Throughout a series of recent examinations, observers have documented how Europe's artificial intelligence sector remains pulled toward dominant American players. When European AI ventures do flourish, the gains typically move upward to American research institutions and ultimately to hyperscalers—the same corporations that serve as both Europe's critical infrastructure backbone and its most formidable rivals.
Brussels has responded by reaching for market-oriented instruments: co-financing large-scale model factories, funding industry consortia to develop new systems, directing institutional capital toward venture and growth equity, and pursuing deregulation alongside labor market loosening. The underlying theory holds that European AI firms, particularly those in healthcare, manufacturing, defense, and legal services, can recreate the self-reinforcing cycle that characterizes Silicon Valley—where successful companies produce lucrative exits, those exits recycle capital and experienced talent back into the ecosystem, and the ecosystem spawns fresh ventures.
Yet these interventions share a common flaw: they treat AI's development as an inevitable force that Europe must accommodate rather than guide. They fail to explain how such measures would alter incentives within a fundamentally captured marketplace—one where startups like Hugging Face end up acquired by Nvidia, where Germany's DeepL aligns with AWS, and where neither consumers, enterprises, nor government agencies retain meaningful control over any layer of the technology stack.
Rather than a single comprehensive solution, four alternative policy directions merit consideration, each encouraging a departure from approaches that entrench European dependence.
1) Take the market's uncertainty seriously
Two contradictory truths coexist: the AI sector exhibits extreme concentration, yet remains fundamentally unsettled. No isolated technological breakthrough or market adjustment will automatically position Europe as a global AI leader or constrain Nvidia and hyperscaler influence. However, the question of which layer—applications, models, or infrastructure—will ultimately capture value remains genuinely open. The foundational technology itself is far from settled: whether dominant AI development will perpetually depend on progressively larger models trained on exponentially more data and compute, delivered via cloud platforms, remains uncertain.
Several plausible futures diverge sharply: open-weight models might achieve lasting technical equivalence and become commodified. Leading laboratories might respond by moving upward into enterprise applications and autonomous agents. Inference expenses, rather than training expenses, might determine where economic value concentrates. Markets might undergo sharp corrections. Each scenario demands substantially different policy responses.
Market signals already point toward one scenario worth taking seriously: customers increasingly adopt open-weight models because they perform adequately and cost less. Many vendors are similarly shifting toward this position. This represents an early manifestation of what commodification would entail: numerous providers offering functionally equivalent models, with minimal switching expenses. Two distinct versions of this outcome exist.
In one version, models cease functioning as scarce resources as inference expenses converge toward pure compute costs. Value then flows to whoever operates the hardware supporting inference—ASML, TSMC, Nvidia, memory manufacturers—and to whoever runs models most efficiently at scale through services, currently the hyperscalers (Amazon, Google, Microsoft). Europe remains dependent, though the dependency shifts downward through the stack, from models to compute.
According to researchers Arvind Narayanan and Akash Kapur's formulation, laboratories respond to commodification by moving upward toward applications such as enterprise products and "agents" embedded so thoroughly in business processes that removal becomes prohibitively costly. Europe would remain dependent on dominant laboratories and the hyperscalers whose compute infrastructure they require. This directly conflicts with enterprise and public sector interests in choice and autonomy.
Both scenarios necessitate fundamentally different policy approaches than current debates centered on frontier access at any cost assume. Whether models commodify hinges partly on whether major purchasers demand portability or tolerate lock-in (and whether regulation enables switching). Whether open-weight models retain relevance depends partly on whether well-resourced actors continue developing them, and whether government purchasing creates demand. Whether laboratories' upward migration results in entrenchment depends on whether interoperability mandates arrive before organizations become locked-in.
Europe must urgently develop capacity to track market direction closely, enabling policy adjustment rather than betting on a single trajectory. Rigorous scenario planning represents not wishful thinking but a prerequisite for exercising genuine influence over AI's development path.
2) Actively shape the market to be more open, more competitive, and more interoperable
Regardless of AI's ultimate trajectory, one reality stands: current market structures exhibit horizontal and vertical concentration that no technological development alone can resolve. A small cluster of corporations—Google, Microsoft, Nvidia, Amazon, Meta, and their funded laboratories—monopolizes access to essential inputs for AI development. Critically, they also control consumer access. The capital these firms invest in proprietary ecosystems functions as quasi-governmental industrial policy: subsidized computing, financial assurances, and expanding private networks that fundamentally determine today's AI marketplace.
This reality exposes a flaw in the political discussion about whether governments should engage in more deliberate market intervention. The AI market is already being actively shaped—not only by governments but by a handful of exceptionally powerful corporations advancing their own interests. The genuine question concerns who performs the shaping.
At minimum, Europe should prevent incumbent players from consolidating dominance further: prohibiting self-preferencing by cloud and foundation model firms, or mandating interoperability across cloud infrastructure providers, represent starting points. These constitute conventional competition enforcement. Europe possesses these mechanisms (though they require updating for AI contexts), yet policymakers rarely treat them as central to industrial strategy.
A more aggressive strategy would deploy all available mechanisms to steer the AI market toward European advantage. Futures where AI remains expensive and markets stay concentrated prove worse than those where AI becomes cheaper, more interoperable, and more competitive. But this reshaping must span the entire stack, not just the model layer. Directing markets this way serves the interests of European companies protecting profit margins, European governments pursuing autonomy, and European citizens concerned about employment and dependence on uncontrollable corporations. Market shaping cannot alone guarantee AI benefits people and economies, but it forms a necessary foundation.
The challenge is substantial. European discourse treats market shaping narrowly as either subsidies or occasional antitrust enforcement. Democratic governments actually command a far broader toolkit: procurement decisions, technical standardization, trade mechanisms, tax policy, prohibitions on bundling, coordination among domestic competitors, and only subsequently public investment and competition law. If Europe already contemplates compute expansions of historic proportions, it can equally debate deploying its full range of instruments.
This approach fails unless private actors also shift course. As long as European companies remain deeply integrated into and dependent on existing technological empires, momentum will not change. What remains unstated is that established firms lack incentives to abandon the current arrangement. European incumbents resell hyperscaler capacity, receive their investment, and participate in their partnership frameworks; breaking away means forfeiting current revenue for a market that only materializes if all actors defect simultaneously. This represents a classic collective-action dilemma—precisely the category of problems governments exist to address.
3) Treat cloud dependence and AI sovereignty as the same problem
Europe's reliance on three hyperscalers and AI market concentration constitute one interconnected challenge, yet their relationship remains the most overlooked dimension. Most current AI demand targets proprietary foundation models bundled with hyperscaler infrastructure. Consequently, European cloud providers face a chicken-and-egg dynamic: without demand for open-weight models, they cannot justify AI-grade infrastructure investment, and without such infrastructure, demand flows to hyperscalers.
This coupling of closed models with hyperscalers distinguishes AI dependence from cloud dependence Europe has previously experienced. Databases or email systems could theoretically operate anywhere. What perpetuates public sector and European corporate reliance on identical cloud providers involves inertia, extended contracts, data egress charges, and dependence on bundled services. Using a frontier model today still predominantly means relying on hyperscalers, despite recent multi-cloud diversification efforts.
Though OpenAI and Anthropic have negotiated substantial auxiliary infrastructure agreements with alternative cloud providers, their primary operational infrastructure remains anchored to Microsoft Azure, Amazon Web Services, and Google Cloud. Google further collapses distinctions by delivering models through unified platforms.
The economic consequences of hyperscaler dependence in AI differ from cloud dependence. A company's cloud expenses represent relatively fixed overhead regardless of business performance. AI inference expenses scale with token consumption, and once proprietary AI integrates into products or workflows, the provider extracts fees from every transaction. Should AI achieve the ubiquity its promoters envision, this constitutes a de facto levy on European economic output.
This explains why the Cloud and AI Development Act (CADA) misses the fundamental issue. It treats projected future demand for one specific AI variant—large models operating in cloud environments rather than on devices or premises—as a generic requirement necessitating capacity expansion. Should the technological paradigm shift, that capacity becomes stranded. Should it not, captured-market demand predominantly targets closed models bundled with hyperscalers.
CADA reserves only a portion of public sector demand for European cloud providers while pursuing tripled data center capacity that predominantly advantages hyperscalers. This leaves the market's fundamental structure intact. In scenarios where AI commodifies, this proves catastrophic, as it surrenders one stack layer where economic value concentrates—but not for Europeans, rather for a cloud provider oligopoly continuing to extract rents.
Preventing this outcome requires treating cloud market diversification, not generic growth, as a central AI industrial policy objective. The cloud-AI entanglement also reveals Europe's strategic interest in AI developments where most applications avoid hyperscaler dependence altogether: open models, or smaller models running locally, on-device, or on-premises. It demonstrates why procurement functions as such a powerful mechanism for generating demand European suppliers could satisfy.
4) Decide what AI is actually for
An essential qualification: even markets where AI is inexpensive, open, and stack layers remain competitive do not automatically produce public benefit or serve broader social interests. Market shaping determines who captures value, not how it distributes or what genuinely matters.
Frameworks like Europe 2031 simply presume that deploying progressively larger frontier models generates economic gains exceeding the deepened dependence costs they impose. This represents not merely technological prediction but an economic wager. This bet appears prudent only if one believes an inevitable path toward artificial general intelligence exists (or that models will perpetually improve and remain cost-effective). Under that assumption, European AI policy becomes negotiating the least-damaging dependence terms. Without that belief, whether the wager merits acceptance becomes genuinely open. Throughout this analysis, the contention has been that AI's economic benefit for Europe, under current market conditions, remains far from assured. AI might instead concentrate and extract wealth from Europe toward Silicon Valley's privileged classes.
Rejecting this wager demands more than aggressive market reshaping. It requires an affirmative conception of AI's societal purpose. Europe arguably possesses greater latitude for imagining alternatives than the United States, where economic momentum depends on the frontier cycle proving true. Bain & Company projects that rendering AI infrastructure spending profitable requires $2 trillion in annual revenue by 2030. That revenue must originate somewhere, explaining the push to embed AI everywhere—even where AI tools prove unsuitable—to identify potential revenue sources.
A productive starting point distinguishes augmentation from replacement. Replacement logic treats large language models as worker substitutes. This assumption underlies trillion-dollar private AI lab valuations—replacing, capturing, and extracting value from entire economic sectors. Augmentation instead means deploying AI only when it genuinely enhances worker productivity and service delivery. Rather than replacing customer service representatives, language models could expand their capabilities through new capacities. Choosing between these paths reflects political and power decisions, not technological destiny.
This represents what decentering AI means concretely: rejecting the premise that AI constitutes an exceptional technology around which all other concerns—labor, environmental impact, democratic governance—must reorganize. Instead, it becomes a technology requiring shaping according to constraints reflecting the societies people wish to inhabit.
None of these proposals requires Europe to control every stack layer. They demand sufficient authority, interoperability, and negotiating capacity that individuals, companies, and government agencies can switch providers and exercise genuine choice. They also demand acknowledging AI's fundamentally uncertain and therefore malleable future. AI represents neither wave nor tsunami. To cease passive drift, Europe must determine its destination.
