FAQ

France has fallen behind, but it still has the conditions needed to catch up: an economy large enough to finance the effort; a first-rate scientific and technical base; the ability to attract international talent; genuine strategic autonomy vis-à-vis the United States; and abundant, largely decarbonised electricity. It has, in particular, a player like Mistral and an exceptional pool of researchers and engineers, many of whom now work in American labs. So it is not starting from scratch.

The current gap rests largely on means that can be assembled: compute, talent, capital, energy and the ability to execute fast. But the scale of the investments is rising quickly. The very purpose of Prometheus is to concentrate these means fast enough to reach about 12 GW of compute in 2029.

The other risk is that the lead of the first labs becomes cumulative. Access to the best models already helps write code, automate research, design new experiments and improve the next models. It is the beginning of a form of recursive improvement: AI contributes to producing better AI. If we let three years pass without acting, catching up with the frontier will become economically and technologically impossible for a country like France.

This is not an additional item of current spending but an exceptional investment, limited to a three-year catch-up period. The order of magnitude, as a share of GDP, is similar to what France committed to deploy its nuclear fleet, and that is why we chose it: the state has already sustained an effort of this kind. It represents less than a third of the current annual public deficit. It would raise the public debt ratio by less than four points. In case of success, Prometheus could then finance its operations by selling models, services and compute.

Recall that the alternative is not free. Not investing means buying, for decades, the American services on which our companies, administrations and infrastructure will depend.

Arthur Mensch estimated it, before the Assemblée nationale, at $1,000bn a year: « If we import non-European technology, that is a trillion of trade deficit to add to our existing trade deficit on digital services. If Europe's only role is to be the energy supplier, then ninety percent of the value leaves Europe and gets invested elsewhere. »

The question can also be put more simply: do we want the automation of a considerable share of French work to flow, as through communicating vessels, to American companies? A large fraction of the payroll currently paid in France would be converted into subscriptions, API calls and purchases of foreign services.

The project is not designed to create jobs, and would create few relative to its cost. The direct jobs can be computed: the lab itself would offer about 2,000 jobs, and datacentre operations, taking the usual figure of 300 to 500 people per gigawatt in steady state, about 5,000 jobs. Building the infrastructure would involve about 3,000 to 5,000 workers per GW during construction, that is, over three years, roughly 100,000 cumulative job-years. For indirect jobs, with standard multipliers for construction and technical services in France, the figures above can roughly be doubled. That would put permanent jobs, direct, indirect and induced, in the range of 20,000 to 40,000, on top of the temporary construction jobs.

But that is not the main point; the main point is that we run the risk of seeing hundreds of billions of dollars of French and European value added leave for the United States if we remain dependent on their models. If Prometheus captures even 10% of that flow, the employment effect passes through the spending of its revenues, with potentially hundreds of thousands of indirect jobs.

The social consequences of automation may be comparable, but owning the models would let France keep a share of the revenues, set the conditions of access, serve its critical sectors first, adapt the systems to its companies and its law, and not depend on a foreign authorisation to run its economy. Moreover, the French would benefit from the economic returns in France and, if they take part in the investment, from the financial returns.

The targeted consumption (about 121 TWh per year in 2029, PUE included) would be supplied by the most decarbonised power fleet of the large economies: the carbon intensity of the French grid stands around 20 to 40 gCO₂eq/kWh depending on the year (RTE data), against roughly 350 to 400 for the United States, where most of the world's compute is being built today. The electrical opportunity cost must however be counted: the note establishes that consumption would be covered without new capacity, by raising the load factor of the existing fleet (up to 100 TWh) and the margin of the roughly 90 TWh currently exported. Those exports are not climate-neutral: exported French nuclear electricity substitutes, in some neighbouring countries, for mostly fossil generation. Redirecting part of those flows to Prometheus would therefore carry an indirect carbon cost. Datacentre cooling can be designed as closed-loop or dry cooling, at the price of an energy overhead (depending on the site's location and design), which brings direct water consumption down to marginal levels.

Prometheus would compete with OpenAI, Anthropic, Google, Meta and xAI to recruit some of the rarest researchers and engineers in the world. Many of the talents working in those labs are French, but their skills are now paid at the global market price. Salaries in the tens or even hundreds of millions of euros are the norm for this type of position in the large labs. Since payroll is in any case an almost negligible cost next to the infrastructure, this is not the place to economise.

With Prometheus, France could become the third power of artificial intelligence and the spearhead of European strategic autonomy. It could provide citizens, companies and administrations with advanced AI agents, export intelligence services, develop robotics and use the productivity gains to finance its social model and preserve its standard of living.

Without Prometheus, the French economy would probably use just as much artificial intelligence, but it would buy it mainly from the United States or China. French companies would pay a growing share of their revenues to foreign suppliers, while the most sensitive sectors might receive less advanced models or models subject to political restrictions. The exporting powers could blackmail us on any subject from that lever.

They could benefit in three ways. First as users: every citizen could access agents able to help with their work, their administrative steps and their daily life, then in time robotic systems. Next as workers and taxpayers: the productivity gains could strengthen wages, companies, public revenues and the financing of pensions. Finally as savers: part of the project could be opened to French savings, notably through vehicles compatible with the PEA or the PER.

The game is not over: Meta has just released models of a very good level (Muse Spark), as has xAI, which shows that holding a lot of compute remains an essential determinant of success. Besides, we plan an organisation that leaves the lab itself its full freedom and autonomy: the state's involvement would consist in launching the initiative and then confining itself to building the compute infrastructure, without interference in model development. There are precedents: it was the state that entirely created the Compagnie Française des Pétroles (CFP) in 1924, which later became the private company Total, with the success we know.

The Minitel did not fail through poor execution, but because France had bet on the wrong paradigm at the moment the rest of the world was choosing another. The Prometheus Plan consists, on the contrary, in copying a paradigm already validated by the market: LLMs and the scaling laws. The right parallel is 1970s France copying American light-water reactors rather than persisting with its national designs; the risk of picking the wrong technology is therefore considerably reduced.

In the status quo scenario, every token bought from OpenAI or Anthropic already embeds the cost of Nvidia GPUs, plus the AI lab's margin. The rent paid to an American lab, by contrast, is avoidable in the short term. An embargo on chips would freeze our capacity growth; an embargo on models, as the Fable 5 episode showed, cuts our supply of intelligence instantly and entirely. Prometheus's GPU cost is a discrete import spike and a durable asset. Available data show that accelerators keep value after the next generation ships. Epoch AI observes that the H100's price stayed in a low-to-median range of $20,000 through 2024 and 2025, with no clear downward trend in public prices. In early 2026, H100 rental indices even rose on the back of demand, and the secondary market still attributed a high residual value to them. Conversely, consuming models one does not produce is an additional, continuous burden on the trade balance. The real choice is therefore not between a perfectly autonomous France and a France dependent on Nvidia. It is, in the short term, between two configurations: depending on Nvidia as well as on the labs, their models and American jurisdiction; or depending initially on Nvidia for chips, while mastering in France the infrastructure, the models, the skills, the revenues and the access decisions.

Prometheus does not claim to spare France an ambitious diffusion policy. However, a diffusion policy without production capacity of our own would largely amount to subsidising the purchase of foreign services. Consider the hypothesis of 10% of the payroll being automated by AI agents: the question becomes whether 10% of the European or French payroll is replaced by American AI agents, or whether we want to keep that value.

The frontier cannot be carved up. Every path to a frontier-level model in cyberdefence or mathematics demonstrated to date goes through a generalist base with state-of-the-art pre-training. The capabilities that matter for security, for example long-horizon reasoning, agentic capabilities over multi-hour tasks, command of code, are not obtained by training a small model on a specialised corpus: they emerge from scale, that is, from massive pre-training. The experimental models behind new results in mathematics themselves come from the frontier labs. For instance, the AlphaProof Nexus framework behind the resolution of Erdős problems is based on sub-agents built on Gemini 3.1 Pro. National security requires a continuous flow of capabilities tracking the frontier.

It is key to distinguish the cost of a given level of capability, which collapses over time, from the cost of the frontier, which explodes. Both are true simultaneously, and it is the second that matters for Prometheus. At a given model capability, algorithmic efficiency gains (model architectures, optimisation and training methods) and hardware gains mean a lower training cost to develop the model and lower inference costs (down by a factor of 10 each year for the last 3 years). Performance per dollar has improved by about 30 to 40% per year across accelerators released between 2012 and 2025, a doubling roughly every 2 to 2.5 years. Some observations suggest that, at a given performance, the cost of a model will be divided by 4 each year thanks to technological advances. In other words, if training a model costs $100m today, that cost falls to $25m a year later, then $6m in two years, and so on. However, at the frontier, labs reinvest those efficiency gains, for example by opening new scaling axes such as test-time scaling, which improves results by spending more compute at resolution time. Each fall in cost per token opens uses that were not viable before and can make total demand explode.

On several occasions, the cofounders of the Chinese AI labs have named compute as one of the main limitations of the Chinese labs relative to the American ones. Lin Junyang, former technical lead of Alibaba's Qwen team, puts at less than 20%, an assumption he already deems « very optimistic », the probability that a Chinese company overtakes players like OpenAI within the next three to five years. He attributes that gap mainly to American compute resources one to two orders of magnitude larger.

Moreover, the Chinese labs have more limited compute for inference, notably for consumer uses and B2C deployments. That constraint weighs on the overall quality of the user experience: some models can post excellent scores on certain benchmarks, for example code, while remaining less polished for other uses. Thus, in the Kimi K3 release blog, the developers note that the user experience remains noticeably below that of Claude Fable 5 and GPT-5.6 Sol, despite close aggregate scores on Artificial Analysis.

Having the compute to serve the largest number of users is also an essential economic and technical engine of the frontier. The more users a lab serves, the more it learns to cut its cost per token, to improve its kernels, its routing, its batching or its accelerator utilisation; the more efficient its inference becomes, the more intelligence it can sell, the more revenue it can generate, the more data and usage signal it can capture and reinvest in the next cycle.

A frontier lab can invest on the order of $0.5bn to $1bn a year in data acquisition: licences for external content, human annotation on frontier tasks, where each example can cost tens of thousands of dollars, or RL infrastructure and environments. But to first order, compute is becoming an ever more reliable proxy for the quality of the data a lab really has access to. Several mechanisms contribute to this.

Synthetic data generation is a first source. With good environments and many rollouts, a model can generate, critique, filter and rewrite its own training trajectories. More simulation steps mean richer, more diverse, higher-signal data; compute powers the procedural generation of tasks, error correction and the curriculum's rising difficulty. A powerful model that produces and refines the training examples of the next model replaces armies of annotators.

For a hard mathematics or programming problem, the lab can also generate a hundred different answers, submit them to a verifier (unit test, compiler, formal proof or reference answer), then keep only the correct and most instructive trajectories. The larger the compute budget, the more solutions can be sampled, the more rare reasoning paths explored, and the cleaner the dataset built on problems close to the limit of the model's capabilities. DeepSeek-R1, for example, used outputs produced by its own model and selected by rejection sampling to create new supervised training data.

Model-simulated environments are also a way to convert compute directly into data. Rather than building an executable environment out of tools and APIs, a second model simulates the environment: the agent emits a tool call, and the simulator model generates the tool's response, the state transition or the user's reply, from the tool definitions and the interaction history alone. Work like Qwen-AgentWorld simulates a whole series of agentic environments (MCP, search, terminal, SWE, web, OS, Android) inside a single model.

Finally, larger inference compute allows the models to be deployed to more users. The data from that massive usage then helps identify errors, enrich the training sets and improve the following model generations.

Connecting a large datacentre to the grid often plays out over 4-5 years, sometimes more, whereas the firm commitment of an end client (an AI lab, for instance) can only come at a much later stage of the project, often in the 12 to 18 months before first power-on. The site developer therefore carries the risk over a long period. It must advance cash to RTE to accelerate orders, and spend 3-4 years justifying a site where little is visibly happening, with some €30m to disburse to make it connectable.

Two issues govern the connection: the availability of capacity, and the scale of the investment works.

Inference demand, which grows faster than compute supply, has so far kept equipment value at levels atypical for computer hardware. Three indicators converge.

First, prices for new hardware show no erosion. The H100's sale price held in a range of $25,000 to $40,000 per unit from mid-2024 to early 2026, that is, over most of its commercial life, despite the launch of the next generation (Blackwell), itself sold at a higher unit price, so strong is the growth in demand. This persistence of the H100's value, which should in theory have depreciated in favour of the far more powerful next generation, is explained by the algorithmic gains mentioned above: a single H100 can accomplish more and more tasks.

Second, the secondary market shows an abnormally small discount by enterprise-hardware standards. CoreWeave, the leading specialised compute lessor, indicates that its A100s acquired in 2020 remain fully booked, and that a batch of 2022 H100s reaching the end of their contract was immediately re-leased at 95% of its original rate.

Third, utilisation rates of older generations remain maximal. Nvidia confirmed in November 2025 that A100s, delivered from 2020, still run at full utilisation six years after entering service; Google reports 100% utilisation of its seventh- and eighth-year TPUs.

100%, as the leasing of Colossus 1 to Anthropic shows.