GPAI Systemic Risk: The 10^25 FLOP Threshold and What It Triggers
Training compute above 10^25 FLOP triggers a presumption of systemic risk — and with it, model evaluation, adversarial testing, incident reporting and cybersecurity duties. Here is how the threshold works and what to do when you cross it.
General-purpose AI models are the foundation layer of the EU AI Act’s entire regulatory pyramid, and the Act draws one hard line through that layer: models whose accumulated training compute exceeds 10^25 floating-point operations are presumed to pose systemic risk. Crossing that line changes your compliance profile overnight — heavier evaluation duties, adversarial testing, serious-incident reporting directly to the AI Office, and cybersecurity obligations on top of the baseline GPAI duties everyone carries.
Whether you are building frontier models or buying them, you need to know exactly where the threshold sits, how designation works, and what Article 55 adds to Article 53.
What “systemic risk” means in the AI Act
The Act uses the term deliberately: a GPAI model has systemic risk when its capabilities or scale of deployment could threaten public health, safety, fundamental rights, or the stability of the economy in ways that go beyond the impact of any single application built on top of it. A buggy chatbot is a product defect; a frontier model that half of European industry embeds is infrastructure.
The Act’s answer is proportionate tiering:
- All GPAI models carry baseline provider duties under Article 53 — documentation, downstream transparency, a copyright policy, and a public training-content summary.
- GPAI models with systemic risk carry everything above plus the Article 55 duties — model evaluation with adversarial testing, systemic-risk assessment and mitigation, serious-incident tracking and reporting to the AI Office, and cybersecurity protections for the model itself.
The category matters commercially as well as legally. When a model is designated as systemic risk, downstream providers integrating it inherit a heavier evidence burden — which means your customers’ questionnaires will start asking for proof of Article 55 work.
The 10^25 FLOP presumption (Article 51)
Article 51 gives the threshold its legal force. A GPAI model is presumed to have systemic risk when the cumulative amount of computation used for its training crosses 10^25 FLOP. A few practical notes:
- It is a presumption, not a definition. The Commission may also designate a model as systemic-risk on other grounds — notably when model capabilities, the number of modalities it handles, or its integration into critical services suggest comparable risk even below the threshold. Treat 10^25 as the floor of the conversation, not a safe harbour.
- Compute is measured cumulatively. Training runs, large-scale fine-tunes, and continued pre-training all add to the total. Teams that track compute per run but not per model are measuring the wrong thing.
- There is no public registry of “safe” numbers. Providers self-assess against the criterion and notify; the Commission reviews designation decisions.
| Scenario | Threshold status | What it triggers |
|---|---|---|
| In-house model trained at 10^23 FLOP | Below presumption | Art. 53 baseline GPAI duties (if placed on the market) |
| Frontier model trained at >10^25 FLOP | Presumed systemic risk | Art. 53 + Art. 55, Commission notification |
| Model below threshold but with exceptional capabilities | Designatable | Commission may designate under Art. 51(1) criteria |
| Fine-tune of an existing model | Depends on cumulative compute and role | Recompute the total; check whether you became a provider |
Designation and notification (Article 52)
The procedure is provider-driven. When your model meets the Article 51 criteria — including crossing the compute threshold — you must notify the Commission without delay and, in any event, within two weeks of meeting them. In practice:
- Measure the cumulative training compute across the model’s full training history, not just the final run.
- Document the calculation — the date crossed, the jobs counted, the assumptions. This file is what you will hand an auditor or the AI Office.
- File the notification with the Commission (the AI Office handles the technical side), including the model name, provider identity, and the date the threshold was met.
- Continue to market — notification does not pause deployment, but it starts the clock on Article 55 obligations.
Failing to notify is itself an infringement: GPAI provider violations fall under Article 101, with fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher.
What changes at systemic-risk tier (Article 55)
Article 55 layers four obligations onto the Article 53 baseline:
- Model evaluation with adversarial testing. Not a one-off benchmark: a documented, repeatable programme that probes the model for misuse vectors — jailbreaks, generation of prohibited content, evasion of filters — and records findings.
- Systemic-risk assessment and mitigation. Identify the specific systemic risks the model creates, evaluate their likelihood and impact, and implement proportionate mitigation (access controls, capability restrictions, staged release).
- Serious-incident tracking and reporting. Serious incidents and near-misses must be tracked and reported to the AI Office through the designated mechanism. This is the same “serious incident” concept used elsewhere in the Act, and it requires the logging backbone described in Article 12 — you cannot report what you did not record.
- Cybersecurity of the model and its physical infrastructure. Protection of model weights, training environment, and serving infrastructure against unauthorised access or theft. Stolen frontier weights are a systemic-risk event in themselves.
Each duty produces artefacts: evaluation reports, mitigation plans, incident logs, security assessments. Mid-market teams rarely build these models, but they must be able to ask for them — that asymmetry is where procurement questions come from.
GPAI baseline duties still apply (Article 53)
Systemic-risk status adds to Article 53; it never replaces it. Every GPAI provider must still maintain:
- Technical documentation for the AI Office and downstream providers, including architecture, training data summary, and evaluation results.
- Information to downstream providers, sufficient for them to integrate the model lawfully and to meet their own obligations — a legitimate reason your AI vendor asks for detail that feels excessive.
- A copyright compliance policy respecting EU law, including machine-readable rights reservations under the text-and-data-mining regime.
- A sufficiently detailed public summary of the training content, so the market can see what went in.
Downstream users — the mid-market companies fine-tuning or reselling models — inherit the duty not to circumvent opt-outs encoded in that training data, and they must use the model within the information limits the provider gave them.
The Code of Practice as your compliance route
The Commission’s GPAI Code of Practice is the pragmatic compliance instrument for Articles 53 and 55: sign it, and you have a concrete, recognised way to demonstrate adherence; refuse it, and you must demonstrate an equivalent (and arguably harder-to-prove) route to the same outcome. For model providers, signing is usually the cheaper path. For downstream buyers, “our provider is a Code signatory with systemic-risk measures documented” is the fastest answer to a due-diligence questionnaire.
Models already placed on the market before the GPAI obligations bit benefit from a transitional window under Article 111: compliance for pre-existing models is required by 2 August 2027, which is why 2026 still feels early — until you look at how long a real evaluation programme takes to stand up.
What this means if you buy or fine-tune models
Mid-market companies sit downstream of all of this, and three questions decide your position:
- Do we fine-tune in a way that changes our role? Substantial modification or putting your own name on a model can make you the provider of that model under Article 25 — with the compute question re-evaluated on the modified system.
- Is our vendor’s model designated? If yes, demand evidence of Art. 55 evaluation, mitigation, and incident reporting; it flows straight into your own vendor questionnaire responses.
- Can we rely on the provider’s outputs? Article 53 information duties are designed so downstream integrators can comply without re-running the provider’s evaluations — but only if you actually collect and store what they hand over.
How to check your own position in three minutes
The free exposure scan asks twelve questions about your AI usage and governance and returns an exposure score, your ranked exposures with the exact article that fired, and your first three moves — mapped to the dates that apply. Confirm your email and the full report unlocks.
Frequently asked questions
Is 10^25 FLOP the only way a model becomes systemic risk?
No. It is a rebuttable presumption: crossing 10^25 FLOP of cumulative training compute puts the model in scope automatically, but the Commission may designate models below the threshold when their capabilities, modalities, or integration into critical services suggest comparable systemic risk. The safe working assumption is that the threshold is the floor of the designation conversation, not a ceiling that protects anything under it.
Who has to notify the Commission when the threshold is crossed?
The provider of the GPAI model — the entity that develops it and places it on the EU market under its own name or trademark. Deployers and downstream fine-tuners do not file the notification for someone else’s model, but they should verify that their provider has filed it when the model is designated, because their own due-diligence answers depend on it.
Does fine-tuning push us over the threshold?
Cumulative training compute counts every substantial training stage a model has been through, so a large fine-tune or continued pre-training run contributes to the total. If your post-fine-tune model crosses 10^25 FLOP — or if by fine-tuning and marketing the result you become the provider under Article 25 — the notification duty lands on you. Record compute for every training job, including fine-tunes, from day one.
What are the penalties for missing the notification?
GPAI obligations, including systemic-risk duties, are enforced under Article 101: fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. Beyond the fine, an unnotified designated model invites information requests from the Commission and reputational damage at exactly the moment enterprise buyers start asking for your Code of Practice signatory status.
We put our model on the market before August 2025 — when do we comply?
Article 111 provides a transitional regime: GPAI models placed on the market before the obligations applied must comply by 2 August 2027. That is not a grace period from the market’s perspective — large buyers are already asking for evidence of evaluation and mitigation practices now, so waiting for the legal deadline puts commercial contracts ahead of regulatory ones.
Does systemic-risk status change anything for our downstream customers?
It changes what they can expect from you: enough technical information to integrate safely, evidence of adversarial evaluation and mitigation, incident awareness, and cybersecurity assurances for the model layer. Downstream providers building on a designated model should request these artefacts contractually — they will need them for their own AI Act answers and for their customer questionnaires.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min