Article 51(2) of the EU AI Act (Regulation (EU) 2024/1689) presumes a general-purpose AI model has high-impact capabilities once the cumulative compute used to train it, measured in floating-point operations, exceeds 10^25. What the Act never spells out is how to actually count that number. Do earlier versions of a model contribute their compute to a later one’s tally? If a model is branched, is each branch’s compute counted separately or added together? A threshold this consequential, sitting with no counting methodology attached to it, is the whole story here.
What the threshold is and isn’t
Article 51(1) classifies a general-purpose AI model as presenting systemic risk if it meets either of two conditions: it has high-impact capabilities established through appropriate technical tools and methods, or the Commission has designated it as comparable, ex officio or following a qualified alert from the scientific panel, based on the criteria in Annex XIII. The 10^25 FLOP figure in Article 51(2) only speaks to the first route — it’s a rebuttable presumption that high-impact capabilities exist, not a definition of systemic risk itself, and not the only way a model ends up in scope. Article 51(3) lets the Commission adjust the threshold by delegated act to reflect the state of the art, including algorithmic improvements and more efficient hardware — so even the number itself isn’t fixed forever, only fixed until the Commission decides otherwise.
The contradiction at the centre
Recital 111 defines high-impact capabilities relative to the frontier: capabilities that match or exceed those recorded in the most advanced general-purpose AI models. That’s a moving target by design — as the frontier advances, the bar for “high-impact” moves with it. The 10^25 figure Article 51(2) actually operationalises this with is nothing of the sort: it’s fixed, absolute, and doesn’t move on its own between one delegated act and the next.
The practical consequence is exactly the one the Act’s own drafting implies but never resolves: a model that clears the relative, frontier-based bar this month may not clear it next month once competitors release something more capable, while the absolute FLOP number attached to that same model never changes at all. There’s a second problem sitting underneath this one — most frontier models are proprietary, so the actual comparator Recital 111’s test asks you to measure yourself against usually isn’t measurable at all. The relative definition and the absolute number were clearly meant to approximate each other. Nothing in the text guarantees they track.
The counting questions the text leaves open
Three practical questions decide whether a lot of real models sit above or below the line, and the Act answers only one of them.
Fine-tuning — this one has a real answer. Recital 111 states that the cumulative compute input used for training includes the total compute used for activities to improve the model’s capabilities prior to deployment, explicitly naming pre-training, synthetic data generation, and fine-tuning. So fine-tuning compute counts toward the cumulative total, at least for the fine-tuning that happens before a model’s initial deployment. What that doesn’t resolve is the harder version of the same question: whether fine-tuning or continued training applied after an earlier version was already placed on the market gets added to that version’s original tally, or starts a new count.
Prior iterations — genuinely unresolved. Nothing in the text says whether compute spent on earlier versions of a model contributes to a later version’s cumulative figure, which matters enormously for any provider that iterates on the same underlying architecture across releases rather than training each version from nothing.
Branches — also unresolved. If a single training run forks into multiple variants, nothing says whether each branch’s compute is counted independently or whether the branches’ compute is summed as one cumulative figure against the threshold.
Annex XIII gives the Commission its own set of proxies for the designation route — estimated training cost, time, or energy consumption, as alternatives to a direct FLOP count — but it’s unclear whether a provider doing its own self-assessment under Article 51(2) has to use the same criteria, and none of these proxies answer the iteration or branching questions regardless of which measure you use.
The other Annex XIII numbers
The FLOP threshold gets all the attention, but Annex XIII gives the Commission six other criteria to weigh when designating a model as comparable under Article 51(1)(b), and one of them is rarely quoted at all:
- the number of parameters in the model;
- the quality or size of the training dataset, for example measured in tokens;
- the amount of compute used to train it, measured in FLOPs, or expressed through a combination of other variables such as estimated cost, time, or energy consumption;
- the model’s input and output modalities — text-to-text, text-to-image, multimodal — along with advanced thresholds for high-impact capabilities specific to each modality, and the particular type of input or output involved, such as biological sequences;
- benchmarks and capability evaluations, including the number of tasks the model performs without additional training, its adaptability to learn new distinct tasks, its degree of autonomy and scalability, and the tools it can access;
- whether it has a large impact on the internal market because of its reach — which is presumed once the model has been made available to at least 10,000 registered business users established in the EU;
- the number of registered end-users.
What you do when you cross it
If a model meets the Article 51(1)(a) condition, the provider notifies the Commission without delay, and in any event within two weeks of the condition being met or of it becoming known that it will be met. The notification has to include the information needed to demonstrate the condition is satisfied. Providers aren’t limited to reporting the fact, either — in that same notification, a provider can put forward sufficiently substantiated arguments that, despite meeting the threshold, the model exceptionally doesn’t present systemic risks because of its specific characteristics, and therefore shouldn’t be classified as a systemic-risk model. If the Commission decides those arguments aren’t sufficiently substantiated, it rejects them, and the model is treated as a systemic-risk model regardless.
The appeal loop
The self-notification route comes with a built-in argument mechanism. The Commission’s own ex officio designation route, under Article 51(1)(b), doesn’t have an equivalent express provision letting a provider make its case before the decision is made — which pushes toward providers over-notifying rather than risk being designated without having had the chance to argue first.
Once a model has been designated this way, though, there is a path back. A provider can submit a reasoned request for reassessment, and the Commission will consider whether the model, measured against the Annex XIII criteria, is still properly considered a systemic-risk model. That request has to rest on objective, detailed, and new reasons that came to light after the designation decision — not a restatement of the original case — and the earliest a provider can make it is six months after the designation. If the Commission reassesses and decides to maintain the designation anyway, the provider has to wait another six months before trying again. There’s no cap on how many times this can repeat, only the six-month cooldown between each attempt.
Frequently asked questions
Is 10^25 a hard line?
No, in either direction. Above it, a provider can still argue its way out under Article 51(2) and 52(2)-(3) if it can substantiate that the model’s specific characteristics mean it doesn’t actually present systemic risk. Below it, the Commission can still designate a model under Article 51(1)(b) if it finds comparable capabilities or impact using the Annex XIII criteria. The number sets a presumption, not a verdict.
Who publishes the list of systemic-risk models?
The Commission does. Article 52(6) requires it to publish and keep an up-to-date list of general-purpose AI models with systemic risk, though that duty is itself qualified by the need to respect intellectual property rights and protect confidential business information and trade secrets under EU and national law.
Does the Omnibus change Articles 51-55?
No. This is one of the more consistently confirmed points across current commentary: the Digital Omnibus reform is squarely focused on the Annex III and Annex I high-risk timeline, registration, and related provisions. The general-purpose AI rules in Articles 51 through 55 have applied since 2 August 2025 and remain untouched by the reform.
What if we don’t know our own FLOP count?
The Act doesn’t offer a substitute self-certification route for this situation — the cost, time, and energy proxies in Annex XIII exist for the Commission’s own comparable-capabilities designation under Article 51(1)(b), not explicitly as an alternative counting method a provider can rely on for its own Article 51(2) self-assessment. A provider without confident compute accounting is sitting in exactly the gap this article describes, without a clearly signposted way out of it.
Can we be designated without ever notifying?
Yes. Article 52(1) states plainly that if the Commission becomes aware of a general-purpose AI model presenting systemic risks that it hasn’t been notified about, it can decide to designate that model as one anyway. Notification is the expected route in; it isn’t the only route the Commission can use to get there.