All summer long, the Grand Continent will keep moving. Every day, we will bring you ideas you won’t find anywhere else (but which will be everywhere come fall), texts that are hard to find or refreshing with our Sundays. To receive them directly in your mailbox and support this momentum, consider subscribing to the review.
The “Prometheus Operation” has the rare merit of quantifying the cost of developing a sovereign French AI. This plan, which would commit France to a multi‑generational project, rests on two bets: on the one hand, that the race to the most powerful model will remain decisive, and on the other, that Washington will give France the time to buy the chips necessary to close the gap. An alternative scenario could, however, emerge: a banalized, commoditized AI, where one would only need to buy models from an ever-broader and more accessible market, akin to the Chinese model Kimi 3. No one can decide this today. Yet the components of such an effort do not share the same horizon: chips are bought in eighteen months, but energy, connected land, training skills, and decision‑making institutions require five to ten years. We must therefore establish an independence strategy based on slow assets that, tomorrow, will give real power to choose. This article therefore proposes a strategy indifferent to the evolution of technology: invest now in what will be useful in all scenarios, map the most critical risks and bring solutions, build the economic and geopolitical levers that will make these solutions credible, prepare “the frontier effort” without blindly committing, and finally equip ourselves with governance tools to take the right decisions at the right moment.
A debate trapped by technological uncertainty
The article published on July 8, 2026 by Le Grand Continent, “Prometheus Operation: France can win the AI race. What would it cost?”, has performed a service to the French debate that few texts have achieved. While most calls for sovereign AI stop at pious wishes, it offers a coherent, quantified strategy that takes on its risks and its limits. Real discussion occurs only with those who set their assumptions on the table, and that is precisely what they did. The present response is addressed to them as much as to the public debate, in the same spirit: our assumptions are on the table, and any nuances we may have, as we will see, concern less the ambition, French and European sovereignty than the order of steps.
Indeed, the French and European debate on sovereign AI has organized itself around two apparently irreconcilable poles. On one side, a sophisticated fatalism: Europe would not have the capital, the talent, or the industrial apparatus needed to catch up technologically. Its economic rationality would thus compel it to buy the best American models and abandon any autonomy ambitions deemed ruinous. On the other, a Promethean stance whose note expresses the most advanced version: $700 billion over three years, 12 gigawatts of compute power by 2029, and a “Prometheus Act” granting exemptions to lift France to the frontier‑level of power.)
These opposing positions are founded on contradictory answers to the same question, rarely stated explicitly: at what pace will models continue to progress?
There is a first possible future: gains become marginal, the gap between the best models and their pursuers narrows, and artificial intelligence gradually becomes a commoditized product, like electricity, chosen by price. Let us call this the commoditization scenario. In this world, the frontier race would be a bottomless pit with no reward, and the fatalist camp would ultimately prevail.
In a second possible scenario, capabilities continue to advance, the best models train the next ones and automate research itself. Technological evolution becomes exponential and the gap widens rather than narrows. Let us call this the runaway scenario. In this world, frontier‑level power becomes a strategic asset akin to nuclear capability, and Promethean thinking would be vindicated on a strategic plane.
A question without a definitive answer
Today, no one knows which scenario will prevail. Proponents of the incremental reading point to the narrowing of gaps on most public benchmarks. The latest wave of cheaper models, including Muse Spark 1.1 launched by Meta on July 9 at a fifth of the price of high‑end models, is the most recent example. Proponents of this thesis also advance deep arguments. First, a given capacity level replicates at a decreasing cost: distillation and compression of models allow achieving, with a fraction of the initial compute, most of yesterday’s performance. Second, barriers to entry shrink in this commoditized segment due to training recipes circulating and open architectures abundant, apart from energy and chips. Finally, the paradigm of current models could hit its limits: Yann LeCun has long argued that large language models are a dead end on the road to artificial general intelligence, and that an architectural rupture will be necessary, which would reset the counters of the race to zero.
Optimistic technologists respond that these benchmarks do not measure what truly matters. Indeed, these benchmarks saturate. They become optimization targets and stop measuring. That is Goodhart’s law, and the best models now reach the limits of what those tests can assess. This is also the finding at the first Forum of frontier AI experts organized by the European Commission in April 2026. They note that these benchmarks do not capture long tasks entrusted to agents, nor leaps of capability like Mythos. They add that client preference for the best models is clearly reflected in revenue figures. This finding is corroborated by the Forum of Experts of the European AI Office, which indicates that industry purchasing behaviors already reveal a strong preference for frontier systems over their immediate pursuers. The gap between Anthropic’s revenue and Mistral’s has indeed widened rather than narrowed. They also observe that the plateau predicted by the benchmarks does not yet appear in data, that end‑to‑end tasks run by agents take about twice as long every seven months according to METR Institute measures, and that entry barriers—both in compute and in exceptional talent—are among the highest in industrial history.
The current situation can be summarized by a paradox that dispels part of the controversy: efficiency gains make a given capacity level cheaper year after year (what Mythos cost yesterday will cost next to nothing tomorrow), but they do not make the frontier itself cheaper, whose cost continues to rise. Both camps can thus be right at the same time: rapid commoditization of yesterday’s capabilities and the continuing escalation of those of tomorrow. The question is whether tomorrow’s capabilities will justify their price and necessity. This disagreement runs through laboratories as well as investors.
Putting in place a method to guarantee the robustness of the European strategy
From this follows a methodological consequence: faced with deep uncertainty, a rational strategy does not necessarily seek the optimal decision in a predefined scenario; it combines robust, reversible, and complementary measures to avoid a single unfavorable scenario from causing a catastrophic collective loss. In decision theory, this criterion has a name: the minimization of maximum regret. The best strategy is not the one that yields the most in one world, but the one that is not catastrophic in any scenario. It consists of buying options rather than taking irreversible commitments, and converting the option into a commitment only once the uncertainty is resolved. It is to this criterion, robustness to both scenarios, that we will submit Prometheus, the frontier fatalism of the commoditization scenario, and our own proposal.
This reading of the moment that Europe is undergoing now has gained institutional support. The Expert Forum convened by the European AI Office, which brings together more than a hundred leading AI specialists, arrives at two conclusions. First, 2026 will be decisive for Europe, which must then formulate and adopt a coherent strategy at the level of the Union and the Member States, taking into account the pace of capability trajectories, markets, and infrastructure, and their implications for European sovereignty. Second, in both scenarios, Europe must strengthen its sovereign capacity to “access frontier models, choose them, control them, and exploit them,” which presupposes a clear view of the levers it has or could put in place at each level of the technology stack. Neither of these two tasks has yet been achieved, and the coming months will tell whether Brussels and the Member States seize them or let them dissolve in another volley of press releases and announcements. The proposal we sketch here aims to respond to both conclusions of the Commission experts: a strategy designed to adapt to the two scenarios and decisively answer, at the right moment, the frontier question, as well as a mapping of dependencies and levers that experts have called for. Let us now turn to the Prometheus proposal.
The Prometheus proposal rightly highlights the risks of European dependency and the limits of fatalistic inaction
The geo-economic diagnosis drawn from the Prometheus proposal regarding the risk associated with AI dependency is correct.
American restrictions, introduced in spring 2026 around Anthropic’s most advanced models, and then partially eased at the end of June, have brought frontier‑model access into a regime similar to that of semiconductors: revocable access without notice, deliberate strategic ambiguity, and administrative discretion. The most plausible reading of this episode is, in fact, worrying: everything indicates that a security alert, probably related to the model’s hacking capabilities, led the administration to want to revoke access for all users, including Americans. Export controls would have been the only legal instrument to achieve this. Europe thus became a collateral victim. Beyond the scenario of targeted coercion, this means that European access to cutting‑edge capabilities is now suspended by the calculation of American domestic risk, with no behavior, no commercial concession, and no diplomatic goodwill able to secure it.
We must also complete this diagnosis by highlighting a mechanism that makes the threat structural rather than merely political: the rarity of compute. Providing a top‑tier model is not the same as selling software, whose copies cost nothing. Each response from a large model consumes tokens, the unit of account for AI consumption (the equivalent of kilowatt hours for electricity), and behind each token hides real compute power. Large laboratories are also chronically short of capacity and must constantly arbitrate among their customers, their consumer subscriptions, and their own research needs. In a world of rationed tokens, foreign customer access becomes the natural adjustment variable for purely economic reasons, before any political intention. The observed allocation hierarchy among leading labs confirms this: the power available first serves American government contracts, then enterprises, and finally the general public, as demonstrated by the differentiated service interruptions during peak demand. Restrictions on foreign client access did not wait for deliberate policy by laboratories to occur. This is precisely what makes the mechanism structural rather than intentional. Anton Leicht had formulated it before the spring episode: AI will not create abundance, but scarcity.
The Prometheus proposal’s analysis of the asymmetry of levers between the Union and the United States is also accurate. Cutting access to a model is instantaneous, whereas blocking ASML exports, the Dutch company that makes essential machines for advanced chip production, is a slow lever whose effects are felt only once stocks are exhausted. We must also consider the risk of retaliation in other areas: the United States could respond to ASML restrictions with trade restrictions or by withdrawing security guarantees for the continent and Ukraine. The European lever exists, but it is insufficient.
Finally, the Prometheus proposal’s warning against a naïve reading of Chinese labs is warranted. Their rapid catch‑up is largely due to distillation, a technique that involves interrogating a leading model en masse to train one’s own model from its responses, i.e., to benefit from compute power funded by others. This does not establish that the frontier would be cheap for a new European entrant. We should draw another consequence not mentioned in the note: American labs and the government now have every reason to close the distillation door, tightening user identity checks and restricting access. The follower route that allows catching up leaders at low cost by piggybacking on their models is closing, which makes independent training capability even more valuable.
The three uncertainties tied to the Prometheus project
The Prometheus project analysis highlights three elements worth discussion.
The first relates to the schedule: the project sets its priorities based on three years of ramping up a continuous stream of U.S. chip deliveries, while Washington’s visibility could prompt a halt. The mechanism is therefore most vulnerable at the moment it is most exposed. It is hard to defend simultaneously that commercial interdependence is powerless to protect access to models, yet will secure a flow of components more strategically important. Two objections to this scenario, raised by Prometheus’ authors, deserve serious consideration. The first is this: until now, Washington has preferred to sell; the current administration has loosened some inherited restrictions and Nvidia effectively pushes for market openness. Yet a discretionary policy remains discretionary, and the risk of it changing is the main risk. The second objection is this: a sovereign effort of this kind should be protected during its construction time, and the United States would likely not take it seriously in its early years. We believe this argument could apply to a politically discreet program, but it does not fit with a plan that claims 12 gigawatts, a derogatory law, and a multilateral treaty, i.e., maximum publicity.
The second element concerns the very economics of catch‑up. The argument is this: capital buys catch‑up, but slowly, while the frontier cost doubles every seven months; in three years, the target will have changed by orders of magnitude. Recent examples abound, with Meta giving the most striking illustration. The company carried out a monumental financial effort that yielded only a repositioning toward mid‑range in terms of value for money: performant and profitable, but far from the frontier. Three years of some of the world’s largest investments have earned a solid place in the group, but far from the lead. The trajectories of xAI and Grok 4.5 confirm this rule: to reach the level of the previous generation of competitors, xAI had to devote three years of development, partner with SpaceX, and acquire key publishers like Cursor. If the appeal of a national mission is a powerful lever for state structures, as with the Manhattan Project or early-stage Atomic Energy Commission, Prometheus’ setup adds organizational drag that the note itself identifies as its main limiting factor. Its three-year horizon corresponds to the period when overcapitalized projects regularly fail for lack of learning to arbitrate under resource constraints.
The third element is the moving target. Aiming for 12 GW by 2029 is equivalent to targeting the world leaders’ level by the end of 2026. At the current expansion rate, the frontier will be well above this level by the plan’s deadline. The EU Frontier Experts Forum describes a global compute capacity that doubles roughly every seven months; the largest projected sites will exceed 3 GW on their own by 2027. A 12 GW target for 2029 thus targets the leaders’ present but probably not their future. The note itself concedes that these few gigawatts will not be enough to meet global demand. The project is therefore under‑sized technologically, while being overwhelming financially, requiring between 4.5% and 8% of GDP per year. By comparison, the 1974 Messmer plan, which enabled France to build its nuclear park, mobilized only about 1% of GDP to deploy a proven, not revocable, large‑scale technology. Add to this another uncertainty: according to the Commission’s consulted experts, no durable business model has yet been demonstrated at the frontier. Some recent generations of models are profitable, but leader revenues are still accompanied by massive capital raises and persistent operating losses. Betting 700 billion on a position whose commercial profitability remains unproven, is adding a bet to a bet. In the commoditization scenario, Prometheus would invest hundreds of billions to conquer a technological frontier whose market value collapses. In the runaway scenario, the project risks being hostage to the supplier of the chips at the most critical moment of its construction.
Recent July‑mid news gives a concrete face to this uncertainty. Moonshot AI has indeed just released Kimi K3, a 2.8‑trillion‑parameter model whose weights are expected to be freely downloadable by the end of the month. Initial independent evaluations place it on par with several recent closed systems, although it remains behind the best in certain tasks. Its forthcoming open availability and competitive price of access nevertheless greatly increase the probability of a commoditization scenario, i.e., a quasi‑monopoly that can reproduce, broadly accessible, and that exerts sustained pressure on the prices of proprietary models.
This opening can be understood as a strategy of industrial power: shifting competition from the proprietary model to the ecosystem, accelerating global adoption of Chinese architectures, and compressing the revenues that fund the compute race for U.S. laboratories. All else equal, the situation resembles that of other European industries that faced the same subsidized‑price competition, triggering a well‑identified geoeconomic spiral: abundant supply, backed by strategic capital, can render competing production economically nonviable, even before a dominant position is achieved.
Kimi K3 thus reveals a third risk for Prometheus. After the risk of technological movement of the frontier and the geopolitical risk of losing access to the chips needed for catch‑up, there is the commercial risk of succeeding too late in a market whose prices have collapsed. The financial closure of Prometheus thus presupposes a gradually self‑sustaining trajectory, in which private demand would take over from the State once the frontier is reached. If frontier‑level capacities remain accessible in open form, the French laboratory could achieve its technical objective without ever achieving economic equilibrium. The 700‑billion asset would then depreciate before it is realized.
This consequence also applies to our proposal. In a market likely to be distorted for a long time by price‑and‑opening strategies backed by geopolitical capital, it is prudent to privilege assets whose value resists price collapse of models: energy and connected land, a multi‑model inference fleet backed by purchase commitments, orchestration, evaluation capacity, and training skills. Above all, we must acknowledge that maintaining a domestic training capacity will probably not rest on the market alone. Like a defense industry, it must be supported by a multi‑year public order.
The increasing strategic role of the model orchestration layer
Over the past few months, value and strategic control have been moving towards the orchestration layer. This term designates the systems that connect models to data and to organizational processes, and that manage memory, permissions, result certification, and task arbitration according to their cost or complexity. If models are the engines of AI, orchestration is the chassis, the transmission, and the dashboard.
Signs of this shift are numerous. OpenRouter, a routing platform between hundreds of models, illustrates how rapidly the market revalues this layer: valued at $1.3 billion in May 2026 in a round led by Alphabet’s growth fund and Nvidia’s venture fund, it would, according to the Wall Street Journal, be the subject of acquisition talks by Stripe at barely $10 billion just two months later. Likewise, the commercial success of laboratories no longer rests on raw token sales, but on deploying agentic products that encapsulate models into structured workflows. Finally, the emergence of low‑cost offerings, such as Muse Spark or the open‑weight Inkling model, accelerates the adoption of multi‑model routers capable of selecting the most economical solution for each request. By making technology architectures interchangeable, orchestration keeps pushing toward commoditization of models. Much of a model’s capacity is in fact latent and revealed or lost by the surrounding system’s quality. However, two nuances hold. At the top of the capability hierarchy, the integrated stacks of leading labs still maintain a clear lead over equivalent assemblies built around open models: interchangeability is fully effective for routine tasks, but not yet for the most demanding autonomous workflows. Further, routing is primarily driven by an economic optimization logic: large French asset managers have thus deployed internal gateways that route each developer’s request to the cheapest model capable of handling it, aligning power consumption and task complexity.
This shift radically redefines dependency relations. It is now at the orchestration level that user adoption, institutional memory, and transfer costs from one system to another crystallize. An economy that masters this layer of applications can thus substitute one model provider for another at lower cost, in effect replacing a motor under the same chassis. Conversely, the one that relinquishes control remains captive, even if it maintains a sovereign model: it is no longer the raw algorithm that keeps the client, but the surrounding ecosystem. Sovereignty of use thus primarily hinges on orchestration.
Nevertheless, two planes deserve careful distinction. Economically, the orchestration layer is indeed where value is created. Geoeconomically, models become substitutable through orchestration, which enables shifting from an American supplier to a Chinese, European, or open supplier, provided a fallback solution exists. That is precisely what the base we propose below guarantees. Two risks specific to this layer remain to be identified and countered: a model supplier could degrade foreign orchestrators’ access to the interfaces they rely on, and orchestration actors tied to laboratories by capital could distort the game in favor of their home company.
This orchestration layer constitutes a favorable terrain for Europe, whereas frontier models have so far escaped it. Enterprise integration falls under business software, process knowledge, and sector expertise—core competencies of European market leaders such as SAP or Dassault Systèmes. Here, the entry ticket is measured in human capital and industrial expertise rather than compute power. The urgency in this domain is just as real: American labs are now aggressively investing in this applicative layer, and the window of opportunity may close quickly.
Moreover, a robust strategy on this crucial topic allows addressing both evolutionary scenarios. In the commoditization case, margins and value will cluster at the orchestration level. In the continual technology rupture case, this layer determines an economy’s ability to convert raw AI power into real productivity gains. With equal access to large models, the value gap will widen as a function of integration quality—a gap already separating American firms from European counterparts.
That said, orchestration is not a panacea: it complements the strategic portfolio, but cannot compensate for lack of access to high‑power models ( Mythos class), which are indispensable for cyberdefense and intelligence. It does, however, allow us to correct a persistent public debate bias: by focusing relentlessly on the frontier, we neglect the segment where our strengths lie and where entry barriers are surmountable. This project does not require unlimited funding; it simply requires unleashing execution levers already available: a Chinese sovereignty‑aware public procurement policy for critical application layers, multi‑supplier obligations to guarantee continuity of operation, and the structuring of sectoral data commons to safeguard our territorial advantage.
Applying the robustness test to the fatalist scenario and the runaway scenario
The fatalist approach is ineffective against the challenges we have identified. In the commoditization scenario, it immediately deprives Europe of the industrial positions it could take in a redefined accessible market. In the runaway scenario, it leaves Europe bare‑the‑bones in the face of sudden access restrictions.
Before presenting our proposal, let us anticipate a first objection that will be raised against it. The robustness requirement we defend must face its most serious critique: would a retreat strategy based on an actor located below the frontier technology, such as Mistral, be valid only in the scenario where it proves useless? If commoditization occurs, this retreat is easy but superfluous; if the frontier technology moves away, an eighteen‑month late model would be obsolete for cutting‑edge uses. This objection requires a two‑step answer.
First, the major risk to protect against is not technological divergence but the rupture of access, a latent threat in both scenarios. In case of blockage, we are no longer competing with the best global standard, but preserving a minimum autonomy. A sovereign frontier‑adjacent model, without reaching it, will always be infinitely more useful for essential state functions and economic continuity than nothing at all. The Ukrainian example is significant in this respect: without mastering the frontier of autonomous systems, its army transformed its military capabilities thanks to drones built from accessible building blocks integrated on the ground.
Finally, we recognize that a simple risk‑management logic is not enough in the runaway scenario. If frontier capabilities constitute a major sovereignty attribute, Europe will have to possess them sooner or later or secure strategic access. That is why our proposal goes beyond a protective policy: it is a doctrine of options, in the financial sense of the term. While insurance compensates a loss after the fact, an option today allows us, for a fraction of its real cost, to gain entry into the race at the moment we choose. Hence, the entire challenge of the trajectory we propose is to determine what must be built now to secure this option, without paying the exorbitant price in advance.
Our proposal: building an indispensable industrial base that holds in all technological scenarios
The plan we propose takes into account an asymmetry of time horizons often ignored in the sovereign AI debate. Indeed, the various components of a frontier effort do not share the same horizon: chips can be acquired in eighteen months once capital and allocation access are secured, whereas energy, connected land, large‑scale training skills, evaluation protocols, and decision bodies require five to ten years of construction. The Prometheus operation proposes immediately financing the fast and expensive brick: frontier training compute. Yet the uncertainty about its value seems too significant. We therefore propose to begin with the rocket’s first stages: now and at controlled cost, build the slow infrastructures that will retain their value regardless of the final scenario.
This alternative trajectory rests on a five‑part priority foundation.
The first concerns energy and connected land. Regardless of the technological scenario, Europe will massively need inference, and decarbonized, abundant electricity is the only value‑chain link in which France has a global comparative advantage, provided it aligns its energy policy. The competitiveness gap is documented: in 2025, industrial electricity prices in the EU were on average twice those in the United States and about half again as high as in China. Energy, permits, and grid connection thus constituted the real bottleneck of compute in Europe, more than capital or chips. To address this, it is essential to accelerate grid connections, purge ghost projects clogging waiting lines, create dedicated industrial zones to reduce permitting times to a matter of months, establish a one‑stop shop, and relaunch nuclear. This portion of the Prometheus plan must be fully executed, because it is required even if no model is manufactured on our soil. Its direct budgetary cost remains modest, since the essential driver is regulatory, while public land acquisitions and network guarantees amount to hundreds of millions of euros per year.
The second task is to deploy a sovereign inference fleet, i.e., a set of compute centers dedicated to the daily execution of models. In the short term, Europe will need to rely on Nvidia for this. Refusing American processors on the ground of dependence would confuse stock with flow: the strategic risk lies in future deliveries and revocable software access, not in installed physical infrastructure. Sovereignty of this fleet does not depend on the origin of its components, but on two imperative conditions: it must be located on European soil and operated exclusively by European actors. A data center in Europe but run by a foreign third party can be disconnected remotely, instantly; a center managed by a European company, with imported hardware, offers, at worst, several years of breathing room before technological obsolescence, a delay sufficient to reverse the balance of power. This sizing is not utopian: the most conservative demand projections indicate that by 2028 the inference needs of a large European country will far exceed any capacities currently planned on the continent. The imminent danger is under‑provision, not overprovision. At this stage, Nvidia and AMD remain essential, but material sovereignty implies, in the medium term, moving an increasing share of inference to specialized accelerators (ASICs), and in the longer term, mastering the full stack: interconnect, memory, packaging, compilers, and libraries, not a single chip isolated. Building now a European accelerator ecosystem is a natural step to extend the inference fleet.
For this fleet, aiming for a capacity of 2 to 3 GW by 2030 would represent an investment of €66 to €101 billion over five years, based on the unit costs advanced by Prometheus itself (around €33.4 billion per possessed gigawatt and €0.8 billion per year for operation). If that magnitude is impressive, it is the rigor of the financing that guarantees credibility.
The economic balance of such a park hinges on two critical variables: the utilization rate of capacities and the accelerated depreciation of chips, which lose most of their value in four to five years. The first risk can be neutralized by anchoring tenants. These strategic clients will secure the infrastructure’s profitability through firm take‑or‑pay contracts (the commitment to pay for reserved capacity whether used or not). These commitments will be undertaken by public procurement, by the competence manager of the third chantier, and by major European companies whose business continuity plans require a back‑up capacity on national soil. The second risk is managed by the capital structure. The core of the funding round should bring together long‑term investors (Public Investment Banks, European Investment Bank, infrastructure funds), backed by a first‑loss tranche borne by the State to reassure private capital, as the Prometheus authors rightly argue. Calibrated to depreciation risk, this public seed would represent 10–20% of the total, i.e., €2–4 billion per year. The remaining capital would be raised from sovereign funds eager to hedge against their own dependence on the Sino‑American duopoly, such as Norway, which can couple the power of its €1.8 trillion fund with its hydroelectric assets, or Gulf states, whose interest is already evidenced by the Emirati MGX fund’s stake in Campus AI alongside Mistral, Bpifrance, and Nvidia.
The Prometheus authors’ analysis of the need to derisk mid‑tier powers is accurate. Our proposal simply channels this dynamic toward a financially sustainable asset under all technology scenarios. This inference fleet serves simultaneously as a shield against access ruptures, as the physical backbone of our option doctrine, and as a tangible asset whose residual value—anchored in buildings, the grid, and land—makes it the most robust investment in the entire portfolio.The third chantier is safeguarding the training competence. It is a more demanding rewrite of the retreat actor’s role. The strategic function of Mistral, from the State’s perspective, lies not in its catalog of models, but in being today the sole place in Europe where the know‑how to train large models—the form of human and collective capital that cannot be bought in eighteen months—exists and is renewed. This know‑how becomes all the more valuable as the route of distillation to catch up closes: tomorrow, one will have to know how to train oneself, or else be completely dependent.
The multiannual public order (defense, security, sovereign functions) must be sized and contracted to fulfill this function: maintaining a world‑class training team in a condition of continuous training, a reasonable distance from the frontier, with verifiable capacity milestones. Concretely, this is a firm purchase commitment of €1.5 to €2 billion per year over five to seven years, plus a dedicated training compute of €0.5 to €1 billion per year, i.e., €2 to €3 billion per year in total. With this lever, the State buys an option. This instrument has a demanding counterpart: support is conditional and reversible. If the designated supplier does not sustain the objectives, the contract can be transferred to another vehicle, consortium, new structure, or recomposed team around the same infrastructures. The plan bets not on a single company, but on a capability, wherever it is located. That said, without a guardian of the capability, the frontier option does not exist, regardless of the capital available when one wishes to use it.
Fourth chantier: sovereign evaluation capacity and a map of dependencies. Independently evaluating the actual performance, weaknesses, and critical uses of the most advanced models costs around €100 million per year at the European level. This evaluation conditions the official categorization of uses (those with acceptable dependencies, those requiring a fallback solution, and those that exclude it), the credibility of access negotiations, and, above all, the reading of indicators on which the decision to exercise the option depends. It is the nervous system of the doctrine, which makes it operational and mobilizable.
Fifth chantier: collective negotiation of access guarantees. It is necessary to spell out precisely what would give Europe a real lever, beyond the received idea that Europe’s market size suffices. In a context of compute capacity shortages, European clients’ loss would marginally affect American laboratories, which can readily reallocate the freed power to domestic demand or to their own research. The mere status of “interested buyer” is therefore structurally harmless.
The real strategic lever for Europe lies elsewhere. It resides first in our infrastructure offer: Europe, and France in particular, can offer American suppliers a resource they sorely lack—a connected and abundantly available decarbonized electricity supply—in exchange for contractual guarantees of access to top models. This mechanism has a precious asymmetric property: once the compute center is implanted on European soil, the invested capital becomes hostage to the relationship. The supplier who reneges on commitments would find itself with billions of energy assets stranded, while the American government, if it sought to impose restrictions, would face the strong lobbying power of its own companies.
This lever then rests on a coalition of industrial bottlenecks. ASML’s monopoly is not isolated; the global semiconductor value chain also depends on Zeiss optics, Trumpf lasers, SK Hynix memory architectures, and Samsung. Coordinated among The Hague, Berlin, Seoul, and Tokyo, this collective weight carries geopolitical heft that ordinary commercial lobbying cannot. To this architecture must be added standard contractual clauses: advance notice, continuity of service, model parameter escrow with a trusted third party in case of access rupture, and a too‑often neglected condition: the absolute security of our own networks, with no laboratory willing to entrust its most precious models to permeable infrastructures. Certainly, none of these guarantees alone amounts to full sovereignty, and recent events remind us that granted access can be revoked. Yet, when combined with our inference fleet and our skill management, they upend the game: an access rupture no longer leaves Europe in paralysis; it imposes a prohibitive switching cost on the supplier. Thus, economic deterrence does not remove dependence, but makes it costly to exploit.
All told, the cost of this base for France and its partners is between €5 and €7 billion per year, i.e., around 0.2% of French GDP. This figure should be compared with the 1.5% required for Prometheus’ public component and the 4.5–8% of its total cost. Over five years, our approach mobilizes $90–$135 billion, financed predominantly by the private sector and backed by operating revenues. While Prometheus mobilizes largely public sums, we propose to mobilize capital that is mainly private, backed by productive assets and operating revenues. Fortified by this robustness trajectory, let us now turn to the maximalist scenario analyses.
Branch A: if commoditization takes hold, the diffusion offensive
If the next eighteen months confirm the incremental hypothesis, characterized by narrowing performance, commoditized capabilities, and a price war—Metas’ repositioning being the forewarning—the strategic center of gravity will definitively shift toward diffusion, i.e., the widespread adoption of AI by the economic fabric. As the Draghi report established, Europe’s productivity lag is as much an adoption gap as a production one. The €1,000 billion of annual spend mentioned by Arthur Mensch will be captured by models actually deployed in the field, not necessarily by the most heavyweight frontier architectures. In this context, cost of use, speed, integration into business processes, regulatory compliance, data localization, and control of agent workflows become the master variables of competition. In all these areas, European actors can build a durable competitive advantage. The provider’s nationality remains strategically important for reallocating value generated in Europe, for exploiting usage data to enrich future models, or for resilience against potential restrictions.
In this technology scenario, several levers must be activated immediately. First, the public levers that accelerate diffusion of AI use: electricity price, the anchor‑client effect of public procurement, administration as Europe’s primary buyer of intellectual services, the structuring of sectoral data commons in health, industry, and mobility, training efforts, and a regulatory overhaul actively encouraging operational deployment.
Second, European preference must be carefully calibrated so as not to impede diffusion. A blanket preference could artificially inflate AI costs for the whole economy, whereas it is essential to promote diffusion at scale. Even worse, blind preference risks equipping our hospitals, networks, and administrations with second‑rate cyberdefense tools at a time when adversaries’ offensive capabilities are being commodified. This preference must therefore be double‑bounded: reserved for identified critical uses within our base and conditioned on sovereign performance thresholds. No preference scheme should apply without an effectiveness floor: if the European actor cannot meet the required level in a given segment, the clause is automatically suspended. This is a key condition to align economic preference with security imperatives it intends to serve.
Building this doctrine of public policies will require leveraging an already engaged balance of power, but still open. The Cloud and AI Development Act, presented on June 3, 2026 as a central piece of the Tech Sovereignty Package, proposes a four‑tier framework of sovereignty requirements for public cloud purchases, but this text is still only a proposal: its real content will be hammered out in negotiations between the European Parliament and the Council, where the degree of effective sovereignty is contested. Our recommendation is to weigh in on that negotiation, rather than try to persuade a Commission that speaks with one voice, knowing that the internal arbitration between a more industrial‑policy‑favorable DG GROW and a more competition‑neutral DG COMP will continue to surface there.
Branch B: the runaway scenario and exercising the option
If performance indicators nevertheless confirm the runaway scenario, characterized by widening gaps on long and agented tasks, repeated performance leaps, a drying up of open architectures, and tightened access controls by American or Chinese decisions, the frontier effort becomes fully justified. Our base would then have been precisely designed to make this ambition possible, at a lower cost and risk. This operational branch does nothing more than reprise the essence of the Prometheus project, but it is activated at a moment when it becomes a certain and winning option (from the standpoint of sovereignty, technology, and trade).
However, this launch would differ from Prometheus’ initial sprint in five fundamental ways.
First, the starting point is cleaned up: energy infrastructures, connected sites, the inference compute fleet, and training teams already exist. The extra cost concentrates exclusively on training compute proper, preserving public funds during the ramp‑up phase. Activating the option does not make the frontier cheap; it guarantees an informed decision, a ready field, and a shareable burden.
Second, the trigger rests on empirical milestones reached, not theoretical bets. Yet, the geopolitical context that would trigger the option would also fundamentally alter the equilibria: a hardening of access dictated by Washington would legitimize our coalition, dispel the hesitations of hesitant partners, and accelerate the emergence of alternative chip architectures.
Third, the coalition would take a tighter, more flexible form, organized around our competence manager rather than around a new body created from scratch. The State would act as guarantor of strategic control, not as chief implementer of science. If Prometheus’ Commissariat‑type philosophy is the right one, the proposed treaty mechanism—guaranteed access rights in exchange for an entry fee—should be broadened beyond the European Union. It should include the United Kingdom, Switzerland, and Norway for their competencies and energy, Japan and South Korea for the semiconductor supply chain, and Gulf states for their capital. All these countries share a vital interest: to circumvent the Sino‑American duopoly.
Fourth, financing would lean on the proven architecture for the inference fleet, with patient capital from large sovereign funds, remunerated by guaranteed usage rights rather than mere promises of return.
Finally, the strategic objective would no longer rest on the uncertain promise of “joining the leaders in three years,” whose market dynamics have shown fragility. The ambition would be to break into the leading pack within half a decade, following a realistic learning trajectory consistent with the industrial history of technology catch‑ups backed by abundant capital.
Triggers and governance of our proposal’s decision
A doctrine of option has value only if an authoritative body monitors its critical indicators and is empowered to activate its execution. Yet this is precisely the blind spot in the current debate, which amasses strategic analyses without equipping itself with a decision‑making arm. To remedy this, our proposal is built around three concrete provisions, unified by a cross‑cutting principle: variable geometry. None of the following workstreams should be suspended due to the need for unanimity among the Twenty‑Seven, whose national interests are too divergent to sustain a “great power” policy on this issue. We must rely from the outset on a core of willing states.
First, governance requires a public dashboard of scenario indicators, operated by a sovereign evaluation authority in coordination with the European AI Office, but independent of its direct control. This tool would objectivally measure the evolution of performance gaps in executing complex agentic tasks, not merely preference rankings, the frequency and scope of American and Chinese access restrictions, the vitality of the open‑model flux, the hardening of distillation protections, the market penetration of autonomous agents in the real economy, and the industrial maturity of alternative chips to Nvidia. This dashboard would be reviewed quarterly according to predefined criteria, the crossing of which would trigger the activation of protocols.
Second, governance must rest on a clearly identified, deliberately flexible decision mandate. States involved in co‑funding, whether European or not, would form a board of contributors, while the European Commission would provide technical and legal support, without a veto right. This is the proven architecture of Europe’s industrial successes—from Airbus to space cooperation—standing in contrast to disintegrated multilateral governance. The decision to launch the frontier effort belongs to the political and budgetary sphere. It must be based on continuously updated numerical scenarios to avoid improvisation in the face of a sudden access crisis.
Third, the mechanism must integrate a private sector dimension, because the critical nature of these technologies affects companies first and foremost. European boards should treat access to models as a major supply risk. This implies implementing dual‑supplier inference policies, meticulously testing continuity plans that include switching to models operated under European jurisdiction, and introducing binding pre‑notice contractual clauses. This entire operational toolkit must thus be structured by drawing all lessons from the spring’s baptism of fire, with the partial suspension of access to Mythos and Fable.
The price of France and Europe’s independence
To invoke de Gaulle’s Rambouillet motto—“It is very expensive, but it is the price of independence”—is powerful, but it also highlights what could undermine Prometheus. If nuclear deterrence was a bold technological gamble, its strength lay in France’s ability to transform an initial dependency (the Norwegian heavy water, exfiltration in 1940; the ballistic expertise recovered in 1946) into a closed domestic chain within fifteen years, from Marcoule’s plutonium (1956) to the Albion plateau missiles (1971). This closure was achieved by abandoning heavy water technology, which depended on external supply, in favor of the graphite‑gas sector.
The fundamental objection to seeking a frontier model at short term is not only the project’s cost but also the sequencing of investments. We believe it essential first to build the base that makes it possible: energy, infrastructure, competencies, and decision institutions. We must prioritize what takes time. If the promoters’ hypotheses prove correct, our approach will at least have the merit of preparing the ground: the project will then find resolute partners and a market for components potentially freed from current monopolies.
In an uncertain world, technological and industrial sovereignty does not come from a maximalist bet but from a robust plan capable of accounting for all technological scenarios. It implies the ability to thrive if technology becomes ordinary, to defend oneself if access closes, and to enter the frontier race at the right moment. This industrial base, which we propose as an option value, costs nearly ten times less than Prometheus and does not require anyone in Washington’s permission.