In Brief
This note presents for the first time a three-year plan for France and its partners to develop a frontier AI laboratory capable of rivaling the best laboratories in the United States, with the aim of ensuring strategic independence.
AI will be the main engine of the economy by 2030
- R&D expenditure on AI, as a share of wages, already exceeds 10% in the most advanced firms and will continue to rise, potentially overtaking wage share in the long run. AI is already today, by a large margin, the primary driver of American growth.
Not being autonomous in AI would expose us to imperial powers
- Importing our AI capabilities translates into relying on others for an increasing portion of our productivity, our industrial strength, and our security.
- France currently has no frontier model (the highest level of capability), and recent American export controls on the best Anthropic models illustrate the danger of this situation. Yet no frontier lab of this caliber will emerge organically without deliberate state action.
- Apart from the United States and China, France is the only country today that possesses the conditions to become the third frontier nation, provided that we make it an absolute priority.
The technical objective
- Aim for 12 gigawatts (GW) of computing power by 2029 (trajectory: 2 GW in 2027, 7 GW in 2028, 12 GW in 2029), i.e., on par with the major American laboratories.
- Attract the world’s best researchers into a tight team (1,700 people).
- Be capable by 2029 of training frontier models at the border. We show that this objective is attainable.
The cost
- 170 billion dollars (Md$) by 2027, over 300 Md$ in 2029, i.e., a cumulative around 700 Md$ over three years. The calculation concentrates the bulk of the cost (95%).
- This would represent 4.5 to 8% of French GDP, including 1.5% public investment per year, constituting a historically large effort: a central budgetary decision in the mandate of the next President of the Republic.
- The remainder of the funding would be provided by private sector investments, but also by private or public capital from other middle-power countries with an interest in the project (to reap direct benefits or to avoid their own dependence on the Sino-American duopoly, at a moment when these two powers are closing the taps of AI to their customers).
The proposed architecture
- We propose to distinguish two blocks: a scientific autonomous laboratory, where the State would own only a minority stake (25%) but would retain strategic control instruments; and a broad program of infrastructure (compute, energy, real estate), led by the public sector, backed by a permissive “Prometheus Law” to accelerate procedures and a financing planning component.
Main blockers
- Beyond the sheer cost and the political acceptability in a democracy of such an investment to support a technology facing strong pushback, the second difficulty lies in supply of chips, and thus in initial dependence on Nvidia and the American administration.
Strategic conclusion
- Less costly alternatives (betting on open source, or negotiating interdependence with the United States and China) offer no guarantee of real autonomy.
- The crucial question remains: are the champions of French sovereignty willing to pay the price?
The Only Question That Matters
The United States administration has recently imposed export controls on the largest language models developed by Anthropic, Mythos, and Fable 5. The decision to restore access on June 30, 2026 marks a new regime: no administration will want to be locked into rigid transparency criteria to grant the market release of the best AI models. A kind of strategic ambiguity will prevail. The government will seek to preserve maximum maneuvering room, including the possibility of withdrawing access to a model without notice and discretionarily, or limiting its use cases.
This 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 permanently irrigated by large language models (LLMs), not having autonomous access to the best frontier models means depending on others for a growing part of our productivity, our industrial power, and our national security. If certain capabilities of the models cross a threshold (autonomy on long tasks, greater reliability, lower inference cost), adoption could undergo a brutal rupture, potentially triggering a rapid economic decoupling for countries that do not have it. No economic question, and likely no strategic one, today is more important than this: will we be subjected to the will of foreign powers to feed our society with mechanical intelligence, or will we be able to produce it ourselves?
When it comes to AI, France has many advantages, beginning with its nuclear fleet and the excellence of its researchers and engineers. It hosts the only European laboratory capable of producing large language models of a reasonable size, Mistral. Yet it is still far from having a frontier lab capable of delivering top-tier AI and following or leading frontier advances. On this front, the country essentially starts from zero.
If France and Europe have not managed to produce champions as powerful as Anthropic or OpenAI, it is due to structural reasons linked to capital dispersion, regulatory environments, and possibly cultural behaviors. It is surely urgent to reduce these frictions, but that will likely take much longer than the short window we have left to put in place a cutting-edge frontier AI laboratory: unifying the European capital market is a long-term challenge, as is simplifying national and European law. Any political will to get France back into the AI race can only come from a deliberate and concentrated State effort.
But what effort and what scale are we talking about? Most who call for sovereign AI today do not grasp the price. The political equation varies greatly with scale. Is it comparable to the cost of a new aircraft carrier, for example? Is it closer to the Messmer plan, which mobilized for a time more than 1% of French GDP to build the fleet of nuclear reactors that still underpins our energy base? Does it go even further?
In this note we propose an “Prometheus operation”: we compute the amounts to invest to create and sustain a frontier lab in France by 2029 (three years), then determine by what means to achieve it and derive the strategic conclusions that follow.
Create a frontier laboratory for France
We deliberately focus on a laboratory dedicated to large language models, because today it is the most mature and hence the most reproducible form of general AI. Exploring unproven paradigms like world models remains valuable and desirable, but as a complement to a project of this industrial scale (just as in the 1970s it was necessary in France to copy and deploy the American model of light-water nuclear reactors, independently of French research into alternative reactor designs).
What is a frontier laboratory?
The term “frontier” in AI refers more to a dynamic over time than to the instantaneous performance of a model: the duration, measured in expert human time, of tasks an agent can complete with a given reliability, has roughly doubled every seven months. Most benchmarks, quickly saturated, become less about fine-grained measurement of model performance and more about minimum entry conditions at the frontier. Access to the frontier cannot be thought of as a one-off purchase of a model. It requires a durable capacity to track and absorb the dynamics. At the frontier, the computing power needed to train a model doubles every 5.2 months since 2020. Over the same period, the training cost of frontier models doubles every seven months. Hardware and algorithmic efficiency gains are also rapid but do not fully offset the rising ambition: they mainly enable aiming for more capable models, longer to train, more agentive, and more demanding in inference.
Frontier AI laboratories rely on a fully integrated set of resources to finance subsequent model iterations and ensure large-scale deployment. Their advantage lies as much in the scientific quality of the technical teams as in ongoing access to capital, data, and computing infrastructure needed to push the frontier. Increasingly, they also rely on access to AI models themselves, which assist developers in producing next-generation models, or even run their own experiments to accelerate R&D automatically; this is the process known as recursive self-improvement and which could rapidly widen the gap between laboratories that have access to the best models (theirs) and their competitors.
Compute: the nerve of the war
The fundamental principle of AI today—and the sole reason why firms developing this technology have embarked on an enormous investment race—is the scale laws.
These are empirical laws according to which a model’s intelligence grows roughly in proportion to the logarithm of the computational power used, whether for training the model or later for its use.
Admittedly, this formula becomes increasingly expensive as more compute is deployed, but the promise is staggering: to achieve unlimited intelligence, vaster than the world’s greatest polymaths, capable of contributing to scientific and technical progress more than all Nobel laureates combined. Nothing guarantees that these laws will hold indefinitely. But they have held so far. It is this promise that makes the investment so significant: AI, becoming a super-intelligent entity capable of making leaps in all scientific and technical domains—and thus in weapons and sovereignty—requires sovereign mastery of this technology.
Winning the race is thus primarily a matter of compute, as the current geopolitics of AI confirms. The two frontier laboratories today, Anthropic and OpenAI, also hold the largest compute reserves. They command several gigawatts of power. And it is indeed compute power that makes the difference:
A common way to measure a frontier laboratory’s “lag” is to choose a set of benchmarks for comparison. None of them alone captures what the frontier truly is, and the resulting figure depends heavily on the chosen metric. Using the Artificial Analysis index (a useful, though incomplete, indicator of the frontier), one finds roughly a four-month gap between Fable 5 and Chinese laboratories. The real gap is almost certainly larger, as shown by applying the same reasoning to FrontierMath, a benchmark for mathematical research problems. Benchmarks like Artificial Analysis can saturate the top: they compress the gaps between the best models. This compression can give the impression of a modest lag, even as differences remain far more pronounced on the most challenging tasks (on ARC-AGI-2, for example, Chinese models still lag by about eight months). On the other hand, an aggregated average masks the actual shape of the frontier. Having four months of lag does not tell us which capabilities remain inaccessible today, nor which task classes are unlocked only by the best available model. At the frontier, the lead is not linear but cumulative: the best model trains, distills, and accelerates the next, captures the most profitable usages and the best talent, and sets the standard others chase. Then, many benchmarks do not account for the tokens consumed to respond. Mobilizing more compute at inference often significantly improves results. Frontier closed models are also tested for safety (red-teaming) or reliability before broad public deployment, which can delay public availability even when the internal capability within the lab already exists.
Given the performance of models developed by Chinese pure-play firms, with ostensibly more limited resources than the major American labs, some now argue that frontier models could be trained with far less compute (and cost) than the American standard. GLM 5.2, published on June 16, 2026 by Zhipu, rivals Claude Opus 4.8 on certain benchmarks, while Zhipu would have only a fraction of the compute capabilities of Anthropic (which is allocated, rather than having compute under its own direct control). Yet, on the one hand, the true financial and compute capacities of Chinese firms are hard to gauge. On the other hand, a portion of their model performance results from distilling American commercial models—using those models to generate data and training environments. To take the example of a DeepSeek or Zhipu and conclude that Europe could develop a frontier lab at low cost would be a mistake. Distillation, in a sense, is access to compute by proxy, where the frontier’s cost has been paid elsewhere. Moreover, a model like GLM 5.2 does not match Anthropic and OpenAI on other benchmark suites and appears overall less versatile and capable.
It is also true that a large share of the compute available to OpenAI and Anthropic is devoted to inference—i.e., the use of models by users—and not to training. In itself, one might be tempted to think that training could be achieved with far less compute, but that argument would be flawed: it is notably the data resulting from massive user interactions that feed the training of subsequent models. We must therefore distinguish training from inference in architecture, but not separate them strategically: for a frontier laboratory, compute is a strategic portfolio to be balanced across training, R&D, internal inference (RL, synthetic data, automated research), and client inference. Scaling effects apply to each link: the more a lab serves users, the more it learns to reduce the cost per token, to improve its kernels, routing, batching, and accelerator utilization rates; the more efficient its inference becomes, the more it can monetize intelligence, generate revenue, capture data and usage signals, and reinvest in the next cycle. Finally, inference enables test-time scaling—the improvement of performance through more computation during task resolution (see the scaling laws supra). Inference is therefore a key economic and technical engine of the frontier.
If France wishes to have its own frontier lab capable of delivering frontier AI in the long term, it needs an organization with scale and means approaching those of the major American labs.
What would be the price of such a project?
For simplicity, at this stage we abstract from existing vehicles capable of hosting and driving this project, and reason from first principles: how many GPUs, how much energy, and how many researchers would be needed to create this lab?
Compute power
By late 2025, OpenAI had about 1.9 GW and Anthropic about 1.4 GW; both labs were expected to reach around 5 to 6 GW each by the end of 2026. We propose targeting a compute load of about 12 GW in 2029. This would bring us to the frontier actors’ anticipated level, by moving quickly: 2 GW in year one, 7 GW in 2028, 12 GW in 2029. By comparison, recall that the Stargate program budgets $500 billion for 10 GW, and that Anthropic has reserved around 10 GW with Amazon, Google and Broadcom.
The first cost covers only the compute base: purchase and construction of capacity owned outright, temporary capacity rental to bridge the startup gap, and then annual operating costs of the owned capacity. Using the order-of-magnitude figures from Epoch AI for simplicity, we estimate roughly 38 Md$ per GW purchased, 8.5 Md$ per GW rented and per year, and 0.9 Md$ per GW owned per year in operating expenses (OpEx). The financing need for compute would therefore be about 161 Md$ in 2027, 219 Md$ in 2028, and 299 Md$ in 2029, i.e., around 678 Md$ cumulatively over three years. Of this total, the bulk corresponds to CapEx for construction and equipment, about 606 Md$, with the remainder covering rentals (60 Md$) and OpEx of owned capacity (13 Md$).
The associated electricity consumption in 2027 would be about 20 TWh per year, including PUE, for 2 GW of available compute. With the planned ramp, consumption would reach around 71 TWh in 2028, then 121 TWh in 2029, for 12 GW available. This massive consumption could, in principle, be fully supported by overbuilt French electricity production, even before the arrival of new nuclear capacity (which will be necessary in the longer term): by 2029, simply increasing the load factor of the existing fleet (about 71% in 2024, versus around 90% for the best operators worldwide, partly due to renewables’ modulation) could free up to 100 TWh per year, in addition to the roughly 90 TWh currently exported.
We also model a transitional compute rental of 1 to 3 GW per year during the period, at a cost of 8.5 Md$ per year per GW. This rented capacity would complement the owned capacity during ramp-up, and notably allow researchers to begin their work earlier in the first year.
In France, five sites are planned to host more than 700 MW each by 2030-2032. This timeline reflects four-to-five-year grid connection times for major sites, including those carried by the public sector. Securing 1 GW aggregated across these sites in the first year would be feasible by going beyond the current fast-track approach and imposing a site-by-site tranche logic on the most advanced projects. Examples like Colossus show that it is possible on a single site to add 300 MW in about seven months by combining a reusable industrial site, mobilized suppliers, accelerated grid connection, and broad regulatory tolerance. The objective of 12 GW by 2029 is an exceptional catch-up trajectory; the five fast-track sites should be pushed close to their target regime by 2029. Moreover, this trajectory implies expanding the portfolio beyond the existing major sites by combining site extensions, industrial conversions, and possibly a few European capacities controlled by French or European actors. It also requires reducing time-to-market by expediting grid connection work as much as possible, or clearing PTF (Propositions Techniques et Financières) queues, the connection commitments made with RTE. By shortening the time-to-market, the State would make these sites far more attractive to investors and developers.
Recruitment
Compute power is an essential baseline, but it is not enough to reach the frontier. Meta and xAI’s inability to compete with Anthropic and OpenAI, despite significant capacities of around 4 and 1.5 GW respectively, demonstrates the importance of the quality of researchers and the lab’s culture.
Although French researchers and engineers are often celebrated for their excellence, Europe still lacks the experience and thus the expertise in frontier-model training. To claim it, we must pay the price to attract researchers from the major American labs (which, in some cases, may involve bringing back European talents).
You do not need a large headcount: we propose a tight team of about 1,700 people, with Anthropic at 3,000 employees and OpenAI at 5,000 (for product and marketing scales we set aside for now). However, salaries must be high, especially at the top: we must mobilize around sixty elite researchers, drawn from the best competitors, for which we anticipate total annual compensation of about $3 billion. That is the market price: nine-figure salaries have become ordinary for poaching between labs. It is worth noting, even if we focus on financial means here, that money alone is not enough: the difficulties faced by xAI show that beyond salary levels, attracting and retaining the best researchers also depends on scientific culture and governance.
The organizational backbone would include, provisionally, 300 senior researchers and engineers (around $4 million per person), 400 specialists in distributed training, systems and infrastructure, and 300 people dedicated to data, post-training, and evaluations (roughly $2 million per person), plus 150 in R&D tooling, 150 in security and cyber defense, 250 in product, and 150 in legal, communications, HR, and support roles.
In total, the payroll would run a little under $7 billion in 2027, rising to about $8 billion in 2028 and $9 billion in 2029, i.e., roughly $24 billion cumulatively over three years. It would remain well below the compute cost: about 3.5% of the cumulative compute. Talent is thus far less expensive than GPUs, justifying paying them very well. Payroll grows more gradually than compute during the build-out phase, around 15% per year: the team remains tight, but salaries keep rising due to competition.
Total
On this basis, the total annual cost would be about $170 Md in 2027, $229 Md in 2028, then $310 Md in 2029, i.e., a cumulative roughly $710 Md over three years. The compute cost alone accounts for the overwhelming majority: around $678 Md cumulatively, i.e., 95% of the total, with the rest covering payroll and various operating costs (about $24 Md and $7 Md cumulatively).
The investment is therefore colossal: this effort would represent about 4.5% of French GDP in 2027, 6% in 2028, and 8% in 2029.
After three years: a self-sustaining trajectory and financial benefits
Although massive, the effort required for a frontier race would be bounded and time-limited, especially from the public sector’s perspective. If frontier attainment proves extremely difficult for an AI lab, such a breakthrough in France would require sustained and multifaceted state support for several years. If successful, private demand for the lab’s services and its attractiveness to international investors would create a self-sustaining trajectory. Private demand would grow as the project benefits from European players and other countries seeking to distance themselves from dependence on American and Chinese export controls. Once the startup phase is over, a frontier European lab, if well managed, should be able to stay at the frontier and likely widen the gap with competitors that have not reached it, without requiring further public funding.
Public and private investments in such an effort would, moreover, be particularly financially rewarding. Beyond the political and strategic value of the project, a frontier lab is a productive asset. The investment is not a loss: it can generate revenue if successful and retain substantial value even if it fails.
Most of the compute power of the major labs is allocated to inference, monetized through API access (per token), public and enterprise subscriptions, and increasingly through agentic products that automate large swaths of intellectual work. The leading labs’ revenues grow at doubling rates or more, reaching tens of billions of dollars per year. A French frontier lab would also enjoy a structural advantage—a largely sovereign European market. Ultimately, if successful, most annual costs could be funded by revenues, with the remainder backed by private investment.
Furthermore, even if the frontier lab failed to reach the frontier, France would still own multiple gigawatts of data-center capacity. These are durable, valuable assets. The compute could be rented (on-demand inference, sovereign cloud), deployed for the European research and startup ecosystem, or repurposed for other high-intensity workloads. It should be noted, however, that GPUs themselves depreciate rapidly, with a useful life of only a few years, which justifies the amortization built into the model. The real patrimonial value lies mainly in the footprint—the energy and land—but it is not the least difficult or least valuable piece to secure.
How to carry out this operation?
Given the magnitude of the resources required for the project, the architecture chosen to implement it is decisive. How should we organize the inputs for this Herculean effort? Should we pit several companies against each other or concentrate resources in a single one, and if so, which? And what role should the State play?
What vehicle around the State?
If one were to reduce the performance of models from a lab to a single equation, one might write, based on the observations above:
performance = compute × data × organization × brains
OpenAI, Anthropic, Meta, or xAI are notably ahead on compute. The organizational factor may be the bottleneck for some firms, such as the overly heavy hierarchy at Meta and, to a lesser extent, Google. By contrast, Chinese firms with compute limitations defend themselves on the brains factor, with brilliant researchers and top-tier engineering.
But once this equation is posed, how to maximize the product? Should we invest in a single champion, or several? Naturally there is tension between competition and concentration of resources. In the United States, competition has allowed multiple actors to emerge and contribute to the country’s progress: Google discovered Transformers, OpenAI discovered the scaling laws on inference with o1, Anthropic published the open-code agents paving a path toward self-improvement. The Chinese example is peculiar: the government maintains a fleet of data centers that grow in power. Neo-labs like Zhipu, Moonshot, or MiniMax then compete for grants and time-limited access authorizations; obtaining a grant is, of course, conditioned by past performance, which enables competition to generate new ideas, but disperses the effort.
However, for a State with limited resources, the scale laws constrain options: any organization that aspires to a 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. This effectively excludes France, alone or even in coalition, from funding multiple champions at once. The investment would therefore have to be focused on a single vehicle. This national champion logic is indispensable but would require governance carefully designed to avoid stifling innovation potential.
Two options emerge. One is to rely on an existing laboratory, with Mistral naturally appearing as the only credible European candidate, or to create a wholly new vehicle. The first option would make Mistral the project’s anchor, absorbing all resources allocated to the Prometheus operation and adopting the frontier race roadmap. The second would be to create a dedicated structure, possibly buying out Mistral’s researchers and assets if this form would enable faster execution, tighter governance, or better alignment with the program’s objectives. Mistral is currently valued at about 20 billion euros, which would represent only about an eighth of the current project’s annual cost.
Then comes the question of the State’s role.
Two conditions are clear. First, the project will never succeed without massive public commitment—financial first, but also political, diplomatic, and regulatory. Only the State, willing to deploy its full power and ability to concentrate resources, can make the Prometheus operation a national priority on the scale of the 1960s-70s French nuclear deterrence program. Second, the State should not add value to the management of the frontier lab itself; it could even be detrimental. It would be politically difficult to defend financing of this magnitude without any public oversight.
The core design would therefore split the project into two parts. On one side, the frontier laboratory proper: researchers with their own culture and scientific freedom, entirely dedicated to training models and governed only by an overarching roadmap to the frontier. On the other side, the rest—a massive infrastructure project delivering, in terms of quantities, costs, and timelines, compute to the lab, which the State would manage.
At the helm of the public segment would be a state-run program directorate tasked with project management for the State. Its role would be entirely facilitative: speeding up administrative procedures and implementing the numerous derogatory regimes indispensable for the project. A “Prometheus Law” equivalent to the “Notre-Dame Law” should be adopted in matters of land, electrical connections, environment, and labor law.
The Prometheus Law would also include a financing programming component setting a substantial public investment in the project for three years at a rate of 1.5% of GDP per year. Historical precedents from major national techno-industrial projects (the French nuclear deterrent, the Manhattan Project, the Apollo program) tend to show that this is the maximal credible scale for a very ambitious program designated by the State as a top national security priority. Essential to give the project credibility and attract French and foreign private investments to supplement it, these public funds would, if successful, be investments likely to yield substantial returns. In the event of success, France would possess an invaluable strategic asset within a very exclusive club of frontier powers (the United States, China, and France) and a means to act as a catalytic driver of European strategic autonomy. Investing 1.5% of GDP in public money per year for three years in such a project appears, under these conditions, more than justified. It is all the more true that, once the bet is won, no further public funding would be required. In terms of procedure, mobilizing such sums through the State would require deft use of European state aid law: the security of the nation would have to be invoked without réserve to keep the Commission at bay, possibly challenging it through a legal confrontation.
It should be noted that while this effort is substantial, it is proportionate to the project’s stakes and should be viewed against existing French budgetary choices. 1.5% of GDP corresponds to less than one-third of the current public deficit, which primarily finances the social security system, and to about 10% of pension expenditures. In such a context, adding 4.5% to the stock of France’s debt/GDP (117.5% of GDP in Q1 2026) would raise it by less than 4%, while decisively improving the country’s prospects for prosperity and sovereignty for the decades to come. It is not our role here to determine the exact budgetary choices to finance such a project if needed, but it should be emphasized that it is neither impossible nor disproportionate.
The frontier laboratory 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 meaningful minority stake, around 25%, with the remainder largely private. The ideal would be to diversify the participation of funds and large European industrial groups: not for capital alone, but for the direct interest in having a sovereign frontier model that cannot be unplugged. The founders and leaders of the lab would not necessarily have to be French; they just need to be first-rate and experienced in training frontier models. The State would, however, guarantee national control of the project, with standard corporate-law instruments: a special voting share giving the State a targeted veto on sensitive operations based on national security imperatives, localization of headquarters and non-delocalization of critical assets, and guarantees of model non-debranchability, etc. The multiple voting rights allowed in France since the Attractivité law of June 2024 would allow separating the economic and strategic aspects within the lab, with the State, BPI, and the founders holding voting-strong shares while foreign capital enters with high financial returns but low or zero voting rights.
The compute itself would be housed in separate structures, subsidiaries or dedicated joint ventures, whose sole objective would be to deliver the anticipated power to the lab. They would gather private investors from all origins and public investors from willing countries, with compensation calibrated to be genuinely attractive: given the capital-intensive nature of these vehicles, the return must be commensurate to attract the needed funding. There are precedents. American Poolside separated its lab from its infrastructure company, whose leadership includes data-center construction and operation veterans from major cloud players. Mistral structures its access to compute through joint ventures with the BPI, MGX, or Nvidia.
A final public, distinct structure would have the mission of putting the new nuclear “on steroids” so that energy does not become the bottleneck after 2030. This is not the subject of this note, but it is essential nonetheless.
Backed by public investment that lends credibility to the project, national savings would be massively mobilized for Prometheus through mandatory allocations in retirement savings plans and life insurance, a mechanism akin to a type of Tibi vehicle directing insurers and institutional investments, publicly traded in a PEA for the general public. The risk-sharing arrangement whereby the State would assume the first-loss tranche would, while accepting sovereign risk, attract large volumes of private capital. Allocating 2% per year of the national life-insurance stock toward Prometheus could match a public investment of 1.5% of GDP, though a complete, ambitious and well-thought-out financial arrangement would be required.
Given the scale of private capital needed, the project’s success would hinge on its ability to quickly demonstrate its ambition to private and international actors. The scale of public and private inputs from France, as well as its willingness to enact major regulatory exemptions to serve the project and recruit top international talent to lead it, would be essential. Early rapid results on models would be imperative to generate a virtuous circle.
Who should lead with France?
In these kinds of projects, a reflex to start from a European logic—without always proving its necessity—often leads to seeking an immediate geographic payoff (as in space), and thus to fragmentation of effort and, ultimately, to failure to scale. Making it a multinational European project would inevitably lead to governance that is plural, lacking clear leadership, full of varied veto rights, and likely unable to sustain the continuity and decisional will essential for success. That is why we start from French capabilities, other converging countries, and even non-European partners as needed and as their interests dictate.
France is indeed structurally the best positioned today to attempt to become the third frontier AI nation because, apart from the United States and China, it is the only power that combines four conditions:
- a sufficient economic critical mass (unlike the Netherlands, Israel, Switzerland, or Singapore);
- a domestic competency base in the field (unlike Germany);
- sufficient strategic autonomy from the United States to avoid being subject to intolerable pressures (unlike the United Kingdom, Japan, and Korea);
- credible access to the international AI talent pool (unlike Russia and India).
In addition to this, France benefits from excess nuclear electricity capacity (and decarbonized) and a tradition of major catch-up technological projects through civil and military nuclear programs, which would provide valuable experience and inspiration for such an effort.
Its success would require clear governance based on unmistakable French leadership, credibilized by a massive and enduring financial commitment from France, both public (through its dedicated programming law) and private. Other participating countries would benefit from access to compute, research and enterprise usage credits, shared standards, and the gradual integration of their industries into the value chain. Beyond that, they would gain a guarantee that, in the event of success, they would have permanent access to a frontier model for their needs and those of their businesses, under conditions comparable to those of French firms. Through a multilateral treaty to which they would be invited to become parties, partners would enjoy such commitment from France, in exchange for a significant financial commitment to fund dedicated computing capacity for the project. In addition to paying an entry ticket (for example 0.3% of GDP per year for 3 years) to finance the project, they would likely play a vital role in the first years by facilitating access to compute within their territories, as the amount of compute that can be built in France is limited in the short term by electricity availability, until the new nuclear capacity comes online to take over.
Determining precisely which partners might be interested is not the objective of this note. This is a diplomacy task to be designed meticulously. It should be noted, however, that if the French project appears credible, a number of middle powers could find it attractive as a hedging mechanism to limit dependence on American and/or Chinese frontier models. Japan and especially South Korea and Taiwan stand out as priority targets given their mastery of the semiconductor chain, though their strategic dependence on the United States might make them hesitant. The United Kingdom is also relevant for its AI ecosystem. Other European partners of Paris should be invited to join the project. Beyond the E5 (France, Germany, Italy, Poland and the United Kingdom), given their positions, the Netherlands, Nordic countries, and Belgium appear particularly likely to be interested. Beyond Europe and East Asia, the UAE or Canada could also be serious candidates.
It should be noted also that potential state partners could be pressured by their own national companies to support the French project to ensure guaranteed access to frontier AI models in case of success, especially if export controls tighten. In such a context, the Prometheus operation could leverage France’s solid reputation for reliability in diplomacy and strategy. Faced with the uncomfortable situation of being caught in a critical dependency on the United States and/or China, many foreign actors—public and private—would have a clear interest in the success of the French bet, even if it remains under national control.
The first three years
How would the project be practically run? We propose the following schedule for its first three years.
These forecasts are based on durations observed in the most recent large laboratories:
International opposition
Beyond the administrative sprint and internal political acceptability, one of the main difficulties would be securing GPU supply. Building the data centers themselves, structuring the project legally, and creating regulatory exemptions are within reach, but access to chips is not. No frontier lab trains frontier models today without Nvidia chips, with the exception of Google, which relies on its own Tensor Processing Units created for its internal needs. Even recent Chinese models are trained on Nvidia GPUs. In the short term, and as long as others do not challenge this quasi-monopoly, the project would therefore depend on Nvidia (hence the U.S. administration). Nothing prevents the latter from imposing export controls on Nvidia products to France (as some European countries have already experienced), and it is possible that Washington would be tempted to do so when faced with a frontier effort of this size.
Two options thus present themselves for France.
Or it still tries to buy Nvidia chips, counting on the market being too lucrative for the company to forgo, and would use its influence in Washington to avoid export restrictions; in the worst case, it would also persuade the Netherlands to threaten Nvidia and the U.S. government with retaliatory action by halting ASML lithography machines—indispensable to GPU production—making the Netherlands’ participation even more essential to the project.
Or we look to obtain compute from new producers elsewhere in the world. But today it is hard to find producers not already in the United States’ orbit, directly or indirectly, with the required scale. Even Huawei’s Ascend chips are not enough to meet Chinese labs’ consumption. Perhaps the situation will change soon; for example, South Korea has announced a plan to invest over $1 trillion in semiconductor plants.
As things stand, this is one of the project’s most serious risks. It is not insurmountable, since the United States would bear a heavy cost for overtly hegemonic export control on a rival ally, which could accelerate the emergence of Nvidia alternatives—especially in Asia. In the meantime, Washington could not prevent Prometheus from renting compute to third parties to train its models.
Strategic alternatives
Other paths are also risky
A cost so immense and an effort of such scale could discourage French and European policymakers, who might feel incapable of financing or driving it, or provoke too much opposition in the population. So what alternatives could France pursue?
The first option is a bet on open source: so far open models have trailed commercial models by a few months, and thus Europe could, in the end, rely on initial open models to feed AI, without depending on others. The bet is risky and involves a non-negligible cost. For running these models, energy and infrastructure are required, which themselves require large investments. Most importantly, nothing guarantees that the flow of open models will not dry up. Their developers—currently mainly Chinese—could suddenly decide to commercialize them. In that case, Europe would be as dependent on China as on the United States. These models could also face access restrictions similar to American models that are closed beyond a capacity threshold like “Mythos.” The press has recently highlighted high-level Chinese state efforts to restrict foreign access to the most advanced models, indicating such dynamics appear already underway. If these models were to catch up to the frontier, they could be described as a supply-chain risk by the American administration, leading to restrictions or even prohibitions on purchase, use, or hosting by American firms. By extraterritorial effects, compliance, or regulatory pressure, their adoption would become harder for European firms.
The second viable path contemplates a negotiated dependence on the Americans. The United States builds the models and Europe uses them. To avoid total subordination, some argue that the continent has levers in the value chain it can deploy in turn to pressure the United States; in other words, sovereignty would be found in mutual dependence rather than independence. It is true that production chains are dispersed globally and no one controls unilaterally all the components necessary for the development and diffusion of large language models. But don’t be fooled: the main European asset—the production of lithography machines for the semiconductor industry through ASML—has limited leverage in a long-term interdependence. It is an indispensable link in the chain, but with a long latency: if Europe blocked ASML exports, effects would be felt only after many months, as other actors could use existing stock. By contrast, blocking a frontier model is an instant weapon for the United States, as the Fable 5 example has shown. The ASML lever could be used in a sanctions scenario focused on infrastructure, as described above, but it is less effective for the final product and in ongoing operation. In a world where Europe’s economy relies entirely on the use of American models, can Europe really afford to go a few days without them? It is far from certain.
At best, European states can try to offset this disadvantage by conditioning American suppliers’ access to the European market on the physical presence of models within European data centers controlled by them. South Korea formalized a partnership between Shinsegae, a national conglomerate responsible for physical infrastructure, and the American startup Reflection AI, providing open-weight models and engineering to build a sovereign 250 MW site designed for Korean enterprises and administrations. This would avoid giving the United States an instant “kill switch” as they have today; but it would not prevent Washington from deciding, at will, to halt the deployment of future models in Europe. Moreover, the European Union has not proven its ability to coordinate decisively to respond to such pressures, as its structure invites “free-rider” behavior.
Finally, can we hope to exploit the US-Chinese rivalry to avoid depending on either by pitting them against each other? A risky game: the United States would be aware that Europe threatening to switch to Chinese models would immediately place itself in China’s hands, and vice versa, leaving little room for negotiation and the added risk of dependence on the currents of processes and technologies of private firms, making rapid switchovers costly and difficult.
Perhaps these are the solutions France and Europe will adopt. In that case, we will need at least to drop the claim of sovereignty and European independence and admit that we must manage our dependence as best as possible. Many European countries are already accustomed to this in their energy supply, and this will therefore be the natural path for them as well.
One of the risks of a ramp-up plan for compute is overcapacity. Today, Anthropic already operates at the scale of several gigawatts and is a reference in compute deployed. However, if the goal is not to depend on other powers for the development of AI capable of effectively assisting all jobs, these levels remain insufficient. A model like Fable 5 could automate only a fraction of tasks, with large variance across professions. In a recent hearing before the National Assembly’s Digital Dependencies inquiry committee, Arthur Mensch, founder and CEO of Mistral, noted that while the cost of AI usage for his own employees already represents 10% of his salary bill, extrapolating that to Europe’s wage mass in a few years would mean €1 trillion in value added per year captured by American or Chinese players, in the absence of European supply. The scaling laws suggest that there are still several orders of magnitude of compute needed to cover a meaningful share of economic needs. In other words, even with substantial efficiency gains, e.g., a factor of 100, a few gigawatts would not be enough to meet the entire demand.
Ultimately, it seems to us that the Prometheus operation represents the primary bet because its risks are commensurate with the technological and strategic tipping point that AI is creating before our eyes. When you compare costs and risks of this effort with the stakes of inaction, the path of least action could lead to France’s strategic insignificance in a world permanently dominated by silicon intelligence.
Why act now
Contrary to what is sometimes said, frontier AI is not destined to become banal within a few years, in the sense that capacities at the most advanced frontiers would simply become accessible to all at low cost without strategic cost to a nation. For some common uses (summarization, translation, office assistance, for example), lagging one generation may suffice: open or semi-open models already provide abundant and inexpensive capabilities. But the capabilities that matter for power (cyberdefense, autonomous agents capable of executing long tasks, acceleration of R&D, biological or military applications) are precisely at the frontier. They are also the ones whose access will become more and more controlled.
Developing frontier expertise yields returns beyond the trained model alone. The leaders of today’s major labs were, yesterday, researchers or engineers at the very edge of that frontier: Dario Amodei, Demis Hassabis, Ilya Sutskever, Arthur Mensch, among others. A French frontier lab would build the human, organizational, and industrial base allowing France to stay in the race in 2028, 2030, and well beyond.
Without a national program, a country gradually becomes dependent on external evaluations of the most advanced models, loses the ability to autonomously gauge real model performance, to identify vulnerabilities, and to assess systemic risks to its critical infrastructures, its security, or its strategic interests.
Frontier labs now explicitly seek to automate part of their own R&D: rapid AI research automation and its development cycle supports continuous improvement of frontier models. The earlier generations of models are used to train the next; in this scenario, being absent from the frontier does not cost linearly more: you lose access to the driver of acceleration itself.
Finally, inaction increases an already massive industrial dependence. The United States concentrates most of frontier compute performance today, as well as priority access to the supply chains feeding them. The more France delays, the more these resources—including compute centers, talent, and infrastructure—consolidate elsewhere.
The window opened by France’s position and the state of the field exists for a brief moment. In three years it will close for good, because the costs of catching up will become prohibitive: it is now that history knocks at the door.
Conclusion: a new nuclear moment for France
General de Gaulle recounted Khrushchev’s reaction in 1960 at Rambouillet when he learned of France’s success with the atomic bomb: “I understand your joy. (…) But you know, it’s very expensive.” The French President commented: “My story provoked no reaction from my interlocutors, except this: ‘Oh yes! It’s very expensive.’ ‘It’s very expensive,’ even for the Americans, even for the Russians. (…) But for us, facing these imperial aims, it is the price of independence.”
We believe the same holds for access to AI. The current state of our economies hardly reflects the place AI will occupy everywhere in the years to come. Of course, most countries cannot claim the ability to produce their own models or their own inference, just as most countries cannot produce their own energy. They will have to negotiate—by necessity or desire—with foreign powers to meet their needs. But we can see where energy dependence leads; just think of the difficulties many European states faced during the invasion of Ukraine. Dependence on mechanical intelligence will be at least as deep.
Since 1945, France chose to develop its own energy source through the nuclear sector. It was a massive and risky investment, but it now ensures our energy autonomy, makes us a net electricity exporter, and contributes to our place in the global power game. Now that access to AI is all at stake, a challenge of a similar scale presents itself. Since the bulk of frontier-lab costs lies in compute, and compute ultimately requires energy, one can view the rise of a French AI as a logical continuation of our nuclear endeavor. Even the nuclear program’s method can be inspirational. The 1945 ordinance creating the Atomic Energy Commission stated: “It has become clear that this body must be very close to the Government, and, so to speak, intertwined with it, yet endowed with great freedom of action. It must be very close to the Government because the fate or role of the country may be affected by developments in the branch of science to which it is devoted, and it is therefore 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 of its effectiveness.”
That is the philosophy we favor for building a frontier laboratory.
At the time when France chose, in the late 1950s, to go alone in a huge effort to acquire the atomic weapon and, more broadly, a complete independent nuclear deterrence, many French people and our main allies were convinced that France did not have the means and would fail. The “reasonable” voices of the time urged France to “manage its dependence” by negotiating with the United States on the model provided by the Nassau agreements with Britain. Had it followed this path, France would probably have had its bomb, but the maintenance of deterrence to this day would have depended on the goodwill of the United States. It took the will of General de Gaulle to recognize that the nuclear weapon had become so essential to national sovereignty that its autonomous mastery justified substantial sacrifices. Today, AI places us before a similar choice.
Launching or not the Prometheus operation will very likely be the most important decision the next President of the Republic will face next year. Like the nuclear weapon in de Gaulle’s era, it will determine whether France maintains the role it seeks in the world or gradually slides into strategic insignificance.