Public funding of AI in France: what state money has actually produced

Since 2018, the state has committed approximately 2.4 billion euros to its national AI strategy. The Court of Auditors praises the growth of research and start-ups, but points to overly slow diffusion and inadequate monitoring.

A strategy launched in 2018, in three phases

France did not convert to artificial intelligence in 2021. Its national strategy for AI (SNIA) was launched in March 2018, following the report by mathematician Cédric Villani, with an announced budget of 1,5 billion euros for 2018-2022. A second phase, called "acceleration," was announced in November 2021 and largely implemented from 2023. A third was opened in February 2025, during the AI Action Summit held in Paris.

The accounting assessment by the Court of Auditors in November 2025 is more sober than the announcements. For the first phase, the State ultimately committed less than 1,3 billion euros, with credit monitoring deemed "incomplete." For the second, it programmed 1,1 billion euros over 2023-2025, a third less than announced, and only 35 % of the credits had been consumed by 30 June 2025. In other words, the cumulative public effort amounts to nearly 2,4 billion euros over eight years, and part of the promised money never reached the ground.

What worked: research and the ecosystem

The Court acknowledges real successes. Research has been structured around centers of excellence and computing infrastructures open to researchers; more than 4 000 French researchers now work on AI. France moved from thirteenth place in the Global AI Index in September 2024 to fifth in September 2025. The number of French AI start-ups has doubled since 2021: more than 1 000 were active in 2025, and they raised nearly 2 billion euros in 2024. Sixteen French unicorns integrate AI into their offerings.

The symbol of this success remains Mistral AI. In September 2025, the Parisian start-up raised 1,7 billion euros, including 1,3 billion from Dutch company ASML, for a valuation of 11,7 billion euros. A private success, but one that also owes to the research environment funded by public money.

What stalled: diffusion into the economy

The setback is clear. The second phase was supposed to make the dissemination of AI in businesses its priority; according to the Court, support for demand "has only benefited from very modest measures," and the expected acceleration "did not take place." Training, the transformation of the administration, and action towards the regions are also among the blind spots.

The "Dare AI" plan, launched on 1^er July 2025 with a budget of 200 million euros, was supposed to correct this weakness. One year later, the government claims 35 000 businesses have been made aware, but only 70 Data AI diagnostics have been initiated, for a new target of 1 000. Awareness is progressing; the actual equipping of SMEs remains to be demonstrated.

June 2026: 655 million euros more, directed towards computing

On 16 June 2026, on the eve of VivaTech, Prime Minister Sébastien Lecornu announced 655 million additional euros via France 2030, intended to "support infrastructure, computing capacities, research, businesses, and industrial sectors." The government specifies that the envelope should support French and European capacities, particularly within the framework of the AI Gigafactories of the European Commission and the Important Project of Common European Interest (IPCEI) dedicated to AI.

The same announcements target the State itself: deployment of a sovereign conversational assistant, "The Assistant," for public agents before the end of 2026, and the choice of the French company ChapsVision by the DGSI instead of the American company Palantir. Sovereignty becomes an explicit public procurement criterion.

The real imbalance: public versus private

These public amounts must be compared with the 109 billion euros in private investments announced during the summit in February 2025, primarily for data centers. A large portion comes from foreign capital: up to 50 billion euros from the United Arab Emirates and 20 billion from the Canadian company Brookfield, according to the Élysée. The State sets the course and funds research, but the physical infrastructure of AI on the territory will be largely owned by external investors.

The question of sovereignty is therefore shifting: it is no longer just about how much the state spends, but who controls computing capacity, data, and models. In this regard, the Court recommends investing in sovereign storage capacities, building a "realistic" ambition for electronic components, and better aligning AI policy with energy and connectivity policies.

Measure before spending more

The most structural criticism concerns governance. The Court calls for creating a general secretariat for AI attached to the Prime Minister, better integrating the budgetary constraint, and, above all, conducting a thorough evaluation of results of the first two phases, launched one after the other without prior evaluation. It also advocates for better complementarity with the European level, territories, and the private sector.

For a policy that aims to be strategic, this is a major weakness: without shared result indicators, it is impossible to say whether a euro invested in computing yields more than a euro invested in training or in supporting SMEs. Yet it is precisely in this area of adoption that France is incurring its most costly delay.

What to remember

Public money has indeed served a purpose: more visible research, a start-up ecosystem twice as dense, and a European champion in language models. But funding has been lower and slower than promised, diffusion into the economy has lagged, and evaluation remains to be built. The new envelope of 655 million euros bets on computing; the next test will be to prove that it also benefits businesses and public services.

Sources

Sep 22, 202618 min read

Open or closed models: in Europe, value is no longer determined by size

Open models are now only 4.4 months behind closed ones, according to Mozilla. But the open frontier is Chinese, and value is shifting to computing and platforms: openness alone is not enough for sovereignty.

Deep DiveRead article

Sep 6, 202618 min read

AI in the civil service: one in five employees exposed, and a transformation still without a pilot

The figure of 36% of public jobs exposed to AI is a global estimate from Roland Berger. For France, the inspectorates general estimate that one in five employees is significantly exposed and warn of a massive "shadow AI" and fragmented governance.