How Europe Can Escape a Captured AI Ecosystem

This three-part series is published in partnership with the AI Now Institute.

Across this four-part series, we have argued that Europe's AI ecosystem is subject to the gravitational pull of dominant US players. Even where European AI companies succeed, value tends to flow upstream to US AI labs and, ultimately, hyperscalers — firms that are simultaneously Europe's indispensable infrastructure providers and its competitors.

We also showed how the response from Brussels has been to reach for the tools of the market: boosting supply by co-financing gigafactories or funding consortia to build models, channeling institutional savings into VC and growth equity, paired with deregulation and looser labor restrictions. The hope is that European AI companies, especially in health, manufacturing, defense, or legal, can replicate Silicon Valley's flywheel dynamic, in which successful startups produce exits, exits recycle capital and experienced operators back into the ecosystem, and the ecosystem produces more startups.

The internal logic of all these interventions is the same: once Europe removes regulatory friction, unlocks domestic capital, and concentrates its efforts where it can plausibly win, Europe will become sovereign. In Europe’s response, AI's trajectory is treated as a natural phenomenon: something Europe must respond to but cannot shape. A tsunami to brace for, or a wave to ride, but never a tide that could be turned.

What these interventions never explain is how they would fundamentally change the incentives in a structurally captured market: one in which start-ups like open-source model platform Hugging Face get bought by Nvidia, Germany's AI DeepL partners with AWS, and neither consumers, businesses, nor the public sector can make meaningful choices at any layer of the stack.

To manage expectations: we do not have a silver bullet — a lucid, innovative policy option that solves these problems in one fell swoop without causing pain in the European tech environment. We wish we did. But in contrast to the policy scripts that deepen Europe’s dependence, we have four that encourage a different way of thinking:

1) Take the market's uncertainty seriously

Two things can be true at once: the AI market is heavily concentrated, and it is also in flux. No technological shift or market correction will, on its own, make Europe a global leader in AI or limit the power of Nvidia and the hyperscalers. But which layer of the stack (applications, models, infrastructure) captures value in the long run is a different question, and one the market has not yet answered. Even the underlying technology is anything but fixed: we simply don't know if the dominant approach to building AI will continue to be based on ever-larger models trained on ever more data and compute, and run on the cloud.

What if open-weight models reach durable parity and models commoditize? What if leading labs respond by moving up the stack into enterprise products and agents? What if inference cost, not training cost, determines where value accumulates? What if the market corrects, and corrects sharply? These are just some of the questions pointing to radically different futures that each need a solid policy response.

We recommend that European policymakers start with the scenario the market demand is already signaling: customers are turning to open-weight models because they are often good enough and cheaper. Many suppliers, as well, are moving towards this thesis.

This is an early form of what the commodification thesis predicts: that several providers will offer models that perform roughly the same, or well enough for most purposes, and that switching between them will be cheap. There are two versions of how this scenario could play out:

In the first, the model itself stops being a scarce asset, as the cost of using interchangeable models becomes the cost of compute. Instead, value flows to whoever controls the hardware that inference runs on (ASML, TSMC, Nvidia, the memory makers) and to whoever runs models most cheaply at scale as part of services, which today means the hyperscalers (Amazon, Google, and Microsoft). Europe would remain dependent, but the dependence would simply move down the stack, from models to compute.

In researchers Arvind Narayanan and Akash Kapur's version of this scenario, AI Labs respond to this development by moving up the stack towards applications like enterprise products and "agents" that are embedded so deeply in workflows that removing them becomes prohibitively expensive. Europe would remain dependent on the dominant labs and on the hyperscalers whose compute they come bundled with. This runs directly against the interests of enterprise clients and the public sector who want choice and sovereignty.

Both of these scenarios demand a radically different policy response from Europe than the current frontier-access-at-all-cost debate assumes.

Whether models commoditize depends partly on whether large buyers demand portability or accept lock-in (and whether regulation and enforcement enable switching). Whether open-weight models stay relevant depends partly on whether anyone with resources keeps building on them, and on whether public procurement creates demand for them. Whether the labs' migration up the stack ends in entrenchment depends on whether interoperability requirements arrive before businesses and the public sector get locked-in.

In any case, Europe needs to urgently build the capacity to closely monitor where the market is heading, so that policy can prepare and adjust, rather than place a bet on a single possible future. Real scenario planning is not wishful thinking or apologetics — it is a condition for having agency in shaping the trajectory of AI.

2) Actively shape the market to be more open, more competitive, and more interoperable

Whatever AI's trajectory, one thing is clear: the current market is horizontally and vertically concentrated in ways that no single technological shift can undo. A handful of firms, such as Google, Microsoft, Nvidia, Amazon, and Meta and the labs they bankroll, control access to the key inputs necessary for building AI. Crucially, they also control access to consumers. The money they are pouring into these captive ecosystems functions almost like private industrial policy: subsidized computing power, financial guarantees and expanding private coalitions fundamentally steer what the AI market looks like today.

That is why the political debate over whether governments should engage in more active market shaping misses the point. The AI market is already being shaped, not only by governments but by a small number of companies with unprecedented power who are shaping the market in their own interest. The question is not whether the market gets shaped, but by whom.

The very least Europe can do is to stop incumbent firms from entrenching their dominance further: prohibiting self-preferencing for cloud providers and foundation model companies, or interoperability requirements across cloud infrastructure providers, are two examples of where to start. This is classic competition policy. Europe has these tools (although they need updating for the AI market), but they are still not viewed as a key tool for industrial policy strategy.

The more ambitious approach is to use every available lever to shape the AI market in Europe’s favor. Scenarios where AI stays expensive and the market stays concentrated are worse than ones where AI gets more open, more interoperable, and more competitive. But this needs to happen across the stack rather than at the model layer alone. Shaping the market in this direction is in the long-term interest of European companies worried about their margins, European governments worried about sovereignty, and European citizens worried about their jobs and about depending on tech companies they do not trust or cannot control. Shaping the market is not enough to ensure AI truly benefits people and economies (as we explore below), but it is a necessary condition.

None of this is easy. And the European debate treats market shaping narrowly as either subsidies or the occasional antitrust case. But the toolkit available to democratic governments is far wider: procurement, standardization, trade policy, taxation, prohibitions on bundling and tying, coordination among domestic firms, and only then public investment and competition regulation. If Europe is already debating compute expansions of unprecedented scale, it can debate deploying the rest of its toolkit as well.

None of this works if the state is the only actor that changes course. As long as private actors are overwhelmingly co-opted and vassalized by the existing technological empires, the momentum will not change. What is not often said out loud is that established private companies lack the incentives to jump off the bandwagon. European incumbents resell hyperscaler capacity, take their investment, and sit inside their partnership programs; defecting means forgoing revenue today for a market that only exists if everyone defects together. That's a collective-action problem - precisely the kinds of problems that states are designed to solve.

3) Treat cloud dependence and AI sovereignty as the same problem

Europe's dependence on three hyperscalers and the concentration of the AI market are one challenge, and the relationship between them is the most overlooked piece of the puzzle. As we have argued in our previous piece, most demand for AI is currently for proprietary foundation models that come tied to hyperscaler compute. As a result, European cloud providers face a chicken-and-egg problem: without demand for open-weight models, they cannot justify investing in AI-grade infrastructure, and without the infrastructure, the demand goes to the hyperscalers.

This bundling between closed models and hyperscalers is what makes AI dependence different from the cloud dependence Europe has already known. A database or an email system could in principle run anywhere. What keeps the public sector and so many European companies hooked on the same cloud providers is inertia, long-term contracts, egress fees, and dependence on the services that come bundled with the dominant cloud providers. If you are using a frontier model today, you are still predominantly relying on the hyperscalers, despite a recent wave of multi-cloud diversification.

While OpenAI and Anthropic have signed massive auxiliary infrastructure deals with neoclouds, their core operational backbones remain firmly anchored to Microsoft Azure, Amazon Web Services (AWS), and Google Cloud. Meanwhile, Google collapses the distinction entirely by providing their models through integrated platforms.

The economic harm of hyperscaler dependence in AI is also different. A company’s (often substantial) cloud bill is still an overhead that stays roughly the same whether the company does well or badly. AI inference costs depend on token use, and once proprietary AI is built into a product or a workflow, the provider takes a cut of every transaction that runs through it. If AI becomes as ubiquitous as its makers need it to be, that means a de facto tax on European economic activity.

This is why the Cloud and AI Development Act (CADA) fails to tackle the root of the problem. It treats projected future demand for one particular kind of AI, large-scale models run in the cloud rather than on devices or on premises, as a generic need for which Europe must expand capacity. If the paradigm shifts, that capacity sits idle. And if it does not, the demand in a captured market is predominantly for closed models that come bundled with hyperscalers.

CADA only reserves a fraction of public sector demand for European cloud providers, while aiming to triple data center capacity in ways that will predominantly benefit the hyperscalers. All of this leaves the fundamental structure of this market untouched. Especially in scenarios where AI commoditizes, this is fatal, as it concedes one of the layers of the stack where money is being made. Just not by Europeans, but by an oligopoly of cloud providers that can continue to seek rents.

To avoid this future, Europe needs to treat the diversification of its cloud market, not its generic growth in size, as a key goal of AI industrial policy. The cloud-AI dependency also shows why Europe has a strategic interest in AI trajectories where most uses of AI don’t need a hyperscaler at all: open models, or smaller models that run locally, on device, or on-premises. It also shows why procurement is such an important lever in creating demand that European suppliers could meet.

4) Decide what AI is actually for

Here is an important caveat: even a market in which AI is cheap, open, and the layers of the AI stack are contestable, is not necessarily one in which AI benefits the public, or the broader public interest. Market shaping contests who captures the value, not how it is distributed, or what is truly valuable in the first place.

Scenarios like Europe 2031 simply assume that deploying ever-larger frontier models will produce economic benefits that exceed the cost of the deepened dependence they bring. That is not just a technological prediction, but an economic bet. Only if you believe that we are on an unstoppable path towards AGI (or at least that models will keep improving and remain the cost-effective choice) does this bet seem safe. In that world, Europe’s AI industrial policy becomes about negotiating the least-bad terms of its dependence. If you do not, then the question of whether the bet is worth taking is open, and what we have argued throughout this series is that AI’s economic benefit for Europe, at least under the conditions of the current market, is anything but guaranteed. In fact, AI could end up both concentrating and extracting wealth from Europe to the privileged elites of Silicon Valley.

Declining the bet thus requires more than just aggressively shaping this market. It also needs an affirmative vision for AI’s role in society. Europe arguably has more room to imagine alternatives than the US, where the economic momentum hangs on the frontier cycle being true. Bain & Company estimates that making the spending on AI infrastructure profitable will require $2 trillion in annual revenue by 2030. That revenue has to come from somewhere, which is why we are seeing a push to integrate AI everywhere — even in places where AI tools are not fit for purpose — to see what may stick as a revenue opportunity.

A starting point for a better paradigm could be the distinction between augmenting and replacing. In a replacement logic, LLMs replace workers. This is implicit in the hopes of trillion-dollar valuations of the private AI labs — to replace, capture, and extract value from a whole segment of economic activity. Augmentation instead means adopting AI services only when and if they truly improve the productivity of workers and help them deliver useful services. Instead of replacing a customer service agent, LLMs can expand what that agent can do by providing new capacities. Choosing between these two ends is a question of power and politics, not technological inevitability.

This is also what decentering AI means in practice: refusing the premise that AI is an exceptional technology around which all other politics — labor, sustainability, democracy — must bend. Instead, it is a technology that can, and must, be shaped around the constraints of what kind of societies we want to have.

None of what we propose requires Europe to own every layer of the stack. It requires enough control, interoperability and bargaining power that people, companies and the public sector can switch providers and make meaningful choices. It also requires accepting AI’s fundamentally uncertain — and thus shapeable — future. AI is neither a wave nor a tsunami. To stop drifting, Europe needs to decide where it wants to go.