Compute thresholds are the way regulators tried to say "frontier model" with a number: the total count of floating-point operations (FLOP) used to train a model (compute). The logic comes from scaling laws: if performance improves predictably with compute (Kaplan et al., 2020), compute serves as a proxy, crude but measurable before deployment, for what a model might be able to do.
Two figures became jargon. US Executive Order 14110 (30 October 2023) required reporting on any model trained with more than 10²⁶ operations (10²³ if trained primarily on biological sequence data) and on clusters capable of 10²⁰ operations per second; it was revoked on 20 January 2025. The EU AI Act presumes that a general-purpose model poses "systemic risk" if its cumulative training compute exceeds 10²⁵ FLOP (Regulation (EU) 2024/1689, 2024, Art. 51). California took up the 10²⁶ figure in its 2025 law SB 53. For scale: Epoch AI estimates that GPT-4 took about 2 × 10²⁵ FLOP, and by mid-2025 it counted more than thirty models above 10²⁵.
The song exaggerates and switches units: "One E thirty flops a second" (osmarks, n.d.), that is, 10³⁰ operations per second, a rate rather than a total (its author says he was not alluding to any regulation). The confusion is instructive: FLOP (a quantity) and FLOP/s (a rate) get mixed up all the time, even in the specialized press.
The criticisms: a fixed threshold ages quickly, because algorithmic efficiency makes it possible to reach the same capabilities with less compute, as Chinchilla showed by beating Gopher, four times its size, on the same compute budget (Hoffmann et al., 2022) (see also distillation). Moreover, training compute ignores the test-time compute of reasoning models. Defenders reply that it is imperfect but verifiable, which is rare in AI regulation.