# Prometheus Plan
Prometheus Plan: France can win the AI race. What would it cost?
- This note proposes the Prometheus Plan: a three-year project for France and its partners to build a frontier AI lab able to compete with the best American labs, in order to secure their strategic independence.
Breaking free from AI dependency demands an unprecedented effort
- In an economy irrigated by large language models, not producing one's own flows of artificial intelligence means depending on others for a growing share of one's productivity, industrial power and security.
- France today has no frontier model (the highest tier of capability), and the recent American export controls on Anthropic's best models illustrate the danger of this situation.
- No lab of that rank will emerge organically: it takes a deliberate effort by the state.
- France is the only power, outside the United States and China, that meets the conditions to become the third nation of the frontier, provided it makes this a first-order national priority.
The technical objective
- Aim for 12 GW of compute in 2029, in Blackwell equivalents (trajectory: 2 GW in 2027, 7 GW in 2028, 12 GW in 2029), the level of the large American labs.
- Attract the world's best researchers within a lean team (1,700 people).
- Be able, by 2029, to train models at the frontier.
The cost
- €148bn from 2027, €271bn in 2029, a cumulative total of around €620bn over three years. Compute concentrates most of the cost (95%).
- That represents 4.5 to 8% of French GDP, of which 1.5% of public investment per year: an effort on a historic scale.
The proposed architecture
- We propose to distinguish two blocks:
- An autonomous scientific lab, in which the state would hold only a minority stake (25%) while keeping instruments of strategic control.
- A vast infrastructure programme (compute, energy, land), steered by the public authorities, backed by a derogatory “Prometheus Act” to accelerate procedures and by a financial programming component.
The coalition
- Led by France, with clearly assumed leadership made credible by its financial commitment.
- Open to partners (European, but also South Korea, Taiwan, Japan, the United Kingdom, the Emirates, Canada, Australia…) who contribute to the financing in exchange for guaranteed access to compute and to the models.
Main sticking points
- Beyond the sheer enormity of the cost, the second difficulty lies in chip supply, and therefore in the initial dependence on Nvidia and on the American administration.
The strategic conclusion
- The cheaper alternatives (betting on open source, or negotiating an interdependence with the United States and China) offer no guarantee of real autonomy.
- The decisive question therefore remains: are the defenders of French sovereignty prepared to pay the price?
# The only question that matters
The American administration recently imposed export controls on Anthropic's most advanced large language models, Mythos and Fable 5. The decision to restore access, taken on 30 June 2026, marks the entry into a new regime: no administration will want to lock itself into rigid transparency criteria for granting the best artificial intelligence models access to the market. A kind of strategic ambiguity will be the rule. The American government will seek to preserve maximum room for manoeuvre, including the ability to withdraw access to a model without notice and at its discretion, or to restrict its use cases.
That decision reminded everyone of the urgency of the situation: artificial intelligence is becoming a resource as important as electricity or oil in our economy. Securing its supply is now a decisive strategic question. In an economy now durably irrigated by large language models (LLMs), not having autonomous access to the best models, those we call frontier models, means depending on others for a growing share of our productivity, our industrial power and our national security. If certain model capabilities cross a threshold (autonomy on long tasks, greater reliability, falling inference costs), adoption could break sharply upward, potentially triggering a rapid economic decoupling of the countries that lack them. In fact, no economic question, and probably no strategic one, matters more today than this: will we be subject to the goodwill of foreign powers to supply our society with machine intelligence, or will we be able to produce it ourselves?
In artificial intelligence, France certainly has many advantages, starting with its nuclear fleet and the excellence of its researchers and engineers. It hosts the only European lab able to produce large language models of reasonable size, Mistral. But that is still a long way from having a lab able to supply first-rank artificial intelligence, and to follow or even lead the advances of the frontier. In that respect, we are in fact starting almost from zero.
If France and Europe have failed to produce champions as powerful as Anthropic or OpenAI, it is for structural reasons tied to the dispersion of capital, the regulatory environment, and probably to cultural behaviours. It is hard to imagine changing these parameters quickly: the unification of European capital markets is a perennial mirage, and so is the simplification of national and European law. One can safely say that a cutting-edge artificial intelligence lab will not develop there organically. Any political will to put France back into the artificial intelligence race can therefore only work through a deliberate effort by the state.
But what effort are we talking about, and on what scale? Most of those who call today for sovereign artificial intelligence do so without measuring its price. Yet the political equation changes entirely with the order of magnitude. Is it equivalent to the cost of a new aircraft carrier, for example? Is it rather an effort comparable to the Messmer plan, which for a time mobilised more than 1% of French GDP to build the fleet of nuclear reactors that is still our energy backbone? Does it go further still?
In this note we propose a “Prometheus Plan”: we compute the amounts to invest to create and sustain a frontier lab in France within three years, before determining by what means to get there and drawing the strategic conclusions that follow.
# Creating a frontier lab for France
We deliberately reason about a lab dedicated to large language models, because it is today the most mature, and therefore the most reproducible, general-purpose artificial intelligence technology. Betting on still unproven paradigms, such as world models, remains useful and desirable, but as a complement to an industrial project of this kind, just as it was necessary in the 1970s to copy and deploy in France the American light-water reactor design, independently of French research on alternative reactor designs.
## What is a frontier lab?
The term “frontier”, in AI, describes a dynamic unfolding over time more than the performance of a model at a given moment: the duration, measured in expert human time, of the tasks an agent can complete with a given level of reliability has doubled roughly every 7 months. Most benchmarks, quickly saturated, become less fine-grained instruments for measuring model performance than minimum conditions of entry to the frontier. Access to the frontier cannot be thought of as the one-off purchase of a model. It requires a durable capacity to track and absorb the dynamic. At the frontier, the compute needed to train a model has doubled every 5.2 months since 2020. Over the same period, the training cost of frontier models has doubled every 7 months. Hardware and algorithmic efficiency gains are also rapid but do not offset the rise in ambition: they mostly allow labs to aim for more capable models, longer to train, more agentic and more inference-intensive.
Labs at the AI frontier rest on the command of an integrated set of resources to finance the next iterations of model development and to ensure their large-scale deployment. Their advantage owes as much to the scientific quality of the technical teams as to continuous access to the capital, data and compute infrastructure needed to push the frontier. Increasingly, they also rest on access to the AI models themselves, which assist developers in producing future models, or even run their own experiments to automatically accelerate R&D; this is the process known as recursive self-improvement, and it could lead to a rapid widening of the gap between the labs enjoying the best models (their own) and their competitors.
## Compute, the sinews of war
The fundamental principle of AI today, and the sole reason AI companies are throwing themselves into an outsized investment race, is called the scaling laws. These are empirical laws by which a model's intelligence grows linearly with the logarithm of the compute deployed, whether in training the model or later at inference time.
Scaling laws: chart published by OpenAI in the technical report of their o1 model, December 2024
Granted, this formula gets ever more expensive as the compute invested multiplies, but the promise is fabulous: to reach an unbounded intelligence, vaster than that of our greatest polymaths, able to contribute more to scientific and technical progress than all our Nobel laureates. Of course, nothing guarantees that these laws will keep holding in the long run. But they have held so far, which incidentally suggests that intelligence is a matter of quantity. It is this promise that makes the investment so important: as AI becomes a superiorly intelligent entity capable of giant leaps in every scientific and technical domain, and therefore also in weaponry, a state's sovereignty will not survive without command of AI.
Winning the race is therefore mainly a matter of compute, as the current state of the global race confirms. The two labs at the frontier today, Anthropic and OpenAI, are also the ones holding the largest reserve of computing power (compute). It is now most often measured in energy used: each of these labs controls the equivalent of several GW. And it is indeed compute that makes the difference:
The performance measure here relies on the Artificial Analysis index, a useful though very incomplete indicator of the frontier. The real gap between Fable 5 and the Chinese labs is certainly larger than the roughly 4-month gap observed here. First, the index saturates at the top: it compresses the gaps between the best models. That compression can give the impression of a moderate lag, even though the differences remain much sharper on the hardest tasks (on ARC-AGI-2, for instance, Chinese models are still ~8 months behind). Second, an aggregate average hides the real shape of the frontier. Being “four months behind” says nothing about which capabilities remain out of reach today, nor which classes of tasks are unlocked only by the best available model. At the frontier, the lead is not linear but cumulative: the best model serves to train, distil and accelerate the next one, captures the most profitable uses and the best talent, and sets the standard the others chase.
Given the performance of models developed by Chinese “pure players”, with resources in principle more limited than the large American labs, some now argue that frontier models can be trained with far less compute, and therefore far less cost, than the American norm. GLM 5.2, released on 16 June 2026 by Zhipu, rivals Claude Opus 4.8 on some benchmarks, although Zhipu is thought to hold only a fraction of Anthropic's compute (and as an allocation rather than compute of its own). But for one thing, the true financial and compute capacities of Chinese firms are hard to assess. For another, it is clear that part of their models' performance comes from distilling American commercial models, that is, using those models to generate training data and environments. Taking the example of a DeepSeek or a Zhipu to conclude that Europe could have a cut-price frontier lab would be a mistake. In a sense, distillation is access to compute by proxy, where the cost of the frontier was first paid elsewhere. Finally, a model like GLM 5.2 does not measure up to Anthropic's and OpenAI's models on other benchmark suites, and appears less versatile and less capable overall.
It is also true that a large share of the compute available at OpenAI and Anthropic goes to inference, that is, to users running the models, not to training them. In itself, it might be tempting to think training is possible with far less compute, but here again the reasoning would be fallacious, since it is precisely the data resulting from the massive use of their models by users that eases the training of the next models. Training and inference must therefore be distinguished architecturally, but not opposed strategically. For a lab at the frontier, compute forms a strategic portfolio to arbitrate between training, R&D, internal inference (RL, synthetic data, research automation) and customer inference. Scale effects play at every link: 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. Finally, inference also enables test-time scaling, that is, improving performance with more compute at the moment tasks are solved (see the Scaling laws chart above). Inference is thus an essential economic and technical engine of the frontier.
We conclude that, if France wants its own lab able to supply it with frontier AI over the long run, it needs an organisation whose scale and means are close to those of the large American labs.
# What would such a project cost?
For simplicity, let us set aside at this stage the existing vehicles that could host and run this project, and reason from first principles: how many GPUs, how much energy and how many researchers would it take to create this lab?
## Compute
At the end of 2025, OpenAI had around 1.9 GW and Anthropic 1.4 GW; both labs are expected around 5 to 6 GW each as early as the end of 2026. We propose to set as an objective a size of the order of 12 GW of IT load in 2029. These gigawatts are counted in Blackwell equivalents, the latest chip generation: since each accelerator generation is more efficient than the last, new generations would be adopted as they are released, so that at constant electrical power the effective compute capacity would keep growing. That amounts to catching up with the projected level of the frontier players, moving fast: 2 GW in the first year, 7 GW in 2028, 12 GW in 2029. For comparison, recall that the Stargate programme plans $500bn for 10 GW and that Anthropic has reserved some 10 GW with Amazon, Google and Broadcom.
The first cost covers only the compute base: purchasing and building owned capacity, temporarily leasing capacity to bridge the gap at the start, then the annual operation of the capacity held. Adopting Epoch AI's orders of magnitude for simplicity, we count about €33bn per GW purchased, €7.4bn per GW leased per year, and €0.8bn per GW owned per year in OpEx. The compute financing requirement would then be around €141bn in 2027, €191bn in 2028, then €261bn in 2029, some €593bn cumulatively over three years. Of that total, most is construction and equipment CapEx, around €529bn, the rest covering leases (€52bn) and the OpEx of owned capacity (€11bn).
The associated electricity consumption in 2027 is about 20 TWh per year, PUE included, for 2 GW of available IT load. With the chosen ramp-up, consumption would reach about 71 TWh in 2028, then 121 TWh in 2029, for 12 GW available. IT load counts only the servers' power draw: the total power called by the sites is higher, since ancillary consumption, cooling and electrical conversion losses must be added on top (with the chosen PUE of 1.15, the 12 GW of IT load in 2029 thus correspond to about 13.8 GW of total installed power).
We also model a transitional compute lease of 1 to 3 GW per year over the period, at a cost of €7.4bn per year per GW. This leased capacity supplements owned capacity during the ramp-up, and notably lets the researchers' work start faster in the very first year.
In France, five sites are slated to host more than 700 MW each by 2030-2032. That timeline owes much to grid-connection lead times of 4-5 years on large sites, including those under public sponsorship. Securing an aggregate 1 GW on these sites in the first year would be possible by going beyond the current French fast-track siting regime and imposing a first-tranche logic, site by site, on the most advanced projects. Examples like Colossus show that a single site can add 300 MW in about 7 months by combining a reusable industrial site, mobilised suppliers, accelerated grid connection and broad regulatory tolerance. The 12 GW objective by 2029 is an exceptional catch-up trajectory; the five fast-track sites would have to be pushed almost to their target regime as early as 2029. Moreover, this trajectory requires broadening the portfolio beyond the existing large sites alone, combining site extensions, industrial conversions and possibly some European capacity controlled by French or European players. It also requires cutting time-to-market by accelerating grid-connection works as much as possible, or by cleaning up the queues of PTF (Propositions Techniques et Financières, the grid-connection offers agreed with RTE, the French grid operator). By cutting time-to-market, the state makes these sites far more attractive to investors and developers.
## Recruitment
Compute is an indispensable base, but it is not enough to reach the frontier; the failure of Meta and xAI so far to rival Anthropic and OpenAI, despite sizeable capacity of about 4 and 1.5 GW respectively, shows the importance of researcher quality and lab culture.
While the skills of French researchers and engineers are often, and rightly, praised, Europe nonetheless lacks experience, and therefore expertise, in the current training of frontier models. If we want to compete, we must pay the price to attract researchers from the large American labs (which, in some cases, will in fact mean bringing European talent home).
There is no need for an enormous headcount: we assume a lean team of about 1,700 people, where Anthropic has 3,000 employees and OpenAI 5,000 (with product and marketing dimensions we leave in the background for now). Compensation, however, is high, especially at the top: it is essential to mobilise some fifty elite researchers, drawn from the best competitors, for whom we budget €2.6bn of annual compensation: close to €90m per founder (even if that could take the form of equity) and €40m per world-class researcher. That is the market price: nine-figure dollar packages have become ordinary in poaching between labs. Note in passing, even though we focus here on financial means, that money is not enough: the difficulties encountered by xAI show that beyond pay levels, the ability to durably attract the best researchers also depends on scientific culture and governance.
The base of the organisation indicatively gathers 300 senior researchers and engineers (around €3.5m each), 400 specialists in distributed training, systems and infrastructure as well as 300 people dedicated to data, post-training and evals (about €1.7m each), plus 150 people on R&D tooling, 150 on security and cyberdefence, 250 on product and 150 on legal, PR, HR and support functions.
In total, payroll comes to around €6bn in 2027, then rises to about €7bn in 2028 and €8bn in 2029, €20.8bn cumulatively over three years. It remains far below the cost of compute: about 3.5% of cumulative compute. Talent is thus much cheaper than GPUs, which justifies all the more paying it extremely well. Payroll grows more moderately than compute in the build phase, around 15% per year: the team stays lean, but compensation keeps climbing under competitive pressure.
## Total
On this basis, the total annual cost comes to about €148bn in 2027, €200bn in 2028, then €271bn in 2029, a cumulative total of the order of €620bn over three years. Compute makes up the overwhelming majority: on its own it weighs about €593bn cumulatively, 95.7% of the total, the rest covering payroll and various operating expenses (€20.8bn and €6.1bn cumulatively, respectively).
The investment is therefore more than massive: the effort represents about 4.5% of French GDP in 2027, 6% in 2028, then 8% in 2029.
## Bootstrap, self-sustaining trajectory and financial returns
Massive as it is, the effort required for such a race to the frontier would nonetheless be bounded and limited in time, particularly from the public authorities' point of view. If reaching the frontier is so hard for an AI lab that achieving it in France would demand massive, multiform state support for several years, that bootstrap phase would be brief. In case of success, once the frontier is reached, private demand for the lab's services would be such, and its appeal to international investors so great, that it would naturally find itself on a self-sustaining, sustainable trajectory. The massification of private demand and the project's attractiveness to foreign investors would be all the greater as it would fully benefit from the interest of European players (and of other countries) eager to de-risk from dependence on American and Chinese export controls. Once the bootstrap phase is over, one can therefore reasonably posit that a European frontier lab, if it remains well run by its leaders, should be able to hold durably at the frontier and probably to widen the gap with competitors that have not reached it, without requiring any further public financing to do so.
The investments, public and private alike, made in such an effort would moreover be positioned to be particularly profitable financially. Before even discussing the economic, political and strategic value of such a project, recall that a frontier lab is also a first-order productive asset. The effort is not sunk: it can generate revenue if it succeeds, and retain substantial value even if it fails.
Most of the large labs' compute goes to inference, which is monetised through billed API access (per token), through consumer and enterprise subscriptions, and increasingly through agentic products that automate whole swathes of intellectual work. The revenues of the leading labs are growing at annual doubling rates or better, reaching tens of billions of dollars a year. A French lab would moreover enjoy a structural advantage, a partly captive sovereign market. In time, in case of success, most of the annual costs could thus be self-financed by revenue and the rest by private investment.
Besides, even if the lab failed to reach the frontier, France would remain the owner of multiple GW of datacentres. These are durable, valuable assets. That compute could be leased out (on-demand inference, sovereign cloud), put at the service of the European research and startup ecosystem, or redeployed to other intensive workloads. Note, however, that the GPUs themselves depreciate brutally, with a useful life of only a few years, which justifies the amortisation built into our model. The patrimonial base lies above all in the shell, the energy and the land. But that is not the hardest part to obtain, nor the least valuable.
# How to carry this plan through?
Given the enormity of the means to devote to the project, the architecture chosen to deliver it is decisive. How should the resources invested in this herculean effort be organised? Should several companies be put in competition? Or should the means be concentrated in a single one, and if so, which? And what role should the state play?
## Vehicle
If one had to reduce to an equation the performance of the models coming out of a lab, one could write, from the observations above:
performance = compute × data × organisation × brains
OpenAI, Anthropic, Meta and xAI are notably ahead on compute.
The organisation factor, by which we mean the firm's ability to organise itself properly to run its research effectively, is perhaps the limiting reagent for some companies, such as the overweight hierarchy at Meta and to a lesser extent at Google.
Conversely, the Chinese companies, more compute-constrained, hold their own on the brains factor, with brilliant researchers and first-class engineering.
Note, however, that in the era of Reinforcement Learning, where AI can learn by experimenting on its own, by "playing" in test environments, data increasingly comes down to something the compute factor can produce.
Concretely, several mechanisms convert compute into data. With good environments and many rollouts, a model generates, critiques, filters and rewrites its own training trajectories, replacing armies of annotators. On a hard mathematics or programming problem, one can sample a hundred answers, submit them to a verifier, unit test, compiler, formal proof or reference answer, and keep only the correct and most instructive trajectories (this is the rejection sampling used by DeepSeek-R1). A second model can also simulate the environment directly, tool calls, state transitions, user replies, as in work such as Qwen-AgentWorld. Finally, more inference compute lets a lab serve more users, whose massive usage in turn feeds error identification and the enrichment of training sets.
But once this equation is set, how do you maximise the product? Should one invest in one champion, or several? There is naturally a tension between competition and the concentration of means. In the United States, competition let multiple players emerge, each of which played its part in the country's progress: it was Google that discovered Transformers, OpenAI that discovered inference scaling laws with o1, Anthropic that shipped the coding agents opening a path towards self-improvement. The Chinese example is singular in its own way: the government maintains a fleet of datacentres that is progressively ramping up. Neo-labs like Zhipu, Moonshot or MiniMax then compete for grants, time-limited usage authorisations; obtaining a grant is of course conditioned on past performance, which lets competition surface new ideas but disperses the effort accordingly.
By contrast, for a state with limited means, the scaling laws narrow the options: any organisation aspiring to world-class AI will need roughly as much compute as one of the American leaders, which directly requires several gigawatts, hence hundreds of billions of euros. That rules out from the start that France, alone or even in coalition, could finance several champions at once. The investment will therefore inevitably have to be focused on a single vehicle. This national-champion logic, unavoidable as it is, would nonetheless demand particularly well-designed governance to avoid sterilising its innovation potential.
The choice would then be between two paths: building on an existing lab, with Mistral naturally appearing as the only credible European candidate, or creating an entirely new vehicle. The first option would make Mistral the project's anchor, having it absorb all the means made available to the Prometheus project and adopt the race-to-the-frontier roadmap. The second would consist in creating a dedicated structure, possibly acquiring Mistral's researchers and assets, if that form allowed faster execution, tighter governance or a better fit with the programme's objectives. Mistral is currently valued at about twenty billion euros, which would represent only a seventh of the present project's annual cost.
Then comes the question of the state's role. Two conditions emerge clearly. The first is that the project will never succeed without a massive commitment of the public authorities, financial first, but also political, diplomatic and regulatory. Only the state, deploying its full power and its capacity to concentrate means, can make the Prometheus Plan a priority national effort on the order of what the construction of the French nuclear deterrent was in the 1960s and 1970s. Note, for those who doubt the relevance of state involvement in a frontier lab, that OpenAI has just offered the American administration a 5% stake. The second is that the state adds no value to the running of the lab itself, and would even most certainly be harmful there. Yet it would be politically difficult to defend financing of this magnitude without any public oversight.
The heart of the arrangement would therefore consist in splitting the project in two. On one side, the lab proper: the researchers with their own culture and free in their scientific choices, entirely devoted to training models, with a general roadmap towards the frontier as their only frame. On the other, everything else, which is in reality an immense infrastructure project meant to deliver compute to the lab in the desired quantities, costs and timelines; that is something the state knows how to run.
At the head of the public side, a state programme directorate would act as the owner's representative on behalf of the state. Its role would be entirely turned towards facilitating the project: accelerating administrative procedures and putting in place the many indispensable derogatory regimes. A “Prometheus Act” equivalent to the "Notre-Dame law" should be passed covering land, grid connection, environmental law and labour law.
The “Prometheus Act” would also include a financial programming component setting massive public investment in the project for 3 years at 1.5% of GDP per year. The precedents of the great French and American techno-industrial national projects (the French nuclear deterrent, the Manhattan Project, the Apollo programme) tend to show that this is the maximum credible order of magnitude for a highly ambitious programme designated by the state as an absolute national security priority. Essential to make the project credible and to attract the French and foreign private investment meant to take over from them, these public funds, far from being burned without counterpart, would be investments liable to generate considerable returns in case of success. If it succeeds, France would find itself endowed with a priceless strategic asset within an ultra-restricted club of frontier powers (the United States, China and France) and with a means of playing an essential catalyst role for European strategic autonomy. Investing 1.5% of GDP of public money in such a project appears, under these conditions, more than justified. All the more so since, once the bet is won, the project would no longer require any additional public financing. Procedurally, having the state invest such amounts in the Prometheus Plan would however require playing European state-aid law well: the national security card should be used without reservation to keep the Commission at bay, even at the price of a legal test of strength with it.
The lab would take the legal form of a holding company whose sole mission would be to deploy frontier models, with no interference in its management. The state would hold a significant minority stake, of the order of 25%, giving it a stake in its success, with the rest remaining majority private. Ideally, stakes would be multiplied among European funds and large industrial groups: for the latter, not so much for their capital as for their direct interest in having a sovereign frontier model that cannot be switched off from abroad. The lab's founders and leaders would not necessarily have to be French; it is enough that they be first-class and experienced in training frontier models. The state, on the other hand, would guarantee national control of the project, with classic corporate-law instruments: a golden share giving the state a targeted veto over sensitive operations grounded in national security imperatives, clauses on headquarters location and non-relocation of critical assets, a guarantee that the model cannot be switched off from abroad, and so on. Multiple voting rights, permitted in France since the June 2024 Attractivité corporate-law reform, would decouple the economic from the strategic within the lab, with the state, BPI (France's public investment bank) and the founders holding the reinforced-vote shares while foreign capital would come in through shares with high economic yield but weak or no votes.
Compute, for its part, would be housed in separate structures, dedicated subsidiaries or joint ventures, whose sole purpose would be to deliver the planned power to the lab. They would bring together private investors of all origins and public investors from willing countries, with remuneration calibrated to be frankly attractive: as these vehicles are capital-hungry, the yield on offer must measure up to drain in the money needed. Precedents exist. America's Poolside separated its lab from its infrastructure company, whose leadership gathers profiles with ten to twenty years of experience building and operating datacentres for the big cloud players. Mistral structures its access to compute through joint ventures with BPI, MGX and Nvidia.
One last public structure, distinct, would have the mission of putting new nuclear « on steroids » so that energy does not become the bottleneck after 2030. Developing that dimension is not the purpose of this note, but it is no less essential.
Backed by the public investment that makes the project credible, national savings would be mobilised massively in the service of Prometheus through minimum mandatory allocations in PER and assurance-vie accounts (France's tax-advantaged retirement savings and its main life-insurance savings vehicle), a Tibi-type scheme steering insurers' and institutional investors' allocations, and a listed vehicle that the general public could hold in a PEA (the French tax-advantaged share account). Risk tranching, with the state taking the first-loss tranche if need be, would, by absorbing the uninsurable risk tied to the sovereignty bet, attract private capital in large quantities. At the scale of the plan, reallocating 2% per year of the national assurance-vie stock towards Prometheus would suffice to match a public investment of 1.5% of GDP, but would require a complete, ambitious and well-designed financial architecture.
Given the mass of private capital to attract, the project's success would depend critically on its ability to make its ambition credible as quickly as possible in the eyes of private and international players. The scale of the means, both public and private, committed by France, but also its boldness in putting major legal derogations at the project's service and its ability to recruit first-rank international talent to lead it, would be essential. Fast first results on the models would then be imperative to set a virtuous circle in motion.
## What coalition around France?
In projects of this kind, a recurring reflex is to start from a European logic without demonstrating its necessity. That too often leads to the pursuit of immediate geographic returns (as in space or defence), hence to the fragmentation of the effort and, in the end, to a failure to scale. That is why we start from French capabilities, with other convergent countries, including non-European ones where appropriate, able to join according to needs and interests.
France is indeed today structurally best placed to attempt to become the 3rd country at the AI frontier because, except for the United States and China, it is the only power combining the following four conditions:
- sufficient critical economic mass (unlike the Netherlands, Israel, Switzerland or Singapore);
- a national skills base in the field (unlike Germany);
- enough strategic autonomy from the United States not to risk being blackmailed with the withdrawal of the security guarantee (unlike the United Kingdom, Japan and Korea);
- credible access to the international AI talent pool (unlike Russia and India).
To this is added the French tradition of great technological catch-up projects, through civil and military nuclear power in particular, which would provide precious experience and inspiration for such an effort.
Its success would require clear governance resting on clearly assumed French leadership, made credible by France's massive and durable financial commitment, both public (through its dedicated programming act) and private. The other participating countries would benefit from the project through compute access rights, usage credits for their researchers and companies, shared standards and the progressive integration of their industrial players into the value chain. Beyond that, they would receive from France a guarantee of permanent access, in case of success, to a frontier model for their needs and those of their companies, on conditions equivalent to those of French companies. Through a multilateral treaty they would be invited to join, the partners would receive such a commitment from France, in exchange for a significant commitment to finance compute capacity dedicated to the project. Besides the partner countries' contribution to financing the project through an entry ticket (e.g. 0.3% of their GDP per year for 3 years), they would probably play an essential role in the project's first years by facilitating the lab's access to compute located on their territory, since the compute capacity that can be built in France is limited in the short term by electricity availability, until the acceleration effort on new nuclear bears fruit and can take over.
It is not the purpose of this article to determine precisely which partners might be interested, a subject that should be the object of a meticulously designed diplomatic manoeuvre. One can however note that, provided the French project appears credible, a number of middle powers could see great value in using participation in this project as a form of hedging that limits their dependence on American and/or Chinese models. Japan and above all South Korea and Taiwan appear as priority targets given their command of the semiconductor industry, but their strategic dependence on the United States could make them hesitate. The same goes for the United Kingdom, particularly interesting for its AI ecosystem. All our European partners should also be offered to join the project. Beyond the E5 (France, Germany, Italy, Poland and the United Kingdom), given their positioning, the Netherlands, the Scandinavian countries and Belgium look especially likely to be interested. Beyond Europe and East Asia, the United Arab Emirates, Canada and Australia could also be serious candidates.
## The first three years
How would the project unfold in practice? We propose the following calendar for its first three years.
| Period | Main objective | Compute capacity | Expected model | Comment |
| ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| T0 – T+3 months | Launch politically and legally
Create the vehicle with its various structures, appoint an executive and scientific leadership, pass the Prometheus Act
| Immediately negotiate the lease contracts; reserve the first sites and the energy | · | Without a single vehicle and committed capital, nothing starts. The derogatory act must be passed in the first months. |
| T+3 – T+6 months | Make the lab operational: recruit the founding core (founders and first leaders); launch the elite-researcher packages | Secure 1 GW leased and purchase the first owned 1 GW | First internal models, to test the stack, ablations, reproductions of the recipes that work | The goal is not the frontier but building the “model factory”. Leasing decouples training from datacentre construction. It allows a start at month 6 rather than month 24. |
| T+6 – T+12 months | First credible generation | Leased compute at scale; first owned orders; final site selection; construction starts | First credible public model, strong for its size | A window comparable to a “European Grok-1”: a strong political and technical signal, but not the frontier |
| T+12 – T+24 months | Enter the real race | Mix of leasing + first owned tranches; first datacentre modules delivered | First serious large model, possibly 6-12 months behind the best American models | This is the first credible window to approach the frontier |
| T+24 – T+30 months | From « good lab » to « frontier candidate » | 4 to 9 GW of dedicated owned capacity if the sites were accelerated; leasing to smooth the peaks | 2nd generation; frontier reached on some segments (code, agents, cyber, science) | The window where the project becomes credible. |
| T+30 – T+36 months | Reach or brush the general frontier | Ramp ideally towards 12 GW; then reduce the dependence on leased compute | 3rd generation; model at the frontier or immediately below | Three years is a credible horizon to enter the frontier then hold there |
These projections rest on the durations observed at the most recent large labs:
| Lab | Trajectory | Duration | Decisive accelerator |
| --------------- | ---------------------------------------------- | ------------ | -------------------------------------------------------------------------------------------------------- |
| xAI | founded Jul. 2023 → Grok-3 (Feb. 2025) | ~19 months | Datacentre built in 122 days in a disused factory, mobile turbines, permit derogations |
| DeepSeek | founded 2023 → V3 (Dec. 2024) / R1 (Jan. 2025) | ~18-20 months | Pre-existing GPU cluster assembled before the American export controls (High-Flyer hedge fund) |
| Moonshot (Kimi) | founded Mar. 2023 → Kimi K2 (11 Jul. 2025) | ~28 months | Team of about 80 people at the start, highly efficient MoE architecture, open-weights strategy |
## International opposition
One of the great difficulties would be, beyond the administrative conduct of this sprint, securing the GPU supply. Building the datacentres themselves, structuring the project legally, the regulatory derogations are in our hands, but access to chips is not. No lab today trains models without Nvidia chips, except Google, which relies on its own Tensor Processing Units, built expressly for its internal needs. Even the recent Chinese models are trained on GPUs produced by Nvidia. In the short term, and as long as other players do not challenge this quasi-monopoly, our project would therefore depend on Nvidia, that is, on the American administration. Nothing would prevent the latter from imposing export controls on Nvidia products bound for France (some European countries have already been subjected to them), and it might be tempted to do so upon seeing a competing effort of this size arise. The current American administration, wedded to “AI dominance” and little inclined to spare its allies, could resort to such measures.
Two solutions then present themselves. Either we attempt to acquire Nvidia chips anyway, betting that such a market would be too juicy for the company, which would use its influence in Washington to stave off export restrictions; in the worst case, we would have to convince the Netherlands to threaten Nvidia and the American government with retaliation by halting exports of ASML's photolithography machines, indispensable to GPU production, which makes The Hague's participation in the project all the more essential. There is much uncertainty there. Or we seek to benefit from new producers elsewhere in the world. But it is hard, to date, to find any that are not themselves held by the Americans, more or less directly, while having the required scale. Even Huawei's chips (Ascend) are not sufficient for the Chinese labs' own consumption. Perhaps the situation will change soon; South Korea, for instance, has announced an investment plan of more than $1,000bn for semiconductor plants.
As things stand, this is one of the most serious risks the project faces. It is not prohibitive, however, because the cost to the United States of an openly hegemonic use of its export controls against an ally would probably strongly accelerate the emergence of alternatives to Nvidia, particularly in Asia. In the meantime, the United States could not prevent Prometheus from leasing compute capacity from third parties to train its models.
# Strategic alternatives
## The other paths are just as risky
Such a massive cost and such a defining effort could discourage French and European political leaders, who might judge themselves unable to finance and conduct it. What, then, would France's alternatives be?
The first option is a bet on open source: so far, open models have trailed the performance of commercial models by a few months, and the European economy can therefore, as a last resort, turn to the former to secure its supply of artificial intelligence, without depending on anyone. The bet is risky and, moreover, carries a far from negligible cost. To run these models, sufficient energy and infrastructure are needed, which in itself implies enormous investment. Above all, nothing guarantees that the flow of open models will not dry up. Their developers, today mainly Chinese, can suddenly decide to make them commercial. In that case, Europe would find itself just as dependent on China as on the United States. These models could also suffer the same access restrictions as the American models, closed for instance beyond a « Mythos »-type capability threshold. If these models were to catch up with the frontier, they could be branded supply-chain risks by the American administration, triggering restrictions, or even bans, on their purchase, use or hosting by American companies. Through extraterritoriality, compliance or regulatory pressure, their adoption would also become harder for European companies.
The second conceivable path is a negotiated dependence on the Americans. The United States builds the models and Europe uses them. To avoid total submission, some argue that Europe holds levers in the value chain that it can use in turn to put pressure on the United States; in other words, that sovereignty should be found in mutual dependence rather than in independence. It is true that production chains are dispersed across the world and that no one unilaterally controls all the components needed to develop and diffuse large language models. But let us not delude ourselves: the main brick the Europeans control, the production of photolithography machines for the semiconductor industry (through ASML), is not such an effective lever in a long-run interdependence. It is an indispensable link in the chain, but with high latency: if Europe blocked ASML's exports, the effects would only be felt after many months, since the other players could draw on their existing stocks. Conversely, blocking a frontier model is instantaneous for the United States, as the example of Fable 5 has shown. The ASML weapon can possibly serve in a standoff focused on infrastructure, as noted above, but it is less effective when it comes to the final product and in steady state. In a world where the European economy rests entirely on the use of American models, could Europe afford to go even a few days without them? Nothing is less certain.
At best, European states can try to mitigate this disadvantage by conditioning American suppliers' access to the European market on the physical presence of the models in datacentres located in Europe and legally controlled by them. South Korea, for instance, has formalised a partnership between Shinsegae, a national conglomerate taking charge of the physical infrastructure, and the American startup Reflection AI, bringing the open-weights models and the engineering, to build a 250 MW site presented as a great sovereign “AI factory” for Korean companies and administrations. That would avoid handing the United States an instantaneous “kill switch”, as they have today; but it would not prevent Washington from deciding, at leisure, to halt the deployment of future models in Europe. Furthermore, the European Union has not demonstrated its ability to coordinate a unanimous, vigorous response to pressure of this type, since its structure invites free-rider behaviour.
Can we, finally, hope to play on the rivalry between the United States and China so as to depend on neither, by putting them in competition? That would be a risky game: the United States would be well aware that a Europe threatening to switch to Chinese models would immediately place itself in China's hands, and vice versa, so that we would have little room to negotiate.
Perhaps these solutions are the ones France and Europe will adopt. In that case, we should at least stop invoking the concept of European sovereignty and independence, and admit that the task will be to manage our dependence as best we can. Many European countries are already used to this where their energy supply is concerned, and it will therefore be their natural slope.
In the end, it seems to us that the major bet that the Prometheus project represents is warranted because its risks are commensurate with the technological and strategic tipping point that AI is producing before our eyes. Against the costs and risks of this effort, capable of mobilising the nation's living forces as nothing has in half a century, one must indeed set those of inaction. For we know this road of least effort and we know where it risks leading: to the strategic insignificance of France in a world durably dominated by the masters of silicon intelligence.
## Why act now
Contrary to what one sometimes reads, the AI frontier is not destined to become commoditised within a few years, in the sense that one could simply wait for the most advanced capabilities to become accessible to all, at low cost, with no strategic loss for a nation. For some everyday uses (summarising, translation, office assistance for example), being a generation behind may suffice: open or semi-open models already provide abundant, cheap capability. But the capabilities that matter for power (cyberdefence, autonomous agents able to execute long tasks, R&D acceleration, biological or military applications) sit precisely at the frontier. They are also the ones whose access will be increasingly controlled.
Developing expertise at the frontier moreover produces returns that exceed the trained model alone. The leaders of today's main labs were, yesterday, researchers or engineers placed closest to that frontier: Dario Amodei, Demis Hassabis, Ilya Sutskever, Arthur Mensch, among others. A French lab at the frontier would build the human, organisational and industrial base allowing France to stay in the race in 2028, 2030 and well beyond.
Without a national programme at the frontier, a country becomes progressively dependent on evaluations produced by others. It loses the capacity to measure autonomously the real performance of the most advanced models, to identify their vulnerabilities and to assess the systemic risks they could pose to its critical infrastructure, its national security or its strategic interests.
Frontier labs now explicitly seek to automate part of their own R&D: the rapid automation of AI research and of its development cycle sustains the continuous improvement of frontier models. Previous model generations are used to train the next ones. In this scenario, being absent from the frontier does not cost linearly more: it forfeits access to the very engine of the acceleration.
Finally, inaction deepens an already massive industrial dependence. The United States today concentrates most of the performance of the large AI compute clusters, as well as priority access to the supply chains that feed them. The longer France waits, the more these resources, compute base, talent and infrastructure, consolidate elsewhere.
The window opened by France's position and the state of the field exists for a brief moment. In five years it will have closed for good: it is today that history knocks at the door.
# Our conclusions
General de Gaulle recounted Khrushchev's reaction in 1960 when, at Rambouillet, he told him of the success of the French atomic bomb: « I understand your joy. (...) But, you know, it is very expensive. » The French president commented: « My account provoked no reaction from my interlocutors, except this one: “Ah yes! It is very expensive!” “It is very expensive”, even for the Americans, even for the Russians. (...) But for us, facing these imperial designs, it is the price of independence. »
We believe the same holds for access to artificial intelligence. The current state of our economies only faintly reflects the place it will take everywhere in the years to come. Of course, most countries in the world will not be able to claim to produce their own models or their own inference, just as most countries in the world cannot produce their own energy. They will have to settle for negotiating, willingly or not, with foreign powers to feed their needs. But we already see where energy dependence leads; one need only think of the difficulties of many European states during the invasion of Ukraine. Dependence on machine intelligence will run at least as deep.
Since 1945, France has chosen to develop its own source of energy, through the nuclear industry. It was a massive, risky investment, but it now secures our energy autonomy, makes us a net exporter, and contributes to our place in the game of world powers. Now that everything turns on access to artificial intelligence, a challenge of the same order presents itself again.
In fact, it may be the very same challenge; since the bulk of a frontier lab's costs lies in compute, and compute, in the end, demands energy first, one can see the rise of a French AI as the logical continuation of our nuclear effort. Even the method of the nuclear programme can serve as inspiration. The 1945 ordinance creating the Commissariat à l'énergie atomique stated: “It appeared that this body had to be at once very close to the Government, and so to speak mingled with it, and yet endowed with great freedom of action. It must be very close to the Government because the fate or the role of the country may be affected by the developments of the branch of science to which it devotes itself, and it is consequently indispensable that the Government have it under its authority. It must, on the other hand, be endowed with great freedom of action, because that is the sine qua non condition of its effectiveness.” That is also the philosophy we favour for building a frontier lab.
At the time when France chose, at the end of the 1950s, to embark alone on an immense effort to acquire the atomic weapon and, more broadly, a complete and independent nuclear deterrent, many French people and our main allies were convinced that it lacked the means and was doomed to fail. The “reasonable people” of the hour called, like today's, for France to “manage its dependence” by coming to terms with the United States on the model provided by the Nassau agreements with Great Britain. Had it followed that path, France would probably have had its bomb, but maintaining the deterrent, to this day, would depend on the goodwill of the United States. It took General de Gaulle's iron will to discern that the nuclear weapon had become so essential to national sovereignty that its autonomous command justified considerable sacrifices. Today, AI places us before the same choice.
Yes, “it is very expensive.” It is extremely expensive, and strewn with obstacles at home and abroad. But it is also a great project able to kindle enthusiasm, to carry other partners along with us, to mobilise our best forces. Finally, launching the Prometheus Plan means taking seriously the vocation France has long set for itself, to be the spearhead of European strategic autonomy, by actually giving itself the means. And if we renounce undertaking works of this order, we may as well renounce speaking of sovereignty and of European strategic autonomy.
By that measure, whether or not to launch the Prometheus Plan will very probably be the most important decision facing the next president of the Republic, to be elected next year. It could well make the difference between preserving the role France intends to play in the world and its slide, in time, towards strategic insignificance.