Stanford AI Index 2026: investment tops $581 billion, industry builds every notable model

Stanford's Institute for Human-Centered AI published the ninth edition of its AI Index on 13 April 2026. The report, organised into nine chapters spanning research, technical performance, responsible AI, economy, science, medicine, education, policy and public opinion, is the most widely cited neutral data source on the state of the field, and this year's numbers describe a market that has moved from rapid growth to something closer to a capital super-cycle.

Global AI investment reached more than $581 billion in 2025, more than double the $253 billion recorded in 2024 and far above the previous 2021 peak of $360 billion. The United States accounted for $344 billion of the total. The report counts 50 notable models released by US organisations in 2025, with China narrowing the gap on quality and Europe contributing just two. Perhaps the most striking structural finding is that essentially every notable model of 2025 came from industry rather than academia or government, compared with a roughly even split a decade earlier.

Resource use is now a first-class topic. The Index estimates that AI training compute has grown more than threefold each year since 2022, that training a single frontier model such as Grok 4 emitted on the order of 72,000 tonnes of CO2-equivalent (with one alternative estimate near 140,000), and that the least efficient models at inference consume over ten times the emissions of the most efficient. On the developer side, GitHub hosted 5.58 million AI-related projects, five times the 2020 level and up 23.7% year on year. In public-opinion surveys 59% of respondents now say benefits outweigh drawbacks, up from 55%, but only 31% of Americans trust the government to regulate AI effectively. Chapters on policy and responsible AI track a continued rise in recorded AI incidents and in national legislative activity, though the Index cautions that reporting standards vary.

Read alongside the March policy moves in Washington and Brussels, the report highlights a widening gap between where capability and capital are concentrated and where oversight capacity sits. A handful of companies now set the technical frontier, fund most of the research and, increasingly, own the compute. For boards the Index is useful less as a forecast than as a benchmark: it shows what "normal" now looks like for spend, adoption and risk, against which an organisation's own AI posture can be judged.

What it means for leaders

• Use the Index as an external baseline in board reporting. Investment, adoption and incident trends give directors context for internal AI budgets and risk appetite statements.
• Add energy and emissions to AI impact assessments. With per-model training emissions now quantified, sustainability disclosures under CSRD and similar regimes will increasingly ask about AI workloads; ISO/IEC 42001 impact assessments can capture this.
• Plan for vendor concentration. Near-total industry control of frontier models means dependence on a few suppliers; document exit strategies and multi-model architectures.
• Track incident data systematically. Rising recorded incidents argue for internal AI incident logging aligned with NIST AI RMF Measure and Manage functions, before regulators mandate it.
• Mind the trust deficit. Low public confidence in regulation shifts scrutiny onto companies; transparent model documentation and clear human-oversight policies are reputational assets.