The European Union will, by 31 March 2027, have in force a binding regulation compelling all generative AI models trained with more than 10^25 FLOPs and placed on the internal market to release both their complete training datasets and model weights under an open-source license. This is not a speculative leap. It is the logical endpoint of legal text already on the books, political alignment among key member states, and the structural incentives of the European regulatory machine.

The legal machinery is already in motion

The 2024 AI Act requires high-risk AI systems to meet transparency obligations. The Act's draft implementing rules, circulated in early 2025, explicitly reference compute thresholds to distinguish frontier models from narrow applications. Article 52 and the corresponding annexes do not merely suggest transparency; they create mechanisms for the European Commission to specify technical documentation and data governance requirements through delegated acts. The path to prescriptive open-source mandates is already carved. A regulation requiring dataset and weight disclosure for models above a specific FLOPs ceiling is an exercise of that delegated authority, not a new legislative battle. The 10^25 FLOPs number is already appearing in standards discussions at CEN-CENELEC and in the Joint Research Centre's technical briefings. The regulatory infrastructure is being assembled.

France and Germany removed the political obstacle

For a prescriptive rule to survive, the bloc's two largest economies must not defect. Both have now signalled support for open-source carve-outs only when accompanied by full dataset disclosure. France's AI strategy, coordinated through the Secretariat General for Investment, has linked state funding for foundation models to data transparency. Germany's Federal Ministry for Digital and Transport backed similar conditions in the Council's working party on the AI Act. The political deal is pragmatic: Paris and Berlin want European champions, and they understand that mandating data openness simultaneously weakens non-European competitors who rely on proprietary data moats and creates a compliance advantage for European firms that can build audit-ready data pipelines. There is no blocking minority forming around a softer approach.

Market access forces compliance

Any developer, regardless of where training occurs, must comply with the regulation to sell or deploy their models in the EU market. Withdrawal is not a realistic option for frontier model companies whose enterprise customers demand EU availability. The EU's Digital Services Act and GDPR already proved that market access regulation can reshape global product design. This regulation extends that pattern into AI. The 10^25 FLOPs threshold captures every major general-purpose model released since GPT-4 and will capture all frontier releases through 2027. The choice for labs is binary: publish and stay in the market, or keep proprietary and lose the 450 million consumers and the enterprise contracts dependent on EU legality.

The enforcement architecture is credible

The AI Act created national competent authorities with the power to fine up to 7 percent of global annual turnover. The European AI Office is staffing enforcement teams. The precedent of GDPR enforcement, where initial slowness evolved into multi-billion-euro penalties, informs the strategy. Market surveillance will be triggered by the compute threshold, which is computationally verifiable through public cloud records and chip export filings. Non-compliance will be easier to prove than discrimination or bias cases. The incentives push toward proactive disclosure.

When this regulation takes effect, the default for frontier generative AI becomes open. Developers will preemptively structure training pipelines for auditability. The competitive battleground shifts from data hoarding to data curation quality. Every enterprise customer in the EU will have the right to inspect the datasets and weights behind the models they integrate. That is a structural change in the economics of AI development, and it arrives on a fixed date.

What is driving this

  • The 2024 EU AI Act already contains mandatory transparency requirements for high-risk systems and delegated acts can specify dataset and weight disclosure without new legislation.
  • France and Germany now support open-source mandates conditional on full dataset transparency, removing the threat of a blocking minority in the Council.
  • A compute threshold of 10^25 FLOPs cleanly separates frontier general-purpose models from smaller systems, making the rule enforceable via cloud and chip export monitoring.
  • The EU's demonstrated willingness to enforce market-access rules under GDPR and the Digital Services Act shows that global compliance is the rational path for developers.

What would prove this wrong

If a major European AI company such as Mistral or Aleph Alpha successfully lobbies for a blanket exemption for models partly funded by member-state governments, that would fragment the regulation and break the prescriptive open-source mandate.

The signal

The EU AI Act's 2024 text already mandates transparency for high-risk systems and the 2025 draft implementing rules explicitly reference compute thresholds; France and Germany have signaled support for open-source carve-outs only if accompanied by dataset disclosure.