (31% confidence), while confidence in EU oversight is higher across surveyed nations.

Report link: https://hai.stanford.edu/ai-index/2026-ai-index-report