AI training compute calculator
Estimate how many operations a training run used, then see which compute thresholds in the EU AI Act, California’s SB 53, New York’s RAISE Act and South Korea’s AI Basic Act it would cross. Thresholds are stated as of October 2026.
- BelowEU Commission guidelines (indicative): >1023 FLOP
- BelowEU AI Act, Article 51: >1025 FLOP
- BelowCalifornia SB 53: >1026 operations, including fine-tuning
- BelowNew York RAISE Act: >1026 operations
- BelowSouth Korea AI Basic Act decree: ≥1026 FLOP
- BelowS. 2938 (proposed): >1026 operations
This is a rough estimate. Estimates by outsiders can be off by a factor of two or more, and the laws put the duty on the developer, who knows its own numbers. California counts fine-tuning and later modifications toward the total. New York and California add a revenue test, so the heaviest duties fall on companies above $500 million.
Crossing a binding line triggers notice, documentation or safety duties. The 1023 line is non-binding guidance, and S. 2938 is a bill, not law. None of the enacted laws listed requires a government license before training a frontier model. Thresholds and statuses are stated as of October 2026. General information, not legal advice. Sources and method: compute thresholds.
The interactive calculator needs JavaScript. The method is simple: training compute ≈ 6 × parameters × training tokens. The EU’s systemic-risk line is 1025 floating-point operations; California, New York and South Korea use 1026. The full table is in compute thresholds.
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. For a dense transformer language model, the usual rule of thumb is 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.
The second method multiplies number of chips, peak speed per chip, a utilization rate, and training time. Its weak point is the utilization rate.
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. A compute threshold decides who has to look closely at a model. It does not decide whether the model is dangerous. The criticism, and what each law requires once a model is covered, are in compute thresholds and frontier AI laws.