GPAI: Buying a Foundation Model Means Inheriting Article 51's Obligations

The EU AI Act doesn't only address use cases: it also addresses general-purpose AI models (GPAI). Articles 51 to 55 of Regulation (EU) 2024/1689 set out a specific regime, in force since August 2, 2025. And it carries a threshold every tech acquirer should know exists.

The 10²⁵ FLOPs threshold

Article 51 presumes that a GPAI model presents a systemic risk when the cumulative compute used for its training exceeds 10²⁵ floating-point operations (FLOPs). Above that, the model moves into a reinforced regime. Below it, baseline obligations already apply.

In concrete terms, a GPAI model must, among other things:

• maintain technical documentation and make it available (Art. 53);
• publish a copyright compliance policy and a summary of training data;
• for systemic-risk models: adversarial evaluation, tracking and serious-incident reporting, cybersecurity (Art. 55).

Why it weighs on value

An acquirer who buys a company developing a GPAI inherits these obligations — and the associated exposure. The question isn't incidental: the copyright policy and the training-data summary go straight to the most active litigation in the sector. A model without Art. 53 documentation, or without Art. 55 adversarial evaluation, is an asset part of whose value depends on a not-yet-completed remediation.

The D7™ score addresses this dimension in its analysis of technical maturity and model risk (D2), articulated with regulatory exposure (D1). It indicates whether the GPAI asset crosses into the systemic regime, and where the Art. 51-55 gaps lie — before a buyer or the AI Office does.