Twenty-six days from today, the European Commission's General-Purpose AI enforcement authority activates under the EU AI Act. The legal trigger is not a safety catastrophe. It is a paperwork failure.

Any frontier AI model provider that has not published the required model cards, training data summaries, copyright compliance statements, or compute threshold disclosures by August 2, 2026 becomes subject to fines of €15 million or 3 percent of global annual turnover — whichever is higher [1]. The coverage criterion is jurisdictional in effect rather than in letter: any provider whose model output reaches a European user through an API response, a chatbot reply, or a generated image is covered, regardless of where the company is incorporated or where its servers sit.

OpenAI, Anthropic, Google DeepMind, and Meta all meet that criterion. None has been granted a carve-out from the EU enforcement framework. None is incorporated in a European Union member state. All four are incorporated in the United States, and all four are covered.

The White House issued a voluntary 30-day preview requirement for US frontier AI companies on June 2, asking labs to share capability assessments and training disclosures with federal officials before public release. That requirement is voluntary; the EU's August 2 obligation is mandatory and backed by enforceable financial penalties [2]. Compliance with the White House preview framework does not satisfy any EU transparency obligation; the two regimes require separate filings, separate documentation, and separate compliance responses.

The GPAI transparency requirements apply at two tiers. Models trained with more than 10²³ floating-point operations must publish model cards, training data summaries, and copyright compliance statements — the standard documentation tier [1]. Models trained with more than 10²⁵ floating-point operations face additional systemic-risk obligations: adversarial testing, incident reporting, and notification requirements. GPT-4 class models and Claude 3 class models likely exceed the lower threshold; whether they exceed the higher one has not been publicly disclosed by any of the four companies [2].

The Geneva forum that closed in Geneva today — the UN's inaugural Global Dialogue on AI Governance — produced a co-chair summary with no enforcement mechanism. August 2 is the same calendar with a different legal architecture. Brussels does not require consensus among 193 governments. The Commission requires a compliance filing from every GPAI provider serving European users, and it has the authority to levy fines without convening a dialogue.

The enforcement gap between multilateral aspiration and unilateral binding law is concrete on today's date. The Geneva forum closes with no enforcement power; the EU countdown clock ticks with fines attached. Both are responses to the same problem. Only one of them has a consequence.

The US regulatory response to AI has, to date, operated through a different theory. Voluntary frameworks, executive orders requesting disclosure, industry self-certification, and FTC consumer-protection oversight are the operative instruments. None carries the financial penalty structure the EU activates in twenty-six days. US companies serving European users must therefore navigate a two-layer compliance structure: the voluntary domestic framework they built their internal processes around, and the mandatory European framework that applies to them regardless of where they operate [3].

Compliance lawyers across Silicon Valley have been advising clients on this gap since the EU AI Act's GPAI provisions were finalized in early 2024. The question that August 2 will answer is whether any major US frontier lab has actually filed the required documentation, or whether the industry has collectively assumed that the Commission's enforcement will begin with grace periods, warning notices, and informal guidance rather than immediate fines.

The Commission has not publicly committed to any enforcement grace period for GPAI providers. It has not published a list of compliant providers. It has not issued pre-enforcement safe-harbor interpretations. What exists is the August 2 date and the penalty structure attached to it.

No major US AI lab has publicly confirmed it has filed GPAI transparency documentation in anticipation of August 2. The Geneva forum produced a co-chair summary. August 2 produces a fine.

-- ANNA WEBER, Berlin