On September 8, 2026, Mistral AI announced a €3 billion Series D at a post-money valuation above €21 billion. Mistral's own announcement calls it the largest equity round ever raised by a European technology company. Samsung Electronics led, with the EQT-managed Scaleup Europe Fund and PSG Equity as co-leads. The company says the money will fund frontier research, commercial expansion and its goal of 1 GW of compute by 2030.
TechCrunch reports that President Macron framed the round as part of building a "third way in AI", meaning sovereignty and less dependence on American technology. It also reports that Mistral now hosts open-weight models from several sources, including Chinese ones. That combination, a heavily capitalised open-weight developer with a state-backed sovereignty mission, is what the EU AI Act's open-source provisions were never quite designed for.
The case for strict rules
The strongest argument for treating open-weight models with suspicion is that release is irreversible. A closed model can be patched, rate-limited or withdrawn. Published weights cannot be recalled, and safeguards can be fine-tuned away. Regulators who take that seriously are not being paranoid. They are reasoning about a real asymmetry, and it is why the AI Act does not give open models a blanket pass.
What the Act actually does
Under Regulation (EU) 2024/1689, Article 53(2) relieves providers of a model released under a free and open-source licence, with weights, architecture and usage information public, from two duties. Those are the technical documentation for authorities and the information package for downstream providers, per the text of Article 53. The exemption is narrow. According to the European Commission's general-purpose AI Q&A, open-source providers still owe a copyright-compliance policy and a public summary of training content.
The exemption also does not apply to models with systemic risk. The presumption attaches at 10^25 FLOP of training compute, and the Commission can designate models below that line. The same Q&A says general obligations began on August 2, 2025 and AI Office enforcement begins in August 2026. Fines can reach 3% of global annual turnover or €15 million, whichever is higher.
That timing matters. Enforcement powers went live only weeks before Mistral raised €3 billion to build frontier open-weight models. As its compute grows toward 1 GW, its future flagship models will likely be judged under the systemic-risk track, where the open-source relief does not apply.
Why proportionality matters here
This is not an argument against the Act. Systemic-risk duties such as evaluations, incident reporting and cybersecurity are reasonable for the very largest models, and a compute threshold is at least a legible trigger. The risk lies in how the rules are applied, and there are three points to watch.
- Open-weight deserves recognition as a transparency benefit. Published weights let independent researchers audit, red-team and benchmark a model in ways closed APIs forbid. Regulators should count that openness toward the risk assessment rather than treat it purely as a hazard.
- Compute is a proxy, not a measure of harm. The 10^25 FLOP line is a rebuttable presumption. The Commission's own Q&A notes it can be adjusted by delegated act as technology evolves. Efficiency gains mean capability per FLOP keeps rising, so the AI Office should tie designation to demonstrated capability rather than headline compute.
- Hosting is not authorship. Mistral hosts third-party and Chinese open-weight models. The Commission's guidance treats a downstream party as a new provider only when it substantially modifies a model. It gives an indicative threshold of more than one-third of the original training compute. Hosting a model unmodified should not turn an infrastructure provider into its regulator-of-record. Authorities should say so plainly, so that European hosts are not disadvantaged against US clouds.
The sovereignty tension
Paris wants sovereign AI procurement and open weights at once, and the two can pull apart. Sovereignty rhetoric invites public buyers to favour a national champion. Open weights work best when many actors can adopt, fork and audit a model. If French procurement rewards Mistral because it is French, and the AI Act then imposes its heaviest duties on the same firm as it scales, Europe will have paid for a frontier lab and taxed it for succeeding.
The better course is to make compliance cheap and predictable. That means published evaluation protocols, a fast designation process, and clear guidance that unmodified hosting does not create provider duties. It also means keeping the open-source relief intact for the many smaller developers below the systemic-risk line. A €21 billion valuation shows that investors, Samsung and BlackRock among them, believe the open-weight model can scale commercially. Europe's job is to keep the rulebook from being the thing that decides where that scaling happens.
What to watch
The first real test will be whether the AI Office designates any open-weight model as systemic-risk, and on what evidence. If designation rests on capability evaluations and published methods, Europe will have built a credible template. If it rests on opaque judgment calls, developers will conclude that building open and building big are incompatible in the EU, and they will build elsewhere.