The ambition displayed during the February 2025 AI Action Summit is fading. The British start-up Fluidstack is giving up on establishing its giant supercomputer in Bosquel, near Amiens. This project represented an investment of €10 billion. It was supposed to be the first phase of a massive 1 gigawatt (GW) computing infrastructure.

Bloomberg reports that this withdrawal also affects a second site south of Paris. This infrastructure was intended to serve the needs of Mistral AI, the French champion of generative AI. “A British start-up gives up on its €10 billion data center project in France to concentrate its growth in the United States,” confirms the American agency.

This withdrawal reduces the €109 billion plan announced by the Élysée. Last January, Bercy claimed to have secured €67 billion in investment. Fluidstack’s departure mechanically diminishes this progress. The RTE network manager and the Somme Sud-Ouest Community are already looking for a replacement for the Bosquel site.

Last November, Fluidstack signed a $50 billion contract with Anthropic. This agreement involves the creation of custom data centers in Texas and the state of New York. This contract will generate approximately 800 direct jobs in the United States.

The relocation of Fluidstack’s headquarters to New York in December signaled this shift. Google’s investment intention of injecting $100 million into the company reinforces this trajectory. Proximity to major research laboratories and cloud giants facilitates access to the massive capital required to purchase the latest generation of chips.

France, however, offered technical advantages. The Bosquel site had reserved access to a high-voltage power grid. Emmanuel Macron emphasized in 2025: “Our nuclear energy is controllable, safe, stable, and decarbonized, ideal for expanding our computing capacities.” These arguments were not enough against the depth of the American order book.

Fluidstack’s withdrawal weakens the value chain of French AI. Without local computing infrastructure, companies like Mistral AI risk depending on servers located outside the European Union. This dependence raises questions about data confidentiality and service latency for European clients.

The French plan aimed to create an alternative to American and Chinese cloud providers. Of the 13 strategic sites initially announced, 10 seemed secured by early 2026. However, Fluidstack’s project failure frees up valuable electrical power. RTE can now allocate this capacity to other industrial or computing projects.

Experts point to the absence of major regulatory barriers in France. The choice is strictly economic. The United States offers a denser market where demand for computing capacities saturates the existing supply. For a growing company, priority goes to signed and funded contracts.

The French government must now convince other players to take over the abandoned computing capacities. Global competition for the purchase of GPU chips makes each project dependent on delivery times from manufacturers like Nvidia. A delay of a few months can render a site obsolete even before its inauguration.

The next step for the government will be to announce the new winners for high-performance computing sites by the end of 2026. The revision of Bercy’s investment plan is expected to adjust the objectives of installed computing power. The priority remains the maintenance of a sovereign infrastructure to prevent the brain drain and the flow of algorithms to Silicon Valley.