Skip to main content
Bureau of AI
Category: EU AI Act & GPAI

10^25 FLOP Threshold

Also known as: FLOP Threshold (10^25), 10^25 FLOP systemic risk threshold, AI Act training compute threshold
Simply put

The 10^25 FLOP threshold is a measure of computing power used to train an AI model under the European Union's AI Act. When the total amount of computation used to train a general-purpose AI model exceeds this level, the model is presumed to carry so-called 'systemic risk,' which triggers additional legal obligations for its provider. It is a criterion set by EU regulators and applies within the EU AI Act framework rather than universally across all jurisdictions.

Formal definition

Under the EU AI Act, a threshold of 10^25 floating-point operations (FLOP) of cumulative training compute is used as a presumption criterion for classifying a general-purpose AI (GPAI) model as posing systemic risk (referenced in Article 51 of the AI Act per European Commission materials). Exceeding this threshold typically brings additional obligations, which commonly include measures such as model evaluation, adversarial testing (red teaming), and systemic risk assessment and mitigation. This 10^25 FLOP figure is distinct from a separate 10^23 FLOP figure that appears in the European Commission's guidance on GPAI models as an indicative criterion for whether a model qualifies as a GPAI model at all; based on the evidence and Commission materials, the 10^23 figure originates in non-binding Commission guidelines rather than the binding text of the AI Act, and the two thresholds address different classification questions and should not be conflated. Note also that FLOP-based compute counts are an imperfect proxy for model capability or risk, and commentary has identified limitations of FLOPs as a classification metric; this entry does not address how training compute is measured or attributed in practice, which remains an area of ongoing interpretation.

Why it matters

The 10^25 FLOP threshold matters because it is one of the few quantitative bright lines in the EU AI Act, converting an otherwise qualitative judgment about AI risk into a measurable criterion. When cumulative training compute for a general-purpose AI (GPAI) model exceeds this level, the AI Act (per Article 51(1)(a) and (2) as described in European Commission materials) presumes the model carries 'systemic risk,' which typically triggers additional obligations for the provider such as model evaluation, adversarial testing (red teaming), and systemic risk assessment and mitigation. For compliance officers and legal teams, this makes the threshold a practical trigger point: it can determine which set of obligations a model provider falls under within the EU framework.

The threshold is also significant because it is frequently misunderstood or conflated with a separate figure. Commission guidance materials reference an indicative 10^23 FLOP figure that addresses a different question — whether a model qualifies as a GPAI model in the first place — and based on Commission materials that figure appears in non-binding guidelines rather than in the binding text of the AI Act. The 10^25 figure, by contrast, is tied to the systemic-risk presumption in the AI Act text. Treating the two numbers as interchangeable, or treating the 10^23 guidance figure as though it were binding law, is a common error that can lead to misclassification of obligations.

Professionals should also recognize that a compute count is an imperfect proxy for capability or risk. Commentary, including analysis published in the Harvard Journal of Law & Technology Digest, has identified limitations of FLOPs as a model-classification metric. The threshold reduces ambiguity for regulators but does not by itself measure a model's actual capabilities or harms, and the details of how training compute is counted and attributed remain an area of ongoing interpretation. The threshold applies within the EU AI Act framework and should not be assumed to apply across other jurisdictions.

Who it's relevant to

GPAI Model Providers and Developers
Organizations training large general-purpose AI models need to track cumulative training compute, because exceeding 10^25 FLOP typically triggers the systemic-risk presumption under the EU AI Act and the additional obligations that commonly accompany it, such as model evaluation, red teaming, and systemic risk assessment and mitigation.
Compliance Officers and Legal Professionals
Those advising on EU AI Act obligations use the threshold as a practical classification trigger. They must be careful not to conflate the 10^25 FLOP systemic-risk threshold with the indicative 10^23 FLOP figure that appears in non-binding Commission guidance and addresses whether a model qualifies as a GPAI model at all.
AI Governance and Policy Specialists
Governance teams and policy analysts monitoring how quantitative criteria are used in regulation will note that the threshold applies within the EU AI Act framework rather than universally, and that FLOP-based counts are an imperfect proxy for capability or risk, with limitations identified in published commentary.
Model Risk and Assurance Functions
Teams responsible for evaluating models can treat the threshold as one input among many. It marks a regulatory trigger point within the EU framework, but does not by itself measure a model's actual capabilities or harms, and the details of how compute is measured and attributed remain subject to ongoing interpretation.

Inside 10^25 FLOP Threshold

Compute threshold (10^25 FLOP)
A quantitative measure of the total computational effort used to train a model, expressed in floating-point operations (FLOP). In the EU AI Act, a threshold expressed at this order of magnitude functions as one indicator used to identify general-purpose AI (GPAI) models that may carry so-called systemic risk. The threshold is a trigger for further scrutiny rather than a standalone determination of risk.
Cumulative training compute
The figure refers to compute accumulated during a model's training process, not to inference-time or per-operation compute. It aggregates the operations required to produce the trained model, which is why it is used as a proxy for model scale and capability.
Presumption mechanism
In the AI Act framing, exceeding the compute threshold typically creates a presumption that a GPAI model has high-impact capabilities and may pose systemic risk. Presumptions of this kind are generally rebuttable and can also be supplemented by other criteria, so the number is not the sole basis for classification.
Threshold versus indicator distinction
The 10^25 FLOP figure is best understood as a screening indicator tied to obligations for providers of the most capable GPAI models. It is a threshold for triggering additional obligations, not a measure of a model's actual performance, safety, or fairness.
Relationship to the AI Act text and guidance
The systemic-risk compute figure appears in the AI Act's provisions on GPAI models. A separate, lower compute figure sometimes cited in connection with GPAI classification (on the order of 10^23 FLOP) appears in non-binding European Commission guidelines rather than in the legally binding text of the Act; the two figures should not be treated as equivalent or as carrying the same legal weight.

Common questions

Answers to the questions practitioners most commonly ask about 10^25 FLOP Threshold.

Does crossing the 10^25 FLOP threshold mean a model is automatically classified as dangerous or high-risk?
No. In the EU AI Act framework, the 10^25 FLOP figure is associated with a presumption that a general-purpose AI model carries systemic risk, which triggers additional obligations for the provider. It is a trigger for heightened scrutiny and specific duties, not a determination that the model is unsafe. A compute figure is a proxy indicator; it does not by itself measure a model's actual capabilities, risks, or harms. Providers may also be designated on other grounds, and the presumption can in principle be rebutted, so treating the threshold as an automatic 'dangerous' label mischaracterizes its function.
Is the 10^25 FLOP threshold a universal regulatory standard that applies everywhere?
No. This threshold is specific to the EU AI Act's treatment of general-purpose AI models and applies within that jurisdiction and scope. It is not a globally harmonized standard, and other frameworks address compute or capability differently or not at all. Compute-based thresholds also appear in some non-EU policy instruments, but the specific value and the obligations attached to it differ. Professionals should not assume that meeting or falling below the EU figure carries the same legal meaning in other jurisdictions, nor that it maps onto model risk management guidance such as SR 11-7, which does not use compute thresholds.
How is training compute typically estimated for the purpose of comparing against the threshold?
Training compute is commonly estimated from the model's architecture, parameter count, and the volume of training data, or from hardware utilization records such as accelerator-hours multiplied by effective throughput. There is no single mandated methodology, and estimates can vary depending on assumptions about hardware efficiency and what phases of training are counted. Organizations should document their measurement method, assumptions, and boundaries (for example whether fine-tuning or multiple training runs are aggregated) so that the estimate is reproducible and defensible. Because methodology choices materially affect whether a model appears to cross the threshold, transparency about the calculation is important.
What is a practical sense of scale for 10^25 FLOP?
10^25 floating-point operations is an extremely large quantity of computation, characteristic of large-scale frontier model training runs rather than typical enterprise or fine-tuning workloads. One way some practitioners express such figures is in petaFLOP-days, using the convention that one petaFLOP-day equals 8.64 × 10^19 FLOP; on that basis 10^25 FLOP corresponds to roughly 1.16 × 10^5 petaFLOP-days. Unit conversions should be checked carefully, since errors of orders of magnitude are easy to introduce. The takeaway for most organizations is that the threshold sits well above ordinary model development activity.
Who is responsible for assessing whether a model meets the threshold, and when should that assessment happen?
Responsibility typically falls on the provider of the general-purpose AI model, since obligations under the EU AI Act framework attach to that role. Because training compute can only be fully quantified once a training run is planned or completed, organizations often estimate expected compute during planning and confirm actual compute afterward. Embedding this check into project intake and model governance workflows, rather than treating it as a one-time legal review, helps ensure the assessment is timely. Legal counsel should confirm the current applicable obligations, as regulatory interpretation and guidance in this area continue to evolve.
What should an organization document to support its position on the threshold?
Useful documentation typically includes the compute estimation methodology and its assumptions, the data sources used (such as hardware logs or architectural calculations), the boundaries of what was counted, and the resulting figure with a margin of uncertainty. Retaining this record supports auditability and helps the organization respond if a regulator or internal reviewer questions the classification. Where a model sits close to the threshold, more rigorous documentation and a conservative interpretation are advisable. This record-keeping practice reduces compliance and model risk exposure but does not eliminate it, and it should be reviewed against current regulatory guidance rather than treated as settled once completed.

Common misconceptions

The 10^25 FLOP threshold determines whether a model is high-risk under the EU AI Act.
The threshold relates specifically to general-purpose AI models and a presumption of systemic risk, which is a distinct concept from the AI Act's separate risk-tiering of AI systems (such as its high-risk classifications). Exceeding the compute figure typically triggers a presumption or additional obligations rather than automatically labeling a model high-risk across the whole framework.
Any compute figure mentioned in connection with GPAI models carries the same legal force.
A lower figure on the order of 10^23 FLOP has been associated with GPAI classification in non-binding European Commission guidelines, whereas the higher figure is tied to the systemic-risk provisions. Guidelines and the binding text of the Act do not carry identical legal weight, and conflating the two misstates what the law requires.
Being above or below the compute threshold settles a model's actual risk.
The threshold is a proxy indicator based on training scale, not a direct measurement of capability, safety, or societal impact. Presumptions of this type are generally rebuttable and may be complemented by other criteria, so the compute figure alone does not establish or eliminate risk.

Best practices

Track and document cumulative training compute for GPAI models using a consistent methodology, since the figure serves as a trigger for further obligations and may need to be demonstrated to regulators.
Distinguish the systemic-risk compute figure in the binding AI Act provisions from the lower figure that appears in non-binding Commission guidelines, and record which source each number derives from when advising on obligations.
Treat exceeding the threshold as a trigger for additional scrutiny and obligations rather than as a final risk determination, and prepare to address any rebuttable presumption with supporting evidence.
When converting compute figures between units, verify the arithmetic carefully; for reference, one petaFLOP-day equals 8.64 x 10^19 FLOP, so 10^25 FLOP is on the order of 100,000 petaFLOP-days, not tens of millions.
Monitor evolving guidance and delegated measures, because the practical interpretation of compute-based triggers may be refined over time and current positions should not be presented as fully settled.
Coordinate compute-threshold assessments between AI governance functions (policy, accountability, oversight) and model risk functions (measurement and monitoring), keeping the two domains distinct while ensuring the threshold analysis informs both.