On September 1 the European Commission sent formal questionnaires to more than thirty AI providers. It was the first time it has used its AI Act enforcement powers against frontier model developers. Somebody who supplies your stack got one.
Commission spokesman Thomas Regnier described two strands: one on safety and security of the most advanced models, one on copyright and transparency, the second aimed particularly at companies that skipped the AI Office’s informal compliance dialogues. The recipients have not been named. Henna Virkkunen, the Commission’s vice president for tech sovereignty, said the EU is “ready to take all necessary steps to ensure that companies comply with their obligations.”
The same day, in Chapel Hill, North Carolina, the United States was chairing a G20 ministerial and pitching the opposite.
Why September 1 and not eighteen months ago
The timing is the whole story and it is easy to miss.
General-purpose AI obligations have been law since August 2, 2025. The Commission’s power to actually enforce them against those providers only entered application on August 2, 2026. For a year the obligations existed with nothing behind them. Then the teeth arrived, and four weeks later the questionnaires went out.
Those powers are not advisory. The AI Office can demand information, demand model access for its own evaluations, require specific risk mitigations, fine up to 3% of global annual turnover, and order a model restricted, withdrawn or recalled from the EU market. An incorrect, incomplete or misleading answer to one of these questionnaires carries its own penalty of €15 million or 3% of global turnover, whichever is higher. A company can be fined for the reply before anyone rules on the underlying conduct.
There is also a plausible trigger. Reuters reported that OpenAI acknowledged in July that advanced models had “hacked on their own into a popular platform for programmers” during security testing, and that Anthropic disclosed its models had “gained unauthorised access” to three external organizations in testing. Both disclosures were voluntary. Both are now, in effect, Exhibit A. Anyone drafting their own transparency posture should sit with that for a moment.
The split is not the one you have been told about
The Carolina Principles, pushed by the US at the September 1 and 2 ministerial, commit signatories to reserve new regulation for genuinely novel considerations, fund foundational research, avoid standing up dedicated AI regulators, and collaborate on testing across public and private sectors. Michael Kratsios put it plainly: “Policymakers do not need to approach each innovation in isolation.” Musk and Zuckerberg were in the room arguing for fewer constraints. Kratsios also said the Chinese government had signed, though no signed document was produced and the Chinese embassy did not comment. Canada pushed a middle path. Formal adoption is slated for the leaders’ summit in Miami on December 14 and 15.
So the axis most people are carrying around — democratic AI safety norms on one side, Chinese state AI on the other — is not the axis. Washington and Beijing appear to be on the same page about not building new regulators. Brussels is alone, and knows it.
One more piece makes the EU position sharper than it looks from outside. Brussels is not simply regulating harder. Under the Digital Omnibus, the high-risk obligations were pushed back: standalone Annex III systems from August 2, 2026 to December 2, 2027, and embedded Annex I systems from 2027 to August 2028. The EU delayed the part that lands on European hospitals, manufacturers and HR departments, and is enforcing hard on the part that lands on a few dozen mostly American labs. Call that triage or call it self-interest. Either way it tells you where Brussels thinks its leverage actually is.
What I think
I wrote recently that a vendor’s usage policy is now a live term in your architecture. This is the same lesson arriving from the other direction, and it is the one that will cost people money. Your model provider’s regulatory position is not their problem. It is an input to your system that you did not specify, cannot negotiate, and will inherit whole.
Three things I would do this month.
Ask your provider directly whether they received one. Not whether they are “AI Act compliant” — every vendor will say yes to that. Ask whether they received a request for information from the AI Office on or around September 1, and what it covered. The answer is not confidential to you as a customer, and a vendor that will not answer it has told you something.
Price the withdrawal power, not the fine. The 3% headline gets the coverage, but a fine is your supplier’s balance-sheet problem. The power to order a model restricted or withdrawn from the EU market is yours. If a model your product depends on becomes unavailable to European users on a regulator’s timeline rather than your roadmap, what happens? If the honest answer is that you do not know, that is the same single-sourcing exposure I have written about before, wearing a different hat.
Do not let the Omnibus delay lull you. The high-risk deadlines moved. The Article 50 transparency obligations for AI-generated content did not. If you ship anything that generates or manipulates content for European users, your date did not change, and the eighteen months of breathing room you may have heard about belongs to somebody else’s compliance program.
Two things to watch between now and Miami in December. Whether any of the thirty-odd requests escalates into a formal investigation, which would tell you the Commission is willing to spend political capital on this rather than just paperwork. And whether the Carolina Principles pick up a single European signature. My expectation is one yes and one no, and I would rather be measured on that than hedge it.
The models you build on are becoming regulated products in one of the three markets that matter. Nobody is going to send you the questionnaire. You will find out from a changelog.
If you are working through an AI initiative and want a second technical opinion, or you want to talk through any of this, connect with me on LinkedIn or see the production side at coreagenticgroup.com.
