What compute on European soil might buy Europe and what it depends on

The US is assembling the legal tools to switch off access to frontier AI infrastructure. On January 12, 2026, the House passed the Remote Access Security Act. The bill would let the Commerce Department license remote access to US technology, closing the “cloud loophole” that lets users barred from buying American compute rent it instead. On May 13, 2025, the Bureau of Industry and Security rescinded the AI Diffusion Rule, leaving it free to set advanced chip access through bilateral arrangements and individual licensing decisions. On June 12, 2026, an export-control directive ordered Anthropic to cut non-American access to Fable 5 and Mythos 5, its two most capable models, days after release. Rented compute, chips, and models are increasingly gated by US trust, and nothing guarantees Europe stays on the trusted side.

Europe’s compute plans, meanwhile, are outpaced by buildouts elsewhere. The EU currently holds around 5% of global AI supercomputing capacity, against roughly 75% for the US and 15% for China. In 2025 the EU announced it would at least triple its overall data-center capacity within five to seven years. But some projections estimate global compute capacity growth well beyond threefold by 2030. On that path, Europe’s share declines even if it meets its target.

A growing number of analyses argue that Europe cannot afford this drift, missing out on large benefits on multiple fronts. A large buildout on European soil, they contend, could strengthen Europe’s sovereignty, from securing access to frontier AI to capturing economic value and hardening its security. The most ambitious proposals call for Europe to host 15% of global AI compute within a few years, equivalent to over 50 GW of capacity. 

A second set of arguments turns on improvements for Europe’s capacity to manage the risks advanced AI brings. Recent months have seen security incidents at three AI developers, and a study has documented a terrorist group’s use of AI models. Europe currently has little means to detect, govern, or defend against such events. Proponents argue a European-based compute buildout would help Europe shape international AI rules and governance dialogues, strengthen its defenses against AI-enabled harm, and decrease AI risks overall.

This post gathers the arguments for a large European-based compute buildout and the questions they depend on. These arguments currently sit scattered across reports and commentary, and stand or fall with uncertain questions. This post aims to provide a better overview by comprehensively and descriptively mapping both. Settling the case will require more: testing those arguments against the questions and weighing them against the risks and opportunity costs. The closing section outlines an agenda for that work.

Open questions for a European-based compute buildout

A European-based compute buildout is here meant as an increase in the global share of AI compute capacity sited in Europe. Questions on ownership, scale, design, and other factors are deliberately left open and outlined below, as they often remain contested and each changes what a buildout might deliver. 

We find that disagreement about whether a large European-based compute buildout delivers the benefits set out in the sections below usually traces back to several unresolved questions. The six non-exhaustive questions below are the ones we find to be doing most of that work:

1. How will AI’s value develop? Capabilities have climbed steadily, but forecasts of AI’s economic and security impact diverge sharply. It is unclear how fast capabilities and value creation will scale, whether value will concentrate at the frontier or be captured by open and last-generation models, and whether it will accrue to infrastructure and model providers rather than to those adopting them. Similarly, whether near-frontier open models keep being released bears directly on this, since capable public alternatives likely make securing frontier AI access matter less.
2. How will the compute market develop? Future demand for AI development and inference compute is contested, though many trajectories point to steep growth. Shifts in chip prices and availability could change the calculus considerably, since, for example, the tighter the market, the more leverage a given European share will likely confer. A separate question is whether a European buildout would accelerate AI progress overall, and with it the risks as much as the benefits. That depends partly on whether that capacity would have been built elsewhere anyway, and views diverge.
3. What form should the buildout take? Ownership: Fully securing sensitive data may require a high degree of European control, whereas foreign-owned but European-sited compute might already confer enough leverage to negotiate frontier AI access. Scale: Fiscal returns plausibly grow with size, all else equal, but some forms of leverage may appear only past certain thresholds. Design: Supercomputing-class clusters and small last-generation clouds would bring very different economic returns. Security: Below a certain security bar, foreign firms, or the governments licensing them, may not place frontier AI model weights in Europe at all, fearing adversaries could steal them. 
4. Would foreign actors support and permit it, and at what cost? A large European-based compute buildout depends on political as much as corporate decisions, from chip export controls to permissions for foreign data-center expansion. In a more volatile environment, controls could hit Europe again. Furthermore, what concessions would it take to attract private investment? The main economic incentive for data-center buildout may be time to power, but whether that alone is enough remains unclear. Subsidies, tax breaks, relaxed environmental rules, or softer ownership terms may also be required, each carrying costs and externalities that could offset the gains.
5. Can Europe deliver such a buildout? Even with the will, the buildout faces hard bottlenecks such as grid capacity and connection queues, permitting, relevant labor supply, technical hurdles, and public backlash, among other factors, which could each cap how much actually gets built.
6. Would it convert into leverage? Some benefits of a large European-based compute buildout follow directly from hosting it, such as the expertise gained in building and operating data-centers. Many others run through leverage. Europe can trade access to compute that foreign firms and governments want to use for what it wants in return—from frontier access agreements to a guaranteed share of compute capacity. Whether that leverage materializes is therefore decisive. Would private developers and foreign states have enough reason to grant it? And how much leverage is there if much of the stack, from chip design to models, stays foreign-controlled? Can Europe even act as a unified actor to wield it? Further, could foreign involvement in European compute deepen the dependence it was meant to reduce?

Every benefit mapped below depends on at least some of these questions. A comprehensive case for a European buildout will require justifying the answers it assumes, showing how much weight each benefit carries under them, and comparing the gains against the risks and opportunity costs. We hope this gives that work a better starting point. The benefits proponents claim can largely be characterized into two groups, sovereignty and AI risk mitigation, which the next two sections take in turn.

What proponents say European-based compute might buy Europe

Proponents of a large European-based compute buildout point to a wide range of benefits for European sovereignty, from economic growth to hard security. As mentioned, whether they materialize depends on the questions above and other uncertainties that we may have missed. We cluster the sovereignty benefits below under the five pillars of EU frontier AI sovereignty defined in our recent report.

1. Economic Competitiveness: strengthening innovation and productivity, attracting talent and industry, and increasing AI-driven tax revenue. If frontier AI proves decisive for value creation, European firms and workers left on less advanced systems risk being outcompeted by those who can deploy the frontier. Frontier AI access secured in exchange for European-based compute could avoid this scenario and accelerate growth and innovation of European industry. Compute proximity could also benefit latency-sensitive sectors, from manufacturing to autonomous vehicles and financial trading. Further, together with frontier AI access agreements, European-based compute could also attract and retain talent and startups, while the expertise gained building it could help construct further clusters for European universities, governments, and firms. Compute may also offer fiscal stability: as AI shifts output from labor to corporate income, jurisdictions hosting none of it risk an eroding tax base, while locating those activities in Europe might provide a revenue stake.
2. Resilience and Preparedness: reducing dependence on compute and models Europe neither hosts nor controls. As AI grows more economically and strategically important, so does Europe’s dependence on models, compute, and other upstream components it neither hosts nor controls. European-based compute could reduce that dependence on several fronts. Most consequentially, compute that foreign providers want to use could give Europe something to trade for access agreements to frontier AI models. Such agreements could also secure a guaranteed share of hosted compute capacity that might otherwise be cut, whether through sanctions, export controls, or other retaliatory measures, or simply because a provider facing a tight market serves US demand first and Europe last. With enough usable compute capacity, Europe could go further and train or fine-tune proprietary models, though these are unlikely to be at the frontier. And in a crisis, such compute capacity could work as a strategic reserve, keeping critical public-sector and defense workloads running even if foreign cloud or chip access is curtailed. Finally, frontier AI access could help European governments keep pace with a growing volume of AI-generated activity, from machine-drafted legal filings to synthetic content, that is hard to handle without AI as capable as the systems producing it.
3. Security and Defense: keeping sensitive workloads under European control, and trading compute for frontier AI access that defense requires. Defense-relevant AI and other sensitive workloads run on foreign compute are exposed to interference, espionage, or theft. Routing compute on European soil instead could enhance control of the data, the model artifacts, and the security around them, though those gains depend on the right ownership and security design. And as AI spreads through critical infrastructure and national security applications, some proponents argue Europe will need enough compute of its own to run such workloads securely and independently. Frontier AI, meanwhile, strengthens both attack and defense. Falling behind on access therefore leaves Europe exposed to anyone who has it, and even model weights held by a well-intentioned actor can leak or be stolen by malicious ones. A large European compute base could secure frontier AI access agreements. Those agreements would help close the gap, letting Europe build frontier AI-enabled defensive capabilities.
4. European Values and Identity: strengthening Europe’s ability to enforce its own rules. Sovereignty also means setting and enforcing one’s own rules without undue external interference, whether on privacy, competition, platform governance, or AI. That requires the ability to resist foreign pressure, and today that capacity is thin. When the Commission moved against US providers under the Digital Markets Act and Digital Services Act, Washington answered with threats of tariffs and trade retaliation. Dependence on foreign compute and models adds a further pressure point: the implicit risk that access is curtailed. As frontier AI grows more central, that dependence deepens, and each enforcement decision risks becoming hostage to it. European-sited compute could blunt this on two fronts. Foreign states and firms that rely on it would have more reason to accept Europe’s decisions, and the access agreements such compute can secure would reduce the threat hanging over enforcement.
5. Foreign Relations and Perceptions: improving Europe’s negotiating position and giving it a greater appeal as a coalition partner. Reduced dependence, secured through access agreements, shields Europe at home and changes how it can act abroad. Being less reliant on foreign compute and models could enable it to engage internationally on its own terms, forging partnerships, shaping agendas, and defending its interests, instead of responding to terms set by others. Hosting a meaningful share of the world’s compute could also make Europe a third pole in a field otherwise split between Washington and Beijing. That position would itself be a bargaining asset, with Europe serving as an important ally to a US caught in a bipolar AI race it wants to win. It would also make Europe a more appealing partner to middle powers seeking to strengthen their economic and security standing without locking into a US or Chinese stack. They might value the compute Europe brings, and a Europe sitting at a critical node of the AI stack is harder to leave out of the rooms where global rules are set.

What proponents say a buildout would do for Europe’s ability to manage AI risks

A second set of arguments holds that the same buildout would improve Europe’s ability to manage AI risks. We follow the 2026 International AI Safety Report in defining frontier AI risks using three categories: malicious use, malfunctions, and systemic risks from widespread deployment. Some of these benefits build on the sovereignty gains above, turning enforcement capacity or negotiated frontier access toward risk management. Others follow from what compute on European soil makes technically possible, or simply from where that compute sits. As with the sovereignty benefits, whether they materialize depends on the open questions above, and on whether they justify the investment required.

1. Safety Rules and Standards: more compliance pressure, firmer jurisdictional grounds, and standing to shape international measures. The leverage set out in the European Values and Identity pillar above applies equally to rules written for AI risk, such as the EU AI Act’s targeting of risks from AI placed on the European market. Providers that come to depend on European compute would gain reason to comply, just as foreign governments would be left with fewer pressure points once access is secured by agreement. Hosting, training, or fine-tuning on European soil may also give future legislation firmer jurisdictional grounds than “placing on the market” alone, which matters for attaching safety conditions to the infrastructure itself. Finally, countries that host the infrastructure models depend on tend to get a say in the rules governing them. That could give Europe real standing in international safety measures.
2. Structural Stability: keeping democracies in the lead, and spreading compute so that no single actor is a point of failure. Where the world’s compute gets built will help decide whether democracies keep their lead in AI, a lead that is worth protecting. Expansion in the US is running into hard limits on power and permitting, pushing new capacity abroad. The main alternative considered today, the Gulf states, is judged a poorer one: close ties to Beijing, a taste for AI-enabled surveillance, and a home in a volatile region. A European buildout could be the safer bet. That lead might also reduce risks directly, since systems that suppress bad news and tolerate no dissent are less likely to catch and correct errors in powerful technology. Distributing compute across democracies could reduce risky concentration for the same reason. Capability this powerful held by one or two entities, however well-intentioned, is a single point of failure, vulnerable to unintended behavior, internal power shifts, or other threats with potentially catastrophic results. Spreading it leaves more independent hands able to spot and respond to dangerous developments.
3. Oversight Infrastructure: verifiable enclaves where frontier models can be examined without their weights leaving a trusted perimeter. Independent auditors need deeper-than-black-box access to catch serious risks, but developers routinely refuse it, fearing exposure of trade secrets and model weights. A ring-fenced, attested enclave, buildable largely on existing technology, could substantially reduce that fear and allow risks to be assessed before deployment. The tooling is closer than commonly assumed: hardware attestation combined with activation-commitment schemes can already verify, at roughly 1% overhead, that a given output came from the claimed model and configuration, and open-source integrations with standard inference engines already exist. Applied to rules like the AI Act’s Code of Practice, that could turn compliance from a self-report into verifiable proof. Yet these enclaves need not be large. Tens to low hundreds of GPUs would plausibly suffice, a tiny fraction of the compute capacities envisioned. A large buildout might nonetheless help bring them within reach. If frontier AI weights sit on European-based compute, extending a verification partition inside that perimeter is a far smaller ask for a provider than shipping weights to a standalone facility. And working verification deployments could show policymakers and industry the benefits of verification technology, potentially building the appetite for R&D investments in future verification technology that could ease international AI agreements.
4. Defensive Capacity: strengthening Europe’s defensive capabilities through frontier AI access agreements. Frontier AI access agreements, traded for use of a large European-based compute base, could let Europe build frontier-enabled defensive capabilities to prevent and mitigate AI-enabled harms, assuming capabilities are not overly offense-dominant. That resilience could also become a collective good. Because attacks will likely gravitate to the least-defended target, Europe deciding to extend frontier-derived defensive tools to smaller neighbors and partners could raise the regional floor, closing off weak links adversaries could otherwise exploit. Compute sited in Europe also falls under European jurisdiction, providing the option to impose Know-Your-Customer requirements to screen out actors of concern.

Where future research needs to go

This post has gathered the benefits proponents claim a large European-based compute buildout could deliver and mapped the questions those claims depend on. Settling what Europe’s compute strategy should be requires more than that. We would welcome work on the following:

1. Identify what assumptions each benefit requires. Every benefit above depends on particular answers to the open questions outlined before. For each benefit, further work should establish which answers it requires to manifest, when it stops holding, and where uncertainty remains. That includes naming the failure modes that break a benefit and identifying how to mitigate them.
2. Understand whether access agreements can actually be bought with compute. Many of the benefits above rest on the assumption that hosting compute capacity foreign providers want converts into agreements guaranteeing access to frontier AI and/or a share of that capacity. Further work should set out the assumptions, incentives, and mechanisms that would make such trades work. 
3. Map the risks and opportunity costs. A large buildout might carry certain costs: capital that cannot be spent elsewhere, concessions on environmental and energy rules, and, depending on ownership, deeper dependence on a few foreign firms. They deserve the same treatment as the benefits, including mapping them, establishing the conditions under which they materialize, and judging which costs and risks are worth accepting.
4. Say which scenarios Europe should plan for, and why. Answers to several of the six questions in section two depend on how AI, compute markets, and geopolitics evolve, and some uncertainty will always remain. A European compute strategy will require judging which scenarios are more likely and which are consequential enough to prepare for. 

This post has not settled whether Europe should build a large compute base on its soil. It has mapped what proponents claim such a buildout would deliver, which open questions those claims depend on, and what further work must answer to place a European compute strategy on firmer ground. 

We sincerely thank Alejandro Tlaie Boria, Oscar Delaney, Chloé Touzet, Philip Fox, Daan Juijn, Haakon Huynh, Matthias Samwald, Markus Anderljung, Tillman Schenk, Jitse Goutbeek, and our TFS colleagues Sven Herrmann, George Gor, and Kathrin Gardhouse for thoughtful comments on earlier drafts. Acknowledgment does not imply endorsement of the arguments or conclusions presented here. Any remaining errors are those of the authors alone.