Governance, Policy & LawTraining & Post-trainingSafety & Alignment

compute threshold

nounAlso: training compute threshold, FLOP threshold, compute-based threshold

In plain wordsA legal cutoff based on how much computation was used to train an AI model. Models trained with more than that amount face extra rules, on the theory that bigger training runs produce more capable models.

A quantity of training computetraining compute, n.: The total number of arithmetic operations performed to train a model. It is the standard yardstick for the size of an AI training effort, and some laws use it to decide which models face extra rules., usually the total number of integer or floating-point operations (FLOPFLOP, n.: A single arithmetic step, such as multiplying or adding two numbers. Counting them measures how much computing went into an AI model, and laws use such counts to decide which models face extra rules.) used to train a model, above which legal or voluntary obligations apply. Executive Order 14110Executive Order 14110, prop. n.: A sweeping 2023 order by US President Joe Biden that told federal agencies how to manage AI risks and required makers of the largest AI models to report their safety tests to the government. It was revoked in 2025. required reports for models above 10²⁶ operations (10²³ for models trained mainly on biological sequence data); the EU AI ActEU AI Act, prop. n.: The European Union's law on artificial intelligence, the first broad AI law anywhere: it bans some uses, sets strict checks for risky ones and puts duties on the makers of the biggest AI models. presumes a general-purpose AI model with systemic riskgeneral-purpose AI model with systemic risk, n.: The EU's label for the very largest and most capable AI models, whose failure or misuse could cause harm across society, so their makers must test them for dangers and report serious incidents. above 10²⁵ FLOP; the AI Diffusion RuleAI Diffusion Rule, prop. n.: A 2025 US rule, cancelled before it took effect, that would have sorted the world's countries into three groups with different limits on how many advanced AI chips they could buy. of January 2025, rescinded before it took effect, would have controlled exports of the closed weights of models above 10²⁶; and California's SB 53SB 53, prop. n.: A 2025 California law requiring developers of the most powerful AI models to publish their safety plans, report serious safety incidents to the state and protect employees who raise alarms. and New York's RAISE ActRAISE Act, prop. n.: A New York law requiring the largest AI developers to publish how they manage catastrophic risks and to report serious safety incidents quickly to a new state office. It takes effect in 2027. apply to developers of models above 10²⁶, counting fine-tuning and other later training. Thresholds are attractive because compute is measurable before deployment, correlates with capability under scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments. and is hard to hide. Critics call them a crude proxy that algorithmic progress steadily erodes, that misses small but dangerous specialized models and that ignores inference-timetest-time compute, n.: The computing effort an AI model spends while answering a question, as opposed to while being trained. Letting a model think longer or try many answers can make it markedly smarter. scaling, so most regimes pair them with capability evaluations or allow regulators to update the number.
The model came in just under the 10²⁵ FLOP compute threshold, so the presumption of systemic risk did not apply.

Origin

Proposed in AI-governance research from about 2022–2023 (for example 'Frontier AI Regulation', Markus Anderljung and colleagues, 2023) as a tractable trigger for oversight; first written into binding government policy by Executive Order 14110 in October 2023 and into legislation by the EU AI Act in 2024.

Usage note

For scale, Epoch AI estimates that GPT-4 (2023) used about 2 × 10²⁵ FLOP; OpenAI has not disclosed the figure. Distinguish from a capability thresholdcapability threshold, n.: A level of dangerous ability — such as meaningfully helping someone make a biological weapon — that, once an AI model reaches it, obliges its developer to add stronger safeguards or hold the model back., which triggers obligations based on what a model can do rather than how much computation went into it.

Sources

Anderljung, M. et al. (2023). Frontier AI Regulation: Managing Emerging Risks to Public Safety. arXiv:2307.03718.
Heim, L. & Koessler, L. (2024). Training Compute Thresholds: Features and Functions in AI Regulation.

Defined in law or standards

EU AI Act (Reg. (EU) 2024/1689), Art. 51(2); Executive Order 14110 (2023), Sec. 4.2(b) (revoked)

Mentioned in

AI regulationAI regulation, n.: Laws and binding government rules about artificial intelligence: what AI may be used for, what checks its makers must run, and what penalties apply when they break the rules., FLOPFLOP, n.: A single arithmetic step, such as multiplying or adding two numbers. Counting them measures how much computing went into an AI model, and laws use such counts to decide which models face extra rules., frontier AIfrontier AI, n.: The most capable general-purpose AI systems of the moment — those at or beyond the leading edge — which governments single out for extra safety attention., frontier modelfrontier model, n.: One of the most capable AI models in existence at a given time, built by a handful of leading labs at great cost, and the main target of AI safety rules., capability thresholdcapability threshold, n.: A level of dangerous ability — such as meaningfully helping someone make a biological weapon — that, once an AI model reaches it, obliges its developer to add stronger safeguards or hold the model back., compute governancecompute governance, n.: Using control over the computer chips and data centers that AI depends on — who can buy them, where they go, how they are used — as a way to oversee powerful AI., Executive Order 14110Executive Order 14110, prop. n.: A sweeping 2023 order by US President Joe Biden that told federal agencies how to manage AI risks and required makers of the largest AI models to report their safety tests to the government. It was revoked in 2025., general-purpose AI model with systemic riskgeneral-purpose AI model with systemic risk, n.: The EU's label for the very largest and most capable AI models, whose failure or misuse could cause harm across society, so their makers must test them for dangers and report serious incidents., RAISE ActRAISE Act, prop. n.: A New York law requiring the largest AI developers to publish how they manage catastrophic risks and to report serious safety incidents quickly to a new state office. It takes effect in 2027., risk-based approachrisk-based approach, n.: A way of regulating AI in which the strictness of the rules depends on how much harm a use could cause: bans for the worst uses, heavy checks for risky ones, little or nothing for the rest., SB 1047SB 1047, prop. n.: A 2024 California bill that would have required makers of the biggest AI models to adopt safety plans and take care to prevent catastrophes. It passed the legislature, but the governor vetoed it., SB 53SB 53, prop. n.: A 2025 California law requiring developers of the most powerful AI models to publish their safety plans, report serious safety incidents to the state and protect employees who raise alarms., training computetraining compute, n.: The total number of arithmetic operations performed to train a model. It is the standard yardstick for the size of an AI training effort, and some laws use it to decide which models face extra rules., AI Diffusion RuleAI Diffusion Rule, prop. n.: A 2025 US rule, cancelled before it took effect, that would have sorted the world's countries into three groups with different limits on how many advanced AI chips they could buy., Collingridge dilemmaCollingridge dilemma, n.: The trap that a new technology is easiest to control when little is known about its harms, but by the time the harms are clear it is too widespread to change easily., dual-use foundation modeldual-use foundation model, n.: A US legal term for a general-purpose AI model that serves everyday uses but is capable enough that it could help someone build weapons, mount cyberattacks or evade human control., IAEA for AIIAEA for AI, phr.: The idea of an international agency for the most advanced AI, modeled on the UN's nuclear watchdog, which inspects countries' nuclear programs to keep them safe and peaceful., know-your-customer for computeknow-your-customer for compute, n.: Rules that would make cloud companies check who their customers are before renting them enough computing power to train a large AI model, much as banks must verify who opens an account.