Skip to content
Briefings are running a touch slower this week while we rebuild the foundations.See roadmap
European Tech Sovereignty
13APR

Mistral CEO proposes EU-wide AI levy

2 min read
17:09UTC

Arthur Mensch wants every AI provider operating in Europe to pay 1 to 1.5% of revenue in exchange for legal certainty on training data.

TechnologyDeveloping
Key takeaway

Mistral's proposed AI levy would trade copyright liability protection for a revenue mechanism benefiting European creators.

Mistral CEO Arthur Mensch published an FT opinion piece in March 2026 proposing a 1 to 1.5% revenue levy on all AI providers operating in the EU market 1. The levy would apply equally to foreign and domestic companies, funding European cultural creators. In exchange, AI developers would receive legal certainty on training data use, shielding them from copyright liability for training on publicly accessible content.

Mensch framed the proposal as a level playing field mechanism, and the framing is calculated. A flat revenue levy hurts US incumbents with large European revenue bases (OpenAI, Google, Anthropic) more in absolute terms than it hurts Mistral, whose European revenue is smaller. The copyright certainty benefit, conversely, matters more to Mistral: the company trains on European-language data and faces greater legal exposure to EU copyright holders than US competitors whose primary training corpora are English-language.

Whether the proposal gains traction in Brussels depends on the Commission's appetite for a new revenue instrument during a period of transatlantic trade tension. The US administration has already filed a Section 301 investigation naming DMA rules as economic warfare. Adding an AI-specific levy would sharpen that confrontation. The cultural funding element is designed to win support from France's powerful creative industries lobby, which has resisted AI training on copyrighted works without compensation.

Deep Analysis

In plain English

Training an AI model requires feeding it enormous amounts of text; news articles, books, websites, and other written material. The question of who owns that material, and whether AI companies need permission to use it, is one of the biggest legal disputes in the tech industry right now. Mistral's CEO Arthur Mensch wrote an opinion piece in the Financial Times in March 2026 proposing a deal: AI companies would pay a levy of 1 to 1.5% of their revenue into a fund that compensates the authors, journalists, and publishers whose work was used for training. In return, AI companies would receive legal certainty; they would not face copyright lawsuits for using publicly accessible content. The proposal is notable because Mensch framed it as applying equally to foreign and European AI companies. In practice, because US companies like OpenAI earn far more revenue, they would pay far more into the fund; a structural advantage for smaller European competitors.

What could happen next?
  • Consequence

    If adopted, the levy redistributes AI training data costs in proportion to revenue, systematically advantaging smaller European AI companies over US revenue leaders.

    Long term · 0.6
  • Risk

    The proposal may stimulate a US trade complaint under the USTR's ongoing Section 301 review of EU digital policy, framing the levy as discriminatory tech taxation.

    Medium term · 0.55
  • Opportunity

    A statutory training data licence creates a mechanism for European publishers and news organisations to monetise AI training use, addressing a significant revenue gap in the media industry.

    Medium term · 0.65
First Reported In

Update #1 · Europe's chip ambitions meet reality

White House· 13 Apr 2026
Read original
Different Perspectives
ASML / European tech industry
ASML / European tech industry
ASML's Q2 2026 guidance came in €300m below consensus as China DUV revenue collapsed 17 percentage points; the company's CEO wrote US export-control outcomes directly into 2026 guidance. European tech firms named on the USTR retaliation list alongside SAP, Siemens and Spotify face the same calculus: US trade exposure constrains what Brussels can legislate on their behalf.
France / Anne Le Henanff
France / Anne Le Henanff
Le Henanff chaired the G7 Digital Ministerial at Bercy on 29 May with CAIDA off the agenda, pivoting France's presidency to AI safety principles it had not designed the week around. France backs CAIDA but cannot override Berlin's tariff calculus, so the ministerial produced no new French-led commitment.
Germany / Federal government
Germany / Federal government
Berlin's automotive sector faces up to $200bn in threatened US tariffs, a commercial exposure that dwarfs any benefit CAIDA's public-sector cloud rules would deliver to German digital firms. Federal silence inside the College of Commissioners functions as a block under consensus adoption rules without requiring a formal veto.
USTR / Ambassador Andrew Puzder
USTR / Ambassador Andrew Puzder
Puzder's public warning on 25 May that CAIDA is inconsistent with the EU-US trade framework was the first time Washington made its bilateral pressure visible before a Commission adoption vote rather than after. The USTR Section 301 determination on 24 July provides the enforcement backstop.
European Commission / Henna Virkkunen
European Commission / Henna Virkkunen
Virkkunen framed the third slip as a procedural delay in finalising a 400-page text without addressing Puzder's trade-framework red line publicly. The Commission enforces existing law against Google while losing the legislative timeline on CAIDA, exposing an asymmetric position: enforcement holds; new sovereignty legislation does not.
OpenForum Europe / open-source community
OpenForum Europe / open-source community
The EUR 350m Sovereign Tech Fund has no Commission host, no budget line, and no commissioner's name attached six weeks after the April conference, while Germany is already paying maintainers to staff international standards bodies. The CRA open-source guidance resolves contributor liability but leaves the financial-donations grey area open with the 11 September reporting clock running.