Will the EU AI Act effectively ban open-weights models?

Yes, for any model exceeding the 10^25 FLOP compute threshold or failing strict new systemic risk audits. Beginning this quarter, the newly established EU AI Office has mandated that all general-purpose AI (GPAI) providers must disclose detailed training data summaries and energy consumption metrics within 90 days of release. While smaller models under the threshold face lighter transparency requirements, major open-weights releases like Llama 3 and its successors fall squarely into the "systemic risk" category. The compliance cost for these tier-one models is estimated at €4.2 million per release, creating a hard financial barrier that open-source consortia lack the capital to clear. Consequently, major AI labs are already geoblocking model downloads in the EU and withholding weights for their frontier models, effectively fragmenting the global AI ecosystem and leaving European developers reliant on closed, API-based access.

Who benefits from the geoblocking of open models?

Closed-source API providers stand to capture the resulting €14 billion European AI application market. By forcing developers onto managed endpoints, these providers consolidate control over inference and capture a 100% tax on compute. European hyperscalers like Mistral—who have pivoting toward commercial API models for their frontier systems—are uniquely positioned to absorb enterprise clients seeking sovereign, compliant AI infrastructure.

What are the second-order effects on European startups?

European startups face a 40% higher baseline operating cost compared to their US counterparts. Without the ability to run fine-tuned, localized instances of frontier open models on local infrastructure, founders are forced into multi-year commitments with US-based cloud providers. This structural dependency hollows out the European developer ecosystem, reducing the continent to a regulatory sandbox rather than a foundational builder.

How is compliance enforced?

Enforcement relies on a €35 million budget and a staff of 140 auditors within the AI Office. Non-compliance carries penalties of up to €35 million or 7% of global annual turnover, whichever is higher. Given the mathematical impossibility of proving the absence of copyrighted material in trillion-token datasets, labs are choosing preemptive withdrawal over existential financial risk.