AI Compute Thresholds Compared: How the EU, California, New York, Illinois and South Korea Define Frontier Models
Most laws aimed at the most powerful AI models decide who is covered by counting training compute. The unit is FLOP, the total number of floating-point operations used to train a model. This page compares the thresholds now written into law, what each one triggers, and when the obligations start.
Last updated: 15 September 2026.
The short version
The EU sets its line at 10^25 FLOP. California, New York, Illinois and South Korea all use 10^26, which is ten times higher. The three US states add a revenue test of $500 million a year for their strictest duties. The EU’s number is a rebuttable presumption, and the European Commission can also designate models below it.
Comparison table
| Jurisdiction | Law | Compute line | Other tests | In force | Main duties |
|---|---|---|---|---|---|
| European Union | AI Act, Article 51 | 10^25 FLOP (presumed systemic risk) | Commission can designate below the line; providers can rebut | GPAI duties from 2 Aug 2025; Commission enforcement from 2 Aug 2026 | Model evaluation, adversarial testing, incident reporting, cybersecurity |
| California | SB 53 (Transparency in Frontier AI Act) | More than 10^26, including fine-tuning and RL | “Large” developers: over $500m revenue | 1 Jan 2026 | Published frontier AI framework, transparency reports, incident reporting, whistleblower protections |
| New York | RAISE Act (as amended March 2026) | More than 10^26, including fine-tuning and RL | Over $500m revenue for large developers | 1 Jan 2027 | Framework, transparency reports, 72-hour incident reports, disclosure filing with DFS |
| Illinois | SB 315 (AI Safety Measures Act) | More than 10^26 | Over $500m revenue | 1 Jan 2027; framework and audit duties 1 Jan 2028 | Framework, whistleblower protections, annual independent audits |
| South Korea | AI Basic Act and Enforcement Decree | At least 10^26 cumulative | Plus criteria set by the science ministry | 22 Jan 2026 (fines deferred during a grace period of at least one year) | Lifecycle risk management, user protection, reporting to the ministry |
Jurisdiction notes
European Union
Article 51 of the AI Act presumes a general-purpose AI model has “high-impact capabilities” once its training compute passes 10^25 FLOP. Providers of those models carry extra duties covering evaluation, risk mitigation, serious incident reporting and cybersecurity. A provider can argue its model is less capable than the leading systems and ask to escape the designation. The Commission can go the other way too, designating a smaller model based on its capabilities.
The Commission can change the number by delegated act, and the Act names improvements in algorithms and hardware as reasons to do so. Commission guidelines separately treat models trained above roughly 10^23 FLOP that generate text or images as general-purpose AI in the first place, so 10^25 is the second gate.
Obligations for general-purpose models have applied since 2 August 2025. The Commission’s power to enforce them began on 2 August 2026. Models already on the market before August 2025 have until 2 August 2027 to comply. The Digital Omnibus on AI, in force since 27 July 2026, pushed back several high-risk system deadlines but left the general-purpose model rules on their original schedule.
California
SB 53 defines a frontier model as a foundation model trained with more than 10^26 integer or floating-point operations. The count includes later fine-tuning, reinforcement learning and other material changes, so post-training can push a model over. Any developer that crosses the line is a “frontier developer”. Those with more than $500 million in annual revenue are “large frontier developers” and carry the heaviest duties, including a published safety framework.
Critical safety incidents must be reported within 15 days, or within 24 hours when there is an imminent risk of death or serious injury. Penalties reach $1 million per violation. The state must review the definition every year.
New York
The original RAISE Act, signed in December 2025, used a compute threshold plus a $100 million compute-cost floor and also covered distilled models. Amendments signed on 27 March 2026 removed the cost floor and the distillation clause and copied California’s definition almost word for word.
New York still differs in a few places. Incidents must be reported within 72 hours. Penalties are $1 million for a first violation and $3 million after that. Large developers must file a disclosure statement with a new office inside the Department of Financial Services before operating in the state. The law takes effect on 1 January 2027.
Illinois
Governor JB Pritzker signed SB 315 on 6 July 2026. It follows California’s compute and revenue tests. Its main addition is a requirement for annual independent audits of whether developers actually follow their own safety frameworks, the first such mandate in a US state law. The Act takes effect on 1 January 2027, with the framework and audit obligations starting on 1 January 2028.
South Korea
The AI Basic Act and its Enforcement Decree took effect on 22 January 2026. Systems trained with at least 10^26 FLOP of cumulative compute, and meeting criteria set by the Ministry of Science and ICT, count as high-performance AI. Their operators must run lifecycle risk management and report results to the ministry. Fines are deferred for a grace period of at least a year except in serious cases.
Thresholds that no longer apply
US Executive Order 14110
The October 2023 executive order on AI required developers to report training runs above 10^26 operations, or 10^23 for models trained mainly on biological sequence data. It was the first official use of the 10^26 line. President Trump revoked it on 20 January 2025.
The AI Diffusion Rule
The January 2025 export control framework controlled the weights of closed models trained above 10^26 operations. The rule was rescinded in May 2025 before its compliance date.
Federal proposals
The White House has urged Congress to pass a national framework that would pre-empt state AI laws. A House discussion draft known as the Great American AI Act also uses 10^26 FLOP. Until Congress or the courts act, the state laws above remain in force.
China has no compute threshold in law. Its rules for generative AI services work through registration and security assessments run by the Cyberspace Administration of China.
How much is 10^26 FLOP?
For a dense transformer, a common rule of thumb puts training compute at about six times the parameter count times the number of training tokens.
A 70-billion-parameter model trained on 15 trillion tokens comes out near 6 x 10^24 FLOP. That is below the EU line. A 400-billion-parameter model on 15 trillion tokens lands around 3.6 x 10^25, above the EU line but well short of 10^26. Reaching 10^26 takes something like a trillion-parameter-class model trained on tens of trillions of tokens, or heavy post-training on top of a very large base model.
Grok 3, released in February 2025, is widely cited as the first public model past 10^26. Epoch AI counted more than 30 models above 10^25 by mid-2025 and projects about 30 above 10^26 by 2027, rising past 200 by 2030.
Why compute is an imperfect proxy
Training efficiency keeps improving, so a model trained with less compute today can match one that needed more a year ago. Distillation and pruning let developers build smaller models that inherit much of a larger model’s behaviour. Reasoning models also spend a lot of compute at inference time, when answering questions, which training-compute thresholds don’t measure.
That is why the EU keeps a capability-based route alongside its number, and why California requires an annual review. Expect the lines to move.