SI Glossary · Policy & governance
Compute Threshold
A compute threshold uses the amount of computation spent training a model as a rough proxy for its capability and risk. Training compute is measured in FLOPs (floating-point operations), the total number of arithmetic operations performed.
Thresholds in law
| Rule | Threshold | Effect |
|---|---|---|
| EU AI Act | 10²⁵ FLOPs | General-purpose models presumed to pose “systemic risk”: evaluations, incident reporting, cybersecurity |
| California SB 53 | 10²⁶ FLOPs (plus revenue test for “large” developers) | Publish a safety framework, report critical incidents |
| New York RAISE Act | 10²⁶ FLOPs (plus a $500M revenue test for “large” developers) | Safety protocols, 72-hour incident reporting (from 2027) |
| U.S. EO 14110 (2023, revoked January 2025) | 10²⁶ FLOPs | Reporting to the federal government |
For scale: 10²⁶ is a 1 followed by 26 zeros. Frontier training runs crossed that level in the mid-2020s.
Pros and cons
Pros: easy to measure and verify, known before a model is released, and it only catches the biggest players.
Cons: a crude proxy. Algorithmic efficiency means smaller runs keep getting more capable, distillation can transfer abilities cheaply, and post-training and inference compute aren’t counted. That’s why some laws combine compute with revenue or capability tests.
Written by
Luka Kušec · Editor
Editor of SI.info. Writes about Super Intelligence, technology policy and the people building frontier models.