Examples
Worked examples
- Is an instance
A model card declaring '~3.8 x 10^25 training FLOPs'.
- Is an instance
A regulator's classification of a model as systemic-risk based on the FLOPs estimate.
Counter-examples
Looks similar, but isn't
- Not an instance
Reporting only GPU-hours (related but hardware-coupled).
- Not an instance
Reporting energy consumption (related but not identical).
Editorial commentary
A compute (FLOPs) estimate is the total number of floating-point operations consumed in training a model, conventionally reported as a single figure such as 3.0 x 10^25 FLOPs. It is used as a proxy for training-run scale in both scientific reporting (comparing models across research groups) and regulation, because it can in principle be estimated or disclosed without revealing proprietary details of a model’s architecture or training data.
Its regulatory role, and what changed
The EU AI Act attaches a specific consequence to this figure: under Article 51(2), a general-purpose AI model is presumed to carry systemic risk once its cumulative training compute exceeds 10^25 FLOPs, triggering the Act’s heaviest obligations under Articles 53 and 55 (technical documentation, model evaluation, adversarial testing, serious-incident reporting). This is a rebuttable presumption, not an automatic classification, and the European Commission can revise the threshold by delegated act. A comparable US mechanism existed briefly: Executive Order 14110 (2023) set a 10^26-FLOP reporting trigger for dual-use foundation models, but it was revoked in January 2025 by Executive Order 14179 and not replaced with an equivalent FLOPs threshold. As of this writing, the EU AI Act figure is the only FLOPs-based regulatory line still active among major jurisdictions.
Limitations as a metric
Training-compute FLOPs measures training-time scale only — it says nothing about training-data quality, architecture efficiency, or (see inference) the growing share of total compute some newer architectures spend after training, at inference time. Two models with identical training-FLOPs totals can differ substantially in capability.
References
- EU AI Act, Article 51(2).
- Executive Order 14179 (2025), revoking Executive Order 14110 (2023).
- See also: Frontier model, Parameter count, Training carbon footprint.
Also known as
training FLOPs · training compute
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Compute (FLOPs estimate)"
vocab-term-identifier="https://casrai.org/dictionary/term/compute-flops-estimate" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/compute-flops-estimate",
"name": "Compute (FLOPs estimate)",
"identifier": "https://casrai.org/dictionary/term/compute-flops-estimate",
"description": "The total floating-point operations consumed by training a model, conventionally reported as a single number (e.g., 3.0 x 10^25 FLOPs) used as a regulatory and scientific proxy for training-run scale.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/compute-flops-estimate",
"sameAs": [
"training FLOPs",
"training compute"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-08-22T15:43:59",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
