T-minus One Hundred Days — What the 2 August AI Act Switch Actually Changes

The countdown

On 2 August 2026, the European Commission's enforcement powers over general-purpose AI model providers come into force. That is one hundred days from today. The fines are capped at €15 million or 3% of global annual turnover, whichever is higher. The AI Office has been consulting informally with providers since August 2025. Starting August 2026, the consultation ends and the enforcement begins.

Most of the European AI conversation this spring has treated 2 August as the moment the AI Act becomes "real." It is not. The AI Act has been real since its entry into force in 2024. What becomes real on 2 August is a specific administrative power: the Commission can now open a proceeding, issue a decision, and collect a fine against a GPAI provider. That is a meaningful addition. It is also a very narrow one.

The mistake — visible in almost every ministry briefing I have read this month — is to treat enforcement as if it were governance. It is not. Enforcement is the moment at which paperwork becomes consequential. Governance is the question of whether the paperwork, once consequential, tells you anything worth knowing.

What the switch actually switches

Three things change on 2 August, and only three.

First, the Commission acquires the power to compel disclosure from GPAI providers — training-data summaries, compute footprints, systemic-risk assessments — under Article 92 information requests. Since August 2025 these disclosures have been "expected." Starting August 2026 they are demandable.

Second, the Commission acquires the power to designate a GPAI model as posing systemic risk, triggering the heavier obligations of Chapter V Section 3 — model evaluations, adversarial testing, cybersecurity protection, incident reporting. The threshold remains 10²⁵ FLOPs of training compute, a figure already exceeded by every frontier model shipped since GPT-4.

Third, the Commission acquires the power to fine. The ceiling is €15 million or 3% of global turnover. For an American frontier lab that is roughly a single week of revenue. For a European foundation-model startup, it is a termination event.

That is the switch. It is not a regulatory regime change. It is a procedural floor.

What the switch does not switch

What 2 August does not do is equally worth naming.

It does not require any GPAI provider to open model weights, training data, or evaluation harnesses to independent European audit. It does not give a European ministry that has deployed a GPAI — in a chatbot, a translation pipeline, a legal-review tool — any right to rebuild or fork that model if the provider changes its terms. It does not stop a GPAI provider from updating a deployed model silently, changing its behaviour inside a public service, and disclaiming the change under "continuous improvement."

None of these are oversights. They are deliberate choices from the Act's 2024 text. They reflect the political bargain that made the Act passable: GPAI providers accepted a documentation regime in exchange for not accepting a structural one.

The result is a regulation that polices how providers describe themselves without policing what they can do.

Three tests the enforcement switch leaves unanswered

Every ministry that has deployed a GPAI in a live service should apply three tests that the Act does not require and the Commission cannot compel.

Provenance. Can you produce, on demand, a provenance chain for every output the deployed model has generated in production — prompt, context, model version, training snapshot? The AI Act requires documentation at the model layer. Public accountability requires documentation at the deployment layer. These are different artefacts and the law conflates them.

Portability. If the GPAI provider doubles its price, withdraws from Europe, or is subject to a CLOUD Act order tomorrow, can you run the same service on an EU-hosted open-weight model in under ninety days? The answer is almost never yes. The architectural work to make it yes is not a compliance exercise. It is an engineering decision made eighteen months before you need it.

Pause. Can you disable the GPAI without disabling the public service wrapped around it? Most current deployments fail this test quietly. The model is not a component; it is the service. A pause switch without a graceful-degradation path is theatrical. It lets a minister say the system can be turned off. It does not let the system continue to serve the public if the minister does.

A procurement that passes these three tests passes the AI Act automatically. A procurement that passes only the AI Act does not pass these three.

The 2 August posture

Ministries and regulated deployers in Europe should arrive at 2 August with a posture, not a file. The file — the Article 92 response, the systemic-risk summary, the incident-response plan — is table stakes. Every counterpart will have one. The posture is the part that distinguishes capable deployment from compliant deployment: knowing, on a named person's desk, who answers when the provider changes behaviour, when the provider withdraws, when the provider is compelled by a foreign jurisdiction to do something the European deployer cannot lawfully permit.

That desk exists in exactly one place today, and that is inside the deploying institution. The AI Office cannot occupy it. A fine cannot replace it.

Closing

The enforcement switch is a useful addition to a regime that, without it, had no teeth. It is also the thinnest possible interpretation of what technology governance looks like. One hundred days is enough time to finalise a file. It is not enough time to build a posture. Anyone who intends to have one on 2 August started building it six months ago.

If you have not started, start now. Not because 2 August requires it. Because August 3rd will, and nobody will be left to remind you.

Read the companion essay on why sovereignty has to mean more than residency, or get in touch to discuss applying these tests to a specific deployment.