Summary
- Article 51(2) of Regulation (EU) 2024/1689 presumes high-impact capabilities where the cumulative computation used to train a general-purpose AI model is greater than 10^25 floating-point operations — a strict inequality on a quantity the provider itself measures, not a risk finding reached by the regulator.
- The provider must notify within two weeks, may argue exceptionally under Article 52(2) that a threshold-meeting model does not present systemic risks, and the Commission can amend the threshold by delegated act; the only public artefact is the Article 52(6) list, published subject to intellectual-property rights, confidential business information and trade secrets.
Every regulatory category has a moment of entry, and this one is unusually literal. Article 51(2) of Regulation (EU) 2024/1689 does not ask an official to form a view about danger. It asks whether a quantity exceeds a number: whether the cumulative computation used to train a general-purpose AI model, measured in floating-point operations, is greater than 10^25 (consolidated text, 27 July 2026). The authentic Official Journal edition prints the operative phrase in exactly that shape — "greater than 10^25." (Official Journal PDF). The comparison is strict, so a model trained at precisely the threshold is not caught by it.
One rendering detail is worth pinning down, because the whole argument sits on it. The consolidated HTML drops the superscript and shows the exponent as a flat "10 25"; the notation used here follows the authentic Official Journal presentation, where the exponent is visible. That is a typesetting difference between two official renderings of the same clause, not a difference in the clause.
The unit matters more than the exponent. Floating-point operations used cumulatively for training is a quantity of computation consumed over a training run. It is not a rate, such as floating-point operations per second; it is not electrical power, memory, parameter count or money. Those are different quantities, measured by different instruments, and none of them converts into the statutory unit by arithmetic. A briefing that treats 10^25 FLOP as a power bill or a clock speed has moved to a different subject.
Two doors into one category
Article 51(1) sets out two routes in. A model is classified as presenting systemic risk either because it has high-impact capabilities "evaluated using appropriate technical tools and methodologies, including indicators and benchmarks", or because the Commission decides, on its own initiative or after a qualified alert from the scientific panel, that it has equivalent capabilities or impact having regard to Annex XIII. The compute threshold is not a third category. It is the presumption that feeds the first route, which is why the statutory sentence is about presumed capabilities rather than declared danger.
The second door is the one that survives every arithmetic change. Designation under Article 52(4) rests on Annex XIII criteria, and the Commission holds a separate delegated power to specify and update those criteria. A model below the line is therefore not outside the regime; it is outside the presumption. The text of Annex XIII and the Code of Practice were not read for this piece, so nothing here turns on their contents.
Fourteen days, and what silence costs
Article 52(1) puts the clock on the provider: it "shall notify the Commission without delay and in any event within two weeks after" the condition is met, or after it becomes known that the condition will be met. The trigger is the provider's own knowledge, and it reaches forward to knowledge that the bar is about to be cleared. My conversion, not the Regulation's words: two weeks is fourteen calendar days.
The notification has to carry the information necessary to demonstrate that the requirement has been met. And the same paragraph closes an obvious exit: if the Commission becomes aware of a general-purpose AI model presenting systemic risks that the provider has not notified, it may designate the model anyway. Silence does not preserve the status quo; it transfers the decision.
The rebuttal is where the discretion sits
Article 52(2) allows the provider, with the notification, to present sufficiently substantiated arguments that exceptionally — although the requirement is met — the model does not present systemic risks because of its specific characteristics and should not be classified as one with systemic risk. Article 52(3) then states the consequence of failure: where the Commission finds the arguments insufficiently substantiated and the provider has not demonstrated the absence of systemic risks, it shall reject them, and the model shall be considered a model with systemic risk.
Read those two paragraphs together and the shape of the instrument appears. The threshold decides who must file. The Commission decides who is excused. The Commission's own guidance adds a detail providers will feel: where the arguments are rejected, the additional obligations attach from the moment the model met the threshold, not from the date of the decision (Commission guidance FAQ, 11 November 2025). The same page states that the guidelines are not legally binding and that authoritative interpretation belongs to the Court of Justice of the European Union.
Article 55(1)(a) is what those additional obligations look like in practice for a classified model: model evaluation using standardised protocols and state-of-the-art tools, including conducting and documenting adversarial testing to identify and mitigate systemic risks. Article 53(1)(a) requires the underlying technical documentation, containing at least the Annex XI information, "so that it can be provided upon request to the AI Office and national competent authorities". Requests, not publication. That distinction runs through the whole instrument.
Who can move the number
Article 51(3) is the provision that can change the threshold, and it is worth being exact about which article does what. Article 51(3) directs the Commission to adopt delegated acts in accordance with Article 97 to amend the thresholds in Article 51(1) and 51(2) and to supplement benchmarks and indicators, where necessary for them to reflect the state of the art. Article 53(5) and 53(6) do not carry that power; they concern measurement methodologies and the annexes on documentation.
The procedure is not a formality. The delegations covering Article 51(3) and Article 52(4) run for five years from 1 August 2024 and are tacitly extended for identical periods unless the European Parliament or the Council opposes the extension no later than three months before the end of the relevant period (Article 97(2)). An individual delegated act enters into force only if neither institution objects within three months, and that period can be extended by three months at the initiative of either (Article 97(6)). A published amending act, Regulation (EU) 2026/1744 of 24 July 2026, rewrote that machinery, including the five-year wording and the three-month objection rule.
My reading of the two periods, labelled as arithmetic: three months plus three months is a maximum of six months before an unopposed act can enter into force when either institution extends.
What the public can read
Article 52(6) does require publication, and its first sentence is short: "The Commission shall ensure that a list of general-purpose AI models with systemic risk is published and shall keep that list up to date." The sentence does not end there. The duty is subject to protection of intellectual-property rights, confidential business information and trade secrets.
The vocabulary is precise. The statute calls this a list; the register and the EU database it creates elsewhere belong to high-risk and other AI systems under Articles 49 and 71. Article 52 does not require the notification itself, an acknowledgment, a receipt or a tracking number to be published. Whether any model currently appears on the Article 52(6) list was not established in this review; that is a limit of what was read, not a finding that the list is empty.
What the sender gets is private by design. Submissions run through the EU SEND platform, which the Commission describes as ensuring the confidentiality, integrity and authenticity of the information shared, and lists among the documents to be transmitted the systemic-risk notifications, reassessment requests and serious-incident reports (Commission implementation page, 28 April 2026). The Service Desk guidance is more concrete: every message sent through the platform returns a digitally signed receipt of transmission to the sender, and information received from providers is subject to confidentiality obligations (AI Act Service Desk). A receipt exists. It just is not addressed to you.
The same page carries a small drafting slip worth noticing: it refers to notifications under Article 51(2) and Article 52(2), while the operative duty sits in Article 52(1). The Regulation is the controlling text.
Reach, and two dates
Article 2(1)(a) applies the Regulation to providers placing general-purpose AI models on the Union market whether they are established in the Union or in a third country, and Article 3 defines a provider by reference to placing a model on the market under its own name or trademark (Official Journal text). The compute threshold therefore does not depend on where the training run happened.
Article 113 splits the calendar. The Regulation applies generally from 2 August 2026, while Chapter V, which contains Articles 51 to 55, applies from 2 August 2025, except Article 101. The Commission's implementation page reports the same split and adds that its enforcement powers entered into application on 2 August 2026. The rules on general-purpose models and the power to sanction against them arrived a year apart.
A live number with a soft edge
A Commission fact page on general-purpose AI obligations, last updated 1 August 2025, states that models trained with more than 10^25 FLOP are presumed to pose systemic risk — and then prints a sentence that has never quite gone away: "This threshold is currently under review" (Commission fact page).
More recent pages still use the number. The Commission's "Navigating the AI Act" FAQ, last updated 20 October 2026, states that models trained using more than 10^25 FLOP are currently presumed to pose systemic risks, that the Commission may update the threshold in light of technological developments, that it may rebut the presumption at a provider's request, and that it may designate models below the threshold using Annex XIII criteria (Navigating the AI Act). The consolidated text of 27 July 2026 still displays greater than 10^25, and the EUR-Lex consolidation page identifies that version as 02024R1689 — EN — 27.07.2026 — 001.001 while warning that a consolidation is a documentation tool without legal effect, the authentic texts being those published in the Official Journal.
No delegated act amending the Article 51(2) threshold was obtained for this piece. The review statement is therefore dated 1 August 2025, the operative number is dated 27 July 2026, and the gap between them is the most honest thing in the file.
Why the perimeter is the story
The European AI Office, established inside the Commission, enforces the rules for general-purpose AI models and develops the tools, methodologies and benchmarks used to evaluate capabilities and classify models with systemic risk (European AI Office). It is the body that reads the notification and weighs the arguments. It is not the body that publishes them.
Put the pieces in order and the instrument is coherent, and its coherence is the point. A legislature set a number. A developer measures it. A two-week clock starts on the developer's own knowledge. The Commission alone decides whether the exception applies, and duties reach back to the moment the number was crossed if the exception fails. The Commission can move the number by delegated act that a Parliament and Council can stop inside a fixed window. And the one public object is a list that intellectual-property rights, confidential business information and trade secrets are expressly allowed to filter.
Anyone can check whether a model is on the list. Nobody outside the reporting relationship can check why it is on the list, because the evidence is a document the Commission requests rather than one the provider files. That is not a scandal and it is not a loophole. It is a design choice, printed in the article that governs the list, and it decides which questions about European AI governance are answerable from the outside.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






