Artificial intelligence is changing everything. On our pages, readers find daily the doctrines, images and key data to understand the great metamorphosis. Subscribe to receive them by email.
The Prometheus plan, published in these pages on July 8, 2026, proposed a path for France to seriously enter the AI race within three years, thereby achieving strategic independence and all the economic benefits arising from mastery of frontier AI models.
The article, which sparked a debate urgent in an electoral year, did not detail its funding modalities precisely and asserted only the necessity of raising in three years several hundred billion euros of private capital, drawing on both national savings and international capital.
We develop here the details of this financing and its structuring, so that our readers have a concrete idea of what implementing this plan could look like. It is, as we will see, a completely realistic plan .
1 — The Financing Need
The Prometheus plan comprises two distinct needs:
- Fund the Prometheus laboratory itself, that is, the team of researchers and engineers who will develop and commercialize the models. It must be able to pay its staff, its operating costs, and especially the use of computing power.
- Finance the construction of 12 GW of computing centers by 2029 and be able to continue deploying more later without public investment.
We estimate the total economic need at around €620 billion over three years. Of this total, about €130 billion corresponds to the broad French public intervention and about €490 billion to private financing. Gross fundraisings by all participants would be higher, on the order of €700 billion, due to internal flows within the structure.
This amount must be mobilized while keeping three constraints and objectives in mind: a sovereignty constraint first, since France must maintain a durable guarantee over the laboratory, its models and certain strategic decisions; an obvious budget constraint, given the state of our public finances; and finally an economic constraint: a sufficient share of assets and yields must accrue to France, in light of the national effort it will undertake.
To achieve this, the plan relies mainly on private financing; public money is used only as needed to mobilize private capital.
2 — Private capital is available; how to attract it?
International capital is available to fund large computing-power projects: contracts signed by American hyperscalers bear witness to this. The contract signed in September 2025 between OpenAI and Oracle to build 4.5 GW of compute over five years for $300 billion is almost by itself a half-prometheus plan. The market can fund data centers through private debt, including GPUs themselves (CoreWeave raised in March 2026 $8.5 billion in credit facilities backed by these processors). This availability is not limited to the United States: even without Prometheus, France’s advantages, foremost its electricity surplus, already make it attractive. If we consider the AI Summit 2025 and the Choose France Summit 2026, this already represents about €150 billion of private compute investment announced in France for the coming years. The hurdle for Prometheus is thus not as high as one might think: it’s a threefold increase of private compute investment in France, not a hundredfold (these already-committed projects could be partly integrated into the plan to deliver the first GW).
If the capital exists, what is missing to bring it to finance such a project? A credible offtaker, i.e., a player capable of committing with the builders and operators of compute centers to buy their output over several years, which would then enable them to raise debt to finance construction. The hyperscalers and major American laboratories have the financial credibility to be this offtaker themselves. It would not be the case for the newly created Prometheus laboratory, which in 2027 would have neither revenues nor track record. It would therefore be up to the State to create an entity capable of serving initially as an offtaker vis-à-vis infrastructure companies, then sublet the compute to the laboratory. If Prometheus reaches the frontier, it could progressively take over this function itself.
Twelve gigawatts of compute require several million next-generation chips, today supplied mainly by Nvidia and subject to American export controls. Prometheus should negotiate multi-year allocation agreements with the major suppliers from the outset. The risk of restrictions cannot be entirely ruled out, but it can be mitigated by the scale of the order, by the participation of several allied states, and also, paradoxically, by the involvement of American investors and suppliers in the project, which would raise the economic and political cost of a future blockage by the American administration.
3 – A three-tier structure
To implement the Prometheus plan along these lines, we propose a three-tier system.
Tier 1: The Prometheus Laboratory
For the laboratory itself, national control must be permanent, but with no interference in its operational conduct. It would be a simplified joint-stock company (SAS), functioning like any private AI laboratory. The State would own about 25% of the capital at the outset, invested under the same financial conditions as private investors, with the difference that it would hold a specific share giving it veto rights on sovereignty-related decisions: change of control, location of models and headquarters, disposition of critical assets and guaranteed access for France and partners. Once the laboratory is created and the road map setting the three-year frontier target is fixed, the State would not intervene in scientific policy, recruitment or commercial strategy. To reassure private investors, the State would announce at the laboratory’s creation a set of precise rules governing its interventions, in the spirit of central banks’ forward guidance.
To obtain the resources needed for its ambition, the laboratory must raise €110–135 billion in equity over three years. The State would contribute about €23 billion during the first two rounds to hold roughly 25% of the capital after each round, then a little less after the third round, in which it would not subscribe. The remainder would come from private investors, during three rounds conditioned on achieving technical and financial milestones. They would target €20, €42 and €72 billion of raised capital, for post-money valuations of €40–50, €120–160 and €240–300 billion respectively. This progression aims to preserve the attractiveness of successive rounds while leaving the team with about a quarter of the capital after the third round. The compensation budget would reach €24 billion over three years, including about €12 billion in salaries and charges and €12 billion in stock options valued upon grant, with no disbursement by the company. The success of these fundraisings would come primarily from the founders’ quality, assembled around world-class researchers, and from public commitments ensuring compute availability for the laboratory as it grows. These founders, recruited by the State at launch from France and Europe’s top researchers in American frontier laboratories, would retain control of the company through multiple-vote shares. Preliminary commitments to purchase inference by major European firms would be welcome to help establish the initial economic value.
Tier 2: Compute France
The other two tiers aim to supply the laboratory with its computing power. We propose two distinct tiers, the first of which would be publicly controlled to a majority. It would be a standard common-law public corporation, Compute France, tasked with building and owning about one-third of the program’s compute capacity, i.e., a size compatible with the amount of national savings available to finance it. This entity has a double interest: first, it provides a strategic asset directly controlled by France; second, it combines French savings with a majority public equity backing to obtain a cost of capital well below market, without the State guaranteeing its debt in the base structure.
The State would contribute to Compute France about €28 billion of equity over three years and would own 65% of the capital, alongside private investors of equal standing. The financing would be completed by bonds and loans, notably subscribed by insurers, the Caisse des dépôts, and institutional investors.
However, this is not about diverting a portion of the €2,160 billion life-insurance stock: only a portion of this savings is reallocated each year, mainly through net new money and the refinancing of maturing bonds, i.e., a cash flow well under €100 billion per year. The “euros” fund cannot directly invest in equities, and prudential rules penalize non-listed assets. The instrument would rely mainly on senior secured bonds of Compute France, without State guarantees, enhanced by subordinate tranches absorbing the first losses and by export-credit agency covers. A reserve guarantee capped at €15 billion, partial and remunerated at market rates, would only be allowed if emissions could not find enough buyers. These bonds would be complemented by unit-linked supports, PER-eligible investments, very long-term loans from the Savings Fund of the Caisse des dépôts and an ELTIF fund of equity eligible for the PEA, without requiring listing. The mobilization would rely on the attractiveness of return and risk, the adaptation of prudential treatment for infrastructure investments, and eligibility for existing savings supports. Market-wide agreements would coordinate the initial issuances, without earmarking to a single issuer or default allocation. Public investment would only occur after the laboratory’s first technical milestone. A target of around €40 billion per year in French private capital mobilized, i.e., €120 billion over three years, thus looks feasible.
Compute France would lease its compute to the laboratory at market price determined at contract signing, with a dollar-denominated rent for the contract’s duration and a partial deferral during the three sprint years toward the frontier. This initial fixation protects the laboratory against a potential surge in compute rental prices later. A portion of the rent would not be paid immediately and would become Compute France’s debt to the laboratory. This mechanism would sharply reduce the laboratory’s cash burn during the sprint without constituting a subsidy.
Note that the compute power thus acquired by Compute France would include an enclave of about 300 MW for defense uses.
Compute France’s dollar purchases would be partially offset by its dollar lease revenues and its funding tied to export-credit agencies. The balance could be covered by forward purchases through the Bank of France, at an estimated cost of €1–1.5 billion, included in the construction cost.
Tier 3: Market-grade Compute Power
For the remainder of the compute requirement, we proceed as is customary for funding this kind of project. Private companies build and operate the compute centers and finance them through the usual mix of debt and equity. The Prometheus laboratory would have priority access to this compute power if its research succeeds; if not, the installed capacity could be leased elsewhere.
No French public capital is required for their construction: only France and its partners’ capital would guarantee the sole public element essential, a long-term contract guaranteeing the purchase of capacity (an offtake).
4 – The International Agency
The delicate point remains the identity of this offtaker, which, as noted, cannot be the Prometheus laboratory in the early stages of the operation. It cannot be the French State itself: if the State directly guaranteed all capacity contracts, this commitment would be consolidated into French public debt, with the consequences one can imagine for sovereign borrowing costs.
The proposed solution is to create by treaty an international organization, the Agency, in which France would contribute about 40% of the financing, with voting rights capped at 35%. The Agency would have decision-making bodies, its own capital and balance sheet. The treaty would set its essential parameters, including the capacities reserved for the laboratory and their pricing conditions; its council would decide how operations are run, site and operator selection, procurement procedures and treasury management. The most sensitive decisions would require 85% of votes: France could thus block changes that protect the laboratory, without being able to decide alone. This autonomy, modeled after the European Stability Mechanism, would aim to avoid consolidating the Agency’s commitments into national public debt, to be confirmed before its creation by Eurostat via an ex-ante opinion. Intelsat and Eutelsat also provide precedents of treaty-based organizations that bought, owned and leased capacities for their members. The Agency would sign capacity-purchasing contracts with operators, then sublease the compute to the laboratory at market price set at signing and for the duration of the contracts.
It is to participate in this agency rather than to finance the project itself that a coalition of partners gathered by France proves extremely useful. It remains to recruit other partners. Two counterparty options could be offered in exchange for their participation in the funding:
- An irrevocable guarantee of access to Prometheus models (under the same conditions as French firms, while the U.S. administration has recently shown it can cut off access to top U.S. models);
- Rights to the compute capacity built by the program (excluding the portion allocated to Compute France), pro rata to their contribution. A small portion would be usable from the start of the program and during the frontier sprint, notably for their public and research needs, but most would be delivered later, between 2030 and 2034. If Prometheus reaches the frontier, these rights could largely be converted, on favorable terms, into inference rights for its models; a portion would also be reserved for companies established in each partner state, pro rata to their contribution. If Prometheus fails, the partners would recover their share of the market compute and could use or lease it to improve their bargaining position vis-à-vis U.S. laboratories.
The envisaged ticket for each partner is 0.3% of GDP per year for three years. The set of partners should ideally represent around €7,700 billion of GDP (or less if the EIB participates in the capital). The creation of this coalition would not be an absolute prerequisite. With partners representing €4,000–5,000 billion of GDP at launch, the market tranche would be reduced to 5–6 GW, part of the gap offset by Compute France and transitional leases, at the cost of an estimated one- or two-quarter delay. Below that, a “France-only” variant with 8 GW in two equal tranches would constitute a retreat, within the same 1.5% of GDP per year three-year public funding envelope, but with a Maastricht debt cost. Given the growing demand for compute in the coming years, even if the Prometheus laboratory fails, partners would be able to recover most of their investment, or more, through leasing the capacities corresponding to their share: in our simulations, the recovered value represents 95–113% of the initial ticket depending on price and compute-use assumptions. The Prometheus wager would thus be virtually free for France’s partners.
5 – The Cost for the State
The public financing envelope for France for the frontier sprint period (2027–2029) would amount to around €130 billion, excluding interest on additional public debt raised to fund it. But these €130 billion are not a subsidy to a lab: they are largely investments and include equity stakes, asset acquisitions, pre-purchases of compute capacity, budgetary expenditures and reserves.
The interest charge would run about €0.8, €2.5 and €4.2 billion over the three sprint years, then around €5.5 billion per year. It would be covered in case of success by the dividends of Compute France, the taxes generated by the program and the discount on prepaid capacity. In case of failure, the remaining annual cost would be about €2.5 billion per year.
The additional OAT issuances required by the Prometheus plan would represent an increase of roughly 15% relative to the baseline, limited to three years. Their impact on rates would be mitigated by the plan’s modest deficit impact, the assets acquired in return, and the cap on losses. Market sentiment would depend on the overall trajectory of French public finances.
Note that about two-thirds of the expenditures would be Maastricht-neutral: they would count as a financial operation for Eurostat, i.e., the acquisition of an asset (equity in the laboratory and Compute France and participation in the Agency) or a receivable (payment for compute capacity delivered to participating states after 2029). Including the interest charge, the Maastricht-budget impact would be around €15 billion per year, i.e., about 0.5% of GDP, pending Eurostat’s confirmation of the treatment of Compute France and the Agency.
An option would be to finance €15–20 billion of the roughly €60 billion of public equity stakes by selling other state holdings that are no longer deemed strategic. This redeployment would reduce the need for new public financing accordingly.
6 – The State Aid Question
Opening a Commission inquiry or facing a challenge before the EU Court of Justice on the grounds of prohibited state aid would have deleterious effects on the borrowing costs of the actors involved (illegal aid that must be repaid with interest, a red-flag for lenders), which would weaken the project. We therefore maximize situations that fit the following:
- In the absence of state aid, wherever the public sector acts like a normal market participant and under market conditions;
- The existence of state aid that falls within one of the EU’s exceptions allowing its legality;
- Financing under the scope of essential State security interests as provided by Article 346 TFEU: the EU internal-market rules would not apply. The CJUE jurisprudence, however, is restrictive in its interpretation.
This is why the State increases its stake in the laboratory on market terms, why the deferred payments by the laboratory to Compute France would simply be a supplier credit with interest, why the International Agency would contract at market prices with infrastructure companies and the laboratory; or why the portion of compute allocated to defense would be funded by dedicated credits enacted in the military programming law.
Compute France’s financing would be built without State guarantees on its debt: equity contributions would be made on the same terms as private investors, savings-fund loans would be at reference rates, etc. The partial reserve guarantee, remunerated at market premium, would be designed not to constitute aid and could be activated without EU Commission pre-approval, though it would be pre-notified as a precaution. The path for major European-interest projects would remain available, with European partners, for measures that require it, including if the deferred payments were deemed to constitute aid despite their market-rate remuneration.
Two further points of a similar nature are added. First, the EU-law justification for the special State action in the laboratory so that it is not considered a capital movement obstacle. In line with the Court of Justice’s jurisprudence, the national security justification, real (guarantee of non-disconnectability, sensitive and dual-use capacities, localization of critical assets) and proportionate, would justify this aspect of the operation. Second, the status of Compute France: its standard corporate form, early private-sector shareholders, its debt without State guarantees, its market rents, and the risks borne on its own balance sheet would aim to place it outside public administrations from its inception. This classification should be confirmed by an ex ante Eurostat opinion.
7 – What Happens If the Laboratory Fails?
Even if the frontier laboratory’s success is credible and, we believe, likely with a project carried out to match its ambition, such an effort cannot be conceived without a precise exit framework set at launch to limit the cost of a possible failure for public finances. The “Prometheus Law” should include milestones for project review with objective criteria to decide whether to continue or terminate. An independent board could be formed to determine whether each milestone has been reached.
We propose two main exit milestones:
- At 12 months from the project’s launch, failure would be evidenced by the laboratory’s inability to raise the next round or to produce a first promising model;
- At 30 months from the launch, failure would be evidenced by the laboratory’s inability to produce a model with less than six months’ delay compared with the world frontier, despite the computing capacity made available to it.
In case of failure at one of these milestones, a pivot would be immediately undertaken, according to a pre-established procedure, abandoning the public efforts for the laboratory and reorienting the built compute, while preserving the GW of Compute France and leasing or reserving the remainder for sovereign use. The pursuit of further subsidies would be absolutely ruled out.
In the scenario (the most likely) where global demand for compute remains very strong in 2029–2030, the public dimension and the access rights to market compute would retain considerable value, regardless of the laboratory Sprint toward the frontier. If we assume within the 12 GW targeted by the program about 9 GW implanted in France and 3 GW in partner countries, France would have a substantial strategic asset on its soil. The net total loss to the State in case of failure would be limited to the equity invested in the laboratory and certain non-recoverable costs, around 1% of GDP. Despite this economic loss, France would be far better positioned strategically than if it continued along current trends. It would possess 9 GW of compute capacity on its soil, including 4 GW held by Compute France, under majority public control, a major lever for negotiating access to the models from American or Chinese laboratories.
An global collapse in demand for compute, in the context of a renewed “AI winter,” would be a different scenario: net losses could then reach 2–4% of GDP. The state’s maximum exposure, including sums already paid and contingent commitments, would remain around €160 billion. The roughly €130 billion for the sprint could be complemented by up to €16 billion in French equity callable by the Agency and up to €15 billion under the reserve guarantee, if that guarantee had been activated.
8 – After 2029: Exiting Public Financing
What has been developed so far constitutes an ignition program; the €130 billion of public money is spent only over three years, and no further public funding for the laboratory is envisaged. Yet we must ensure that it can stand on its own afterwards.
If Prometheus reaches the frontier, it will have sufficient revenues and a valuation to finance itself directly. It will become its own offtaker to build further compute capacity as needs grow according to scaling laws. Compute France could likewise fund its renewal and even expansion through rents, like an infrastructure real estate company.
To remain at the frontier, Prometheus will need to scale from 12 GW in 2029 to 18 GW from 2030 and then 24 GW in 2031 — order-of-magnitude targets to stay competitive with international rivals. But building such capacities takes about three years; we must therefore prepare these 12 additional GW starting in 2027, before knowing whether the laboratory will indeed reach the frontier. To protect these capacities against a rental spike, options would be signed during the first six months of the program for 2030 and 2031 tranches, and then between the sixth and twelfth months for an additional tranche deliverable in 2032. With a premium, the exercise price would be set according to a formula based on equipment, site, energy and capital costs, with the planned indexations, rather than the current compute rental price. Firm commitments would remain conditioned on program milestones. They would be financed by the private sector through the same mechanisms as the initial GW, via Agency purchase commitments, but without new public contributions from France or its partners, save for a moderately higher cost of capital. If successful, the entire new capacity would be made available to Prometheus; if not, the already-advanced capacities that cannot be cancelled would be made available to partner states for their use or rental. This extension avoids Prometheus from peaking after 2029 without increasing the public financial exposure planned by the plan.
9 – What Are We Waiting For?
The Prometheus plan at first glance appears extremely costly, but it can be credibly financed, both in its public dimension—requiring a substantial but time-bounded, modestly deficit-creating effort with limited risk of losses—and in its private dimension, where the plan aligns with the scale of today’s investments by involved financial actors and with their expectations, by offering a structure consistent with their usual practices. It also ensures the frontier-laboratory’s durable presence at the frontier in case of success, without further public financial support after three years.
The expected gains are clearly positive for the State and for the French economy as a whole, but also for international partners involved. Not only would the plan equip France and Europe with an invaluable strategic and economic asset, but even its failure to reach the frontier would leave France and its partners in 2029 with a substantial compute capacity, a far more favorable position than the current trajectory to pursue an alternative strategy, whether open models or building a dependency on American models.
One question remains: what are we waiting for?