Why do AI laws count training compute?

Try it: the compute calculator runs both estimation methods below and shows which current and proposed thresholds a training run would cross.

Why do AI laws count training compute?

AI laws count training compute because, unlike capability, it can be written into a statute as a number, known before a model is finished, and checked by outsiders. Lawmakers want to regulate the most capable models and leave the rest alone, and compute is the proxy they chose.

Training compute is the total number of arithmetic operations a computer performs while training a model. It is usually written as a power of ten. A model trained with 10²⁵ operations used ten times more compute than one trained with 10²⁴.

Lennart Heim and Leonie Koessler of the Centre for the Governance of AI set out the case in a 2024 paper. Compute is linked to capability and risk. It can be measured before a model is finished, even before training starts. Outside parties can check it, because it depends on physical chips and electricity.

Their paper also says compute is an imperfect proxy. They argue a threshold should trigger scrutiny, such as evaluations, rather than decide on its own what rules apply.

Put another way, a compute threshold decides who has to look closely at a model. It does not decide whether the model is dangerous.

Which laws use which thresholds?

The EU sets its systemic-risk line at 10²⁵ floating-point operations; the U.S. federal measures, California, New York, and South Korea all use 10²⁶.

Rule	Threshold	What crossing it triggers	Status (Oct 2026)
U.S. Executive Order 14110, sec. 4.2	>1026 operations; >1023 for models trained mainly on biological sequence data	Reporting to Commerce under the Defense Production Act	Revoked January 20, 2025
BIS proposed reporting rule (Sept. 2024)	>1026 operations	Quarterly reports to Commerce	Proposed; its basis, EO 14110, was revoked
EU AI Act, Article 51	>1025 FLOP	Presumption of systemic risk; notice within two weeks	In force
EU Commission guidelines (indicative)	>1023 FLOP	Presumption that a model is general-purpose, if it can also generate language, images or video	Non-binding guidance
California SB 53	>1026 operations, including fine-tuning	“Frontier model” duties	In force January 1, 2026
New York RAISE Act	>1026 operations	“Frontier model” duties	Takes effect January 1, 2027
South Korea AI Basic Act decree	≥1026 FLOP	“High-performance AI” safety duties	In force January 22, 2026
U.S. AI Diffusion Rule, ECCN 4E091	≥1026 operations, closed weights	Export license for model weights	Still in the regulations but not enforced
S. 2938 (proposed)	>1026 operations	Mandatory federal evaluation before deployment	Introduced September 29, 2025

For what each law requires once a model is covered, see frontier AI laws. For the export rule, see AI export controls.

How is training compute estimated?

Training compute is estimated in one of two ways: by counting the operations the training required, or by counting the hardware and how long it ran. Epoch AI, a research group that tracks large models, documents both.

Method 1: count the operations

For a dense transformer language model, the usual rule of thumb is:

Training compute ≈ 6 × parameters × training tokens

The 6 comes from about 2 operations per parameter per token in the forward pass and about 4 in the backward pass. For mixture-of-experts models, Epoch says to count only the parameters active for each token.

A worked example. DeepSeek’s V3 technical report lists 37 billion activated parameters and 14.8 trillion pre-training tokens. Six times those two numbers is about 3.3 × 10²⁴ operations. On this rough estimate, pre-training sat below the EU’s 10²⁵ line. The calculation is ours and covers pre-training only.

Method 2: count the hardware

The second method multiplies number of chips, peak speed per chip, a utilization rate, and training time.

A hypothetical example: 10,000 chips, each with a peak of 10¹⁵ operations per second, running at 40% utilization for 90 days. That is 10,000 × 10¹⁵ × 0.4 × 7.8 million seconds, or about 3 × 10²⁵ operations. That run would cross the EU threshold but not California’s.

The weak point is the utilization rate. Epoch prefers counting operations when the details are known because hardware estimates rely on that assumption.

When the developer does not disclose

Outsiders often have neither the parameter count nor the chip count. Epoch then estimates compute from benchmark scores. Its own tests found those estimates were off by a factor of 2.2 on average, with a 90% confidence interval of 1.1 to 5.0. It calls them speculative.

Regulators do not have to guess. The EU, California and New York put the duty on the developer, who knows its own numbers.

How many models are above the thresholds?

Epoch AI’s dataset of models trained with at least 10²⁵ FLOP, last updated November 21, 2025, lists 33 publicly disclosed models. Its estimates put two of them above 10²⁶: xAI’s Grok 3, at about 4.6 × 10²⁶, and OpenAI’s GPT-4.5, at about 2.1 × 10²⁶, an estimate Epoch labels speculative. Epoch’s June 2025 summary of the same list had put GPT-4.5 below 10²⁶ and Anthropic’s Claude Opus 4, speculatively, above it.

Those 33 models came from 12 developers. The first was GPT-4, released in March 2023. Epoch says a run at that scale costs tens of millions of dollars with current hardware.

Those figures are Epoch’s estimates, not developer disclosures. Developers are not required to publish their exact training compute, so any outside count is an estimate.

That is the practical difference between the two lines. By Epoch’s count, a 10²⁵ line caught 33 models by late 2025; a 10²⁶ line caught an estimated two.

What happened to the 2023 U.S. threshold?

The 2023 U.S. threshold, a 10²⁶-operation reporting trigger in Executive Order 14110, was revoked in January 2025, but the number moved into state law.

President Biden signed Executive Order 14110 on October 30, 2023. Section 4.2 used the Defense Production Act to require companies developing potential dual-use foundation models to report to the government on training plans, ownership and protection of model weights, and red-team results.

Until Commerce set its own conditions, the order applied to any model trained with more than 10²⁶ integer or floating-point operations, or more than 10²³ for models trained mainly on biological sequence data. It also covered computing clusters capable of 10²⁰ operations per second.

In September 2024, the Commerce Department’s Bureau of Industry and Security proposed a rule to make that reporting quarterly.

On January 20, 2025, President Trump signed Executive Order 14148, which revoked Executive Order 14110 by name. We found no final version of the 2024 reporting rule.

California adopted 10²⁶ in SB 53. New York adopted it in the final RAISE Act. A Senate bill, S. 2938, uses it again.

What are the main criticisms of compute thresholds?

The main criticism is that training compute is a weak and aging proxy for risk. Sara Hooker, then at Cohere, made the case in a 2024 essay. She argued that thresholds “as currently implemented are shortsighted and likely to fail to mitigate risk.” Her main points:

• Smaller models keep beating bigger ones. Better data, distillation and training methods raise capability without more training compute.
• Training compute ignores inference. Techniques such as chain-of-thought reasoning, best-of-n sampling and retrieval add capability after training, and a training threshold does not count them.
• The numbers have no published rationale. Neither 10²⁵ nor 10²⁶ was derived from a measured risk level.
• Counting is ambiguous. Mixture-of-experts models, ensembles and multi-stage pipelines raise questions about what to add up.
• Loose definitions invite manipulation. Without clear rules on how operations are counted, Hooker warns, FLOP is “easy to manipulate.”

Laws have tried to answer some of this. California counts fine-tuning and later modifications toward the total. The EU lets the Commission designate a model below the threshold and change the number by delegated act. New York and California add a revenue test, so the heaviest duties fall on companies above $500 million.

Apart from the EU’s power to change the number, those patches add tests around the threshold rather than moving it. Hooker’s central objection is to the fixed number itself: it drifts as hardware and methods improve.

What would replace a compute threshold?

The main alternatives are capability evaluations, revenue or user-count tests, and designation by a regulator. Each has its own weakness. Evaluations happen late, after a model is trained. Revenue tests miss well-funded labs with no revenue. Designation depends on regulator judgment.

Hooker proposed thresholds that move with the frontier, such as a percentile of the most compute-intensive models, combined with other risk measures. Heim and Koessler proposed keeping compute as an initial filter and using evaluations to decide what follows.

The policy fight over whether a license should sit on top of any of these filters is in should superintelligence require a license?

Why do compute thresholds persist despite the criticism?

Compute thresholds persist because they are cheap to apply and hard to dispute. They are criticized because the link between compute and danger keeps loosening, which is why several laws now pair them with a revenue test, a designation power, or evaluations.

Thresholds and statuses are stated as of October 2026. This guide is general information, not legal advice.