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 quickly. But the scale of the investments is rising fast. Prometheus's goal is precisely to concentrate these means fast enough to reach around 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 makes it possible to write code, automate research, design new experiments and improve the next models. This is the beginning of a form of recursive improvement: AI helps produce better AI. If we let three years pass without doing anything, it will become economically and technologically impossible for a country like France to catch up with the frontier.

This is not an additional recurring expense, but an exceptional investment, limited to a three-year catch-up period. The order of magnitude, as a share of GDP (1.5% of GDP over three years, 4.5% in total), is similar to the effort made for the nuclear rollout in France, and that is why we chose it: the state has already backed 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 operation through the sale of models, services and compute.

In the central scenario, the state's contribution would reach about €70 billion a year for three years. That sum is, however, a ceiling rather than an irreducible need for direct budgetary financing. Public intervention can take several forms in order to maximise the leverage on private capital.

The state could first act as guarantor, in order to reduce the risk borne by investors. It could also greatly improve the economics of the projects by securing long-term electricity supply contracts for the datacentres. In a global context marked by the scarcity of quickly available electrical capacity, access to abundant, decarbonised and predictable energy is in itself an asset likely to attract international developers and operators.

Finally, the public contribution would not be solely financial. An exceptional law, inspired by the one adopted for the reconstruction of Notre-Dame, could radically shorten the timelines for authorising, connecting and building datacentres. By reducing regulatory risk and time to commissioning, the state would increase the economic value of the programme while lowering the amount of public capital needed.

One must also distinguish the share of the capital financed by the state from the share of power it retains in governance. A structure with several classes of shares could, for instance, grant the state class A shares carrying prerogatives strictly confined to the protection of strategic interests, while private investors would hold class B shares offering attractive economic rights. Subject to preserving a balance acceptable to those investors, the state could retain a veto over a limited number of essential decisions: change of control, disposal of strategic assets, technology transfers, location of infrastructure or access to compute. These rights would not extend to the lab's scientific, technological, operational or commercial choices, which would remain the responsibility of its management.

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

Arthur Mensch estimated it, before the National Assembly, at $1,000 billion 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, that means ninety percent of the value leaves Europe and is invested elsewhere."

The question can also be put more simply: do we want the automation of a considerable share of French labour to benefit American companies through communicating vessels? A large fraction of the payroll paid in France today 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 estimated: the lab itself would offer around 2,000 jobs and the operation of the datacentres, if one takes the usual figure of 300 to 500 people per gigawatt in steady state, around 5,000 jobs. Building the infrastructure would involve about 3,000 to 5,000 workers per GW during the works, that is, over three years, around 100,000 cumulative job-years. For indirect jobs, using standard multipliers for construction and technical services in France, one can roughly double the figures above. That would give permanent jobs, direct, indirect and induced, of the order of 20,000 to 40,000, on top of the temporary construction jobs.

But the most important point is elsewhere: we run the risk of seeing hundreds of billions of dollars of French and European added value leave for the United States if we remain dependent on their models. If Prometheus captures even 10% of the flow, the effect on employment passes through the spending of its revenue, 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 revenue, set the conditions of access, serve its critical sectors first, adapt the systems to its companies and its law, and avoid depending on foreign authorisation to run its economy. Moreover, French citizens could benefit from the economic spillovers in France and, if they take part in the investment, from the financial returns.

The targeted consumption (around 121 TWh a year in 2029, PUE included) would be powered by the most decarbonised grid among the major economies: the carbon intensity of the French grid stands at 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. One must, however, take account of the electricity opportunity cost: the note establishes that the consumption would be covered without new capacity, through raising the load factor of the existing fleet (up to 100 TWh) and the margin of the roughly 90 TWh currently exported. These exports are not climate-neutral: exported French nuclear electricity substitutes, in some of our neighbours, for mostly fossil sources. Redirecting part of these flows towards Prometheus would therefore carry an indirect carbon cost. The cooling of datacentres can be designed as a closed loop or as air cooling (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. Much of the talent working in these labs is French, but their skills are now paid at the global market price. Salaries in the tens, even hundreds, of millions of euros are the norm for this kind of role in the large labs. Since payroll is in any case a nearly negligible cost compared with the infrastructure, there is no point economising on it.

With Prometheus, France could become the third power in artificial intelligence and the spearhead of European strategic autonomy. It could supply citizens, companies and public 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 revenue 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 using that leverage.

They could benefit in three ways. First as users: every citizen could gain access to agents able to help them in their work, their administrative tasks and their daily life, and in time to robotic systems. Then as workers and taxpayers: the productivity gains could strengthen wages, companies, public revenue and the funding 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 yet over: Meta has just released very strong models (Muse Spark), as has xAI, which clearly shows that holding a large amount of compute remains an essential determinant of success. Moreover, we envisage an organisation that leaves the lab itself full freedom and autonomy: the state's involvement would consist in launching the initiative before confining itself to building the compute infrastructure, without interfering in the development of the models. 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, by contrast, consists in copying a paradigm already validated by the market: LLMs and the scaling laws. The parallel is rather that of 1970s France, which copied American light-water reactors rather than clinging to its national designs; the risk of betting on the wrong technology is therefore considerably reduced.

There is a genuine similarity in the motivations of the two projects (a fear for France's technological sovereignty), but the Prometheus Plan diverges decisively from the Plan Calcul in its design, precisely because it draws the lessons of that failure and does not repeat its mistakes. Five structural differences change everything.

• Scale: the plan aims for compute parity with the leaders (12 GW), whereas the Plan Calcul funded at a tenth of IBM's means. Prometheus is built precisely against that error of under-sizing. In this respect, it is rather the current European initiatives, scattered and sub-critical, that replay the Plan Calcul.
• The depth of the rival's lead: whereas the CII faced twenty years of IBM lead, proprietary architectures and customers locked in by their installed base, the AI frontier is only four years old and its recipes diffuse (publications, open weights). DeepSeek, Moonshot and xAI have shown that it can be approached in 18 to 28 months, which was inconceivable against IBM.
• The nature of the product: in the Plan Calcul era, computers were sold through commercial networks, which allowed lock-in through maintenance and software, whereas the token is sold through an API with low switching costs, a clear asymmetry favouring the challenger.
• Sovereign demand: in the Plan Calcul era, the administration bought Iris machines under duress. Today, demand for AI is already massive, including within the state for sovereign applications bearing on national security. Prometheus's domestic market is of a wholly different scale from that of the Plan Calcul.
• The proposed structure, finally. The Plan Calcul did not merely fail: it failed for decades without anyone stopping it. Our arrangement of clearly established exit milestones is exactly the anti-Bull. As for the unity of command essential to the project, which determines its first and foremost national character, it avoids the mistake of Unidata, the multinational consortium derived from the Plan Calcul in which three rival manufacturers were meant to design together.

The catch-up we are aiming for requires speed, and therefore a clear chain of decision. A programme dependent on the permanent agreement of many states would risk spending several years dividing up the financing, the sites, the jobs and the industrial spillovers, which would be disastrous for the project.

In the status quo scenario, every token bought from OpenAI or Anthropic already incorporates the cost of Nvidia GPUs, plus the AI lab's margin. The rent paid to an American lab, however, 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 off our supply of intelligence instantly and entirely. Prometheus's GPU cost is a discrete import spike and a durable asset. The available data show that accelerators retain value after the next generation is released. Epoch AI observes that the price of the H100 stayed in a low-to-median range of $20,000 in 2024 and 2025, with no clear downward trend in public prices. In early 2026, H100 rental indices even rose under the pressure of demand, and the secondary market still assigned them a high residual value. Conversely, consuming models one does not produce is an additional and continuous drag 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 the chips, while mastering in France the infrastructure, the models, the skills, the revenue and the access decisions.

Prometheus does not claim to exempt France from an ambitious diffusion policy. However, a diffusion policy without a production capacity of one's own would largely amount to subsidising the purchase of foreign services. If one considers the hypothesis of automating 10% of the payroll with AI agents, the question comes down to whether we replace 10% of European or French payroll with American AI agents, or whether we want to keep that value.

The frontier cannot be sliced up. All the paths to a frontier-level model in cyberdefence or mathematics demonstrated to date go through a generalist base and rely on a state-of-the-art pre-training. The capabilities that matter for security, for example long reasoning, agentic capabilities on tasks of several hours or 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 rests on sub-agents built on Gemini 3.1 Pro. National security requires a continuous flow of capabilities tracking the frontier.

It is essential to distinguish the cost of a given capability level, 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, gains in algorithmic efficiency (model architectures, optimisation and training methods) and hardware efficiency reduce the training cost of the model as well as inference costs (a tenfold reduction each year for the last three years). Performance per dollar improves by about 30 to 40% a year on the accelerators released between 2012 and 2025, that is, 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 four each year thanks to technological progress: in other words, if training a model costs $100 million today, that cost will fall to $25 million a year later, then to $6 million two years on, and so forth. However, at the frontier, labs reinvest these efficiency gains, for example by opening new scaling axes such as test-time scaling, which improves results by spending more compute at solving time. Every drop in cost per token opens up uses that were previously unviable and can make total demand explode.

On several occasions, the co-founders of the Chinese AI labs have named compute as one of the main limitations of the Chinese labs relative to the American ones:

• Tang Jie, co-founder of Zhipu AI and formerly of Google DeepMind, stressed that, despite the release by Chinese labs of strong open-source models, the United States would keep a lead thanks to frontier models still unreleased, backed by markedly greater compute and financial means. The efficiency gains imposed by the constraint can partly offset this disadvantage, without however closing the gap.
• Yang Zhilin, founder of Moonshot AI (which develops the Kimi models in particular), presented compute in 2025 as a central factor of production for "Moonshot" projects, that is, the scaling-up of frontier models.
• Lin Junyang, former technical lead of Alibaba's Qwen team, puts at less than 20%, a hypothesis he already deems "very optimistic", the probability that a Chinese company overtakes players like OpenAI within the next three to five years. He attributes this gap mainly to compute resources one to two orders of magnitude greater in the United States.

Furthermore, the Chinese labs have more limited compute for inference, notably for consumer uses and B2C deployments. This constraint weighs on the overall quality of the user experience: some models can post excellent performance on certain benchmarks, for example on code, while remaining less polished for other uses. Thus, in the release blog for Kimi K3, the developers note that the user experience remains appreciably 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, 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 $0.5 to $1 billion a year in acquiring data: licences for external content, human annotation on frontier tasks, where each example can cost several tens of thousands of dollars, or else RL infrastructure and environments. But to first order, compute capacity is becoming an ever more reliable proxy for the quality of the data a lab actually 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 means richer, more diverse and higher-signal data; compute powers the procedural generation of tasks, error correction and the ramping difficulty of the curriculum. A powerful model that produces and refines the training examples for 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 it becomes possible to sample solutions, explore rare lines of reasoning and build a clean dataset on problems close to the model's capability limit. DeepSeek-R1, for instance, used outputs produced by its own model and selected through 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 interaction history alone. Work such as Qwen-AgentWorld simulates a range of agentic environments (MCP, search, terminal, SWE, web, OS, Android) within a single model.

Finally, greater compute for inference lets models be deployed to a larger number of users. The data arising from this massive usage then helps identify errors, enrich the training sets and improve subsequent generations of models.

The connection of a large datacentre often plays out over four to five years, sometimes more, whereas the firm commitment of an end customer (an AI lab, for example) can only come at a much more advanced stage of the project, often in the 12 to 18 months preceding the first energisation. The site developer therefore bears the risk over a long period: it has to advance cash to RTE to speed up the orders. For three to four years, it must thus justify a site where not much is happening and disburse some thirty million euros to be able to connect it.

Two factors condition the connection: the availability of capacity and the extent of the investment works.

• Availability: one can physically route a production surplus towards a zone. Today, Auvergne-Rhône-Alpes and Normandy are under strain, and the Île-de-France is saturated.
• The grid's ability to carry the energy that will be drawn without endangering the rest of the network, hence, often, the need for investment works.

How to speed things up on these two fronts?

• On availability: the problem comes from the congested PTF queues. A PTF (Technical and Financial Proposal) is the connection commitment concluded with the transmission grid operator, RTE in France. Concretely, a land developer asks RTE for a connection for a given power (X MW) and pays to enter the queue. These queues are congested: for example, hydrogen projects conceived three or four years ago around a power of 100 MW block the connection where the capacity exists. As long as such a project has not left the stack, it slows the connection where there is capacity.
• On the speed of the connection works: it depends first on environmental issues, whose very long procedures can take more than a year, then on the ordering of equipment (transformers and cables in particular), which raises issues of cash advances and supply-chain optimisation. RTE itself, as a buyer, must optimise its supply chain for transformers and cables.

Inference demand, which is growing faster than the supply of compute, keeps the value of the equipment at levels atypical for computer hardware. Three indicators converge.

• First, the prices of new hardware show no erosion. The selling price of the H100 held in a range of $25,000 to $40,000 a unit from mid-2024 to early 2026, that is, over most of its commercial life cycle, despite the arrival on the market of the next generation (Blackwell), itself sold at a higher unit price, so much does demand keep rising. This persistence in the value of an H100, which ought nonetheless to depreciate 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 the standards of enterprise hardware. CoreWeave, the leading specialised compute lessor, reports 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, the utilisation rates of older generations remain maximal. Nvidia confirmed in November 2025 that the A100s, delivered from 2020, are still running at full utilisation six years after they were commissioned; Google reports 100% utilisation of its seventh- and eighth-year TPUs.

Let us take as a criterion of competitiveness the fact that our model is several months ahead of the best open-source frontier of the moment. Two cases arise.

• If our lab manages to release, on a regular basis, models clearly better than the best open-source model of the moment, then the mission's objective is at least partly met, endowing France with a cutting-edge power.
• In the contrary case, unlikely, where the lab's models fail to clearly surpass the best open source of the moment, two possibilities would remain: either lease this available compute at a premium (around $50 billion a year per GW, that is, hundreds of billions a year for the whole) to one or more other labs (the limits on available electrical power in the United States guarantee that several American labs will be compute-constrained, and therefore takers for a lease), or use it as we see fit for our own inference needs with the open-source model in question.

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

Even if the lab failed and Prometheus fell back on a compute-leasing activity, the initial public investment could remain exceptional, under two assumptions: that prices do not fall durably below about 35 to 40% of the rate observed on Colossus, and that at least 70% of the fleet is leased on average.

The Colossus contract provides an upper bound: Anthropic must pay $1.25 billion a month for around 325,000 GPUs split between Colossus I and II, that is, $15 billion a year and about $5 to $7 per GPU-hour. The contract is, however, terminable on 90 days' notice and must therefore not be treated as a market price guaranteed until 2029.

Taking, as a hypothesis, the density of Colossus I, around 220,000 GPUs for a little over 300 MW, a 12 GW fleet leased at 70% and billed at only half the Colossus rate would generate around $140 billion in annual revenue. Its fully annualised cost would be of the order of $102 billion, based on the $8.5 billion per GW per year estimated by Epoch AI. The fleet would therefore still yield nearly $40 billion a year after accounting for operations, amortisation and equipment renewal.

The break-even point would sit around $1.9 per GPU-hour at 70% occupancy, that is, about 36% of the Colossus rate. Prices could thus fall by nearly two-thirds before the leasing activity ceased to cover its full cost.