Europe doesn't own
its infrastructure.
And that's a choice.

We built this cage ourselves. Not through negligence, and not through malice. Through rational decisions, made one at a time, over two decades, by people who were right about each individual choice and wrong about the aggregate. This article tries to document what that cage looks like from the inside, and asks honestly whether the door is still open.

In the nearshore hubs I move through in Lisbon every day, the French banks, the consulting firms, the delivery centres, I have not encountered a single company working seriously with a European cloud provider. Sometimes SAP, at the margins. And honestly, can you blame them? When you look at what AWS, Azure and Google have built over the past decade, in performance, in services, in global reach and in price, choosing a European provider isn't a rational technical decision. It's a political conviction. Most organisations don't run on political convictions.

I know this from experience. My first month at Cisco was an immersive training on how networks actually work: the physical layer, the routing logic, the protocols that sit beneath everything we call the cloud. What struck me then, and has stayed with me since, is something simple. Even the competitors in that space were American. Juniper. Arista. The alternatives to Cisco's switches and routers came from the same country. European data centres run on American hardware by default, not because anyone decided it should be that way, but because when the standards were being set, there was no serious European alternative. The dependency was baked in before it was even recognisable as a dependency.

Later, in a sales role at Devoteam, I worked alongside teams officially positioned as "cloud agnostic." The pitch was: we help clients navigate all cloud options. The reality was that the cloud practice ran on three specialist teams, one for AWS, one for Azure, one for Google, and we were premium partners of all three. When I was in client conversations, those were the options on the table. Not because anyone had consciously excluded European providers. Because there were no European providers in the pipeline, in the offer, in the conversation. The market had spoken, and we were operating inside the market's logic.

That's the observation this article starts from. Not an accusation. A description of how a dependency becomes invisible, and what it costs when you finally decide to look at it.

The European cloud market reached €61 billion in 2024, six times larger than it was in 2017. That growth looks impressive until you ask who captured it. Amazon, Microsoft and Google now hold 70% of the European cloud infrastructure market, according to Synergy Research Group. The three largest European players, SAP, Deutsche Telekom and OVHcloud, hold 2%, 2%, and less than 1% respectively. The remaining 15% is divided among dozens of regional players who, as Synergy's chief analyst put it, struggle to be European at all, often being only domestically focused.

European providers held 29% of their home market in 2017. By 2022 that figure had fallen to 15%, and it has stayed there since. Not recovering. Stable. The US hyperscalers invest approximately €10 billion per quarter in European infrastructure. That is, in the words of Synergy's own analysis, "an impossible hill to climb" for any challenger.

But the cloud layer is only part of the picture. The physical infrastructure underneath it, the switches, routers and network fabric on which data centres actually run, is itself manufactured overwhelmingly by US companies. Cisco, Juniper, Arista. That wasn't a conspiracy. It was the consequence of who invested, who standardised, and who had the industrial base at the moment those decisions were being made. By the time European operators were buying, the catalogue was already written in American. Sovereignty isn't a single layer you can swap out. It goes all the way down.

At the very bottom of that stack, today, are chips. Specifically: the advanced semiconductors that make AI compute possible. Taiwan Semiconductor Manufacturing Company produces approximately 90% of the world's most advanced chips, the ones inside Nvidia GPUs and virtually every piece of AI-capable hardware being deployed right now. Taiwan sits 180 kilometres from the Chinese mainland. This single geographic fact may be the most consequential geopolitical variable in the entire dependency question. The CHIPS Act, the Dresden fab, the Arizona facility: these are the right moves. They are also ten to fifteen years from maturity.

This dominance has a legal dimension that goes beyond market share. The US CLOUD Act of 2018 gives American authorities the power to access data held by US-based cloud providers regardless of where that data is physically stored. Microsoft's own legal director, before a French Senate commission, acknowledged that the company cannot guarantee immunity against a request from US authorities. When the ICC's chief prosecutor Karim Khan was temporarily locked out of his Microsoft account following a US executive order, it sent a concrete signal across European institutions: this dependency has consequences that live far outside the balance sheet.

The European Commission, the body that writes the rules, was found in 2024 to be in violation of its own data protection standards for using Azure. It is now in advanced negotiations to migrate to OVHcloud. The institution that built the cage was living inside it.

And the European alternatives have real limits, which deserve equal honesty. When Airbus launched a tender for sovereign cloud migration at €50 million over ten years, its EVP Digital was direct: European offerings are "not mature." They remain largely infrastructure plays. The platform and software layers, where engineers actually work, are still mostly American. This isn't a story of bad faith. It's a story of a technological gap that widens every quarter.

AWS generated $107 billion in revenue in 2024, at a 37% operating margin, representing over 60% of Amazon's total operating income despite being only 17% of its revenue. The e-commerce business nearly breaks even. Cloud is where the money is. When you pay your Azure bill, you're not just buying compute. You're funding the research, the sales force, and the infrastructure of the company that will be your landlord for the next decade.

There are exceptions, and they're worth naming because they illuminate the rule. Over lunch in Lisbon, a C-level at a major French bank told me they had kept their own data centres, not primarily for cost reasons, but because running their balance sheet through infrastructure they don't legally control was never on the table. "Sovereignty for us isn't a concept. It's a risk management line item." What struck me wasn't the decision. It was that it had been made years earlier and barely revisited since. For organisations handling genuinely sensitive financial data, the dependency question gets answered early and firmly. For everyone else, it doesn't get asked at all.

In April 2025, a data centre opened in a small coastal city 150 kilometres south of Lisbon. SIN01 at the Start Campus SINES Data Campus was inaugurated quietly, with 26MW of operational capacity. The scale of what's planned around it is anything but quiet.

The full campus will eventually span 1.2 GW across six buildings, making it the largest colocation site in Europe by IT capacity. Total private investment: €8.5 billion. Energy: 100% renewable, from a country where 87% of electricity already comes from renewables. PUE target of 1.1. EllaLink subsea cable landing on site. By most measurable standards, a genuine European infrastructure achievement.

In November 2025, at the Web Summit in Lisbon, Microsoft's President Brad Smith announced a $10 billion investment at Sines. "This is bigger than all the data centre investments Microsoft has ever made in Spain," he said. The plan: 12,600 next-generation Nvidia GB300 GPUs, deployed through a partnership with British AI company Nscale. The campus's first major tenant is not a European cloud provider, a European bank, or a European institution. It's Microsoft.

And those 12,600 Nvidia GPUs, manufactured by a US company, on chips fabbed in Taiwan, shipped to a Portuguese coast to run Azure workloads, compress the entire dependency chain into one image. The land is Portuguese. The energy is renewable. The cables are European. The compute is American, built on Taiwanese silicon.

This is not a criticism of Sines, or of Portugal, or of Microsoft. The investment creates jobs, tax revenue, and connectivity infrastructure that benefits the whole peninsula. But it illustrates the structural dynamic precisely: Europe provides the land, the energy, the cables, the permits. The value capture flows elsewhere.

What Sines makes visible in infrastructure, the AI gap makes visible in software. The mechanism is the same. Low prices to capture the market, lock-in to monetise it, switching cost too high by the time you notice. Cloud did this between 2010 and 2020. AI is doing it now, but faster, because the models themselves, not just the infrastructure running them, become the dependency.

In 2024, US-based institutions produced 40 notable AI models. China produced 15. Europe produced 3. That's from Stanford's 2025 AI Index. The gap isn't new, but in AI it compounds differently. When you're locked into Azure, you can theoretically migrate. When your organisation's workflows, your staff's habits, your institutional memory are embedded in a specific model's outputs, the switching cost isn't technical any more. It's cognitive. That's a harder dependency to unwind.

Private AI investment in the US reached $109 billion in 2024, nearly 12 times China's $9.3 billion and roughly 24 times the UK's figure. In generative AI specifically, US investment exceeded the combined EU and UK total by $25.4 billion. The cumulative gap from 2013 to 2024: $470 billion for the US versus approximately $50 billion for the entire EU.

The compute gap is even starker. The US controls an estimated 74% of global AI supercomputer capacity. The EU holds 4.8%. And behind this gap sits the same chokepoint we saw at Sines: TSMC in Taiwan. If the Taiwan Strait is disrupted, by conflict, blockade, or trade escalation, the queue for AI compute doesn't just slow down. It becomes political. Who decides who gets the chips? That answer will not be made in Brussels.

The honest reading: the entire stack, from chips to hardware to cloud to AI models, runs through a small number of chokepoints, almost all of them American or American-adjacent. This is not a tech story. It's a geopolitics story that happens to be expressed in technology. And Europe is, for now, a spectator in both.

Mistral AI is worth pausing on, not just as a data point but as a case study in what European AI looks like when it works. The French startup, valued at €11.7 billion as of early 2026, published 22 measures for European AI policy in April, calling for procurement preferences, fiscal incentives for local infrastructure, and an "AI blue card" for international talent. CEO Arthur Mensch: "Europe is not a market to dominate."

One note of transparency: Mistral has obvious commercial interest in European procurement favouring European providers, and they're raising €830 million in debt to build data centres in France and Sweden. Read their policy document with that in mind. But what's more interesting than the lobbying is the actual business model. Mistral's revenue isn't primarily from consumer chat. It comes from enterprise consulting, helping organisations implement, customise and govern AI systems under European legal jurisdiction. That's precisely the business that a genuine European sovereignty push would make structurally larger. They're not just calling for a level playing field. They're building the professional services layer that would run on it.

In November 2022, querying an AI model at the performance level of GPT-3.5 cost approximately $20 per million tokens. By October 2024, the same query cost $0.07 per million tokens. That's a reduction of more than 280 times in roughly 18 months, documented by Stanford's AI Index. The speed of that collapse is worth sitting with for a moment.

This isn't purely innovation giving value back to users. It's also a classic land-grab. OpenAI was reportedly on track to lose approximately $5 billion in 2024. Anthropic projected losses of around $2.7 billion. These are not sustainable operating models. They're market acquisition costs, the same logic that built Amazon Prime, Uber, and every other platform that subsidised usage until alternatives became unthinkable.

AWS in 2024 operates at a 37% margin. But in its early years, the pricing was deliberately aggressive. Once enterprises had built their architectures on it, once the switching cost was high enough, pricing power followed. There is no reason to believe AI inference will be different.

The risk for nearshore operations, including the ones I observe from Lisbon, is the FinOps problem replaying in AI. Unchecked API usage with bills that don't arrive until the architecture is embedded. The companies that learned this lesson the hard way with AWS are, in some cases, relearning it now with AI inference. The smart ones are mapping their exposure before the land-grab pricing ends.

What concerns me is not that the discourse is missing. At Munich in February, the German Chancellor said something unusually direct: "Nobody forced us into this excessive dependency on the United States. This dependence was self-inflicted." That's not a think-tank paper. That's the head of Europe's largest economy, on the continent's most visible security stage, naming digital dependency as a strategic failure in the same breath as defence and geopolitics. The language has shifted.

The European Parliament voted 471 to 68 in January 2026 for a report demanding European preference in public procurement, a €10 billion Sovereign Tech Fund, and legislation to triple European compute capacity within seven years. The Commission published a Cloud Sovereignty Framework in October 2025 with eight measurable objectives and a €180 million tender where any offer falling below the minimum threshold is automatically rejected. The Cloud and AI Development Act, expected in 2026, would establish eligibility requirements that could restrict non-European providers from public contracts. These are not aspirations. They are institutional movements.

What concerns me is the space between those movements and the room where the actual procurement decisions get made. That space, in the history of European technology policy, tends to be very large.

Gaia-X was supposed to be our structural answer. The project was serious, the ambition was real. Then Amazon, Microsoft and Google asked to participate in the governance. We let them in. And the project lost its meaning before it produced anything of consequence. That's not a failure of vision. It's a failure of will at the precise moment when will was required. It's also a pattern.

The question I keep coming back to, and I genuinely don't have the answer, is this: suppose a major European technology company, say Spotify, decided tomorrow to migrate its entire infrastructure from AWS to OVHcloud or Deutsche Telekom. It would be a powerful political signal, a demonstration that digital sovereignty can exist at industrial scale. It would also, very probably, appear in the next quarterly report under operational risks, and the market would react. Not because OVHcloud is bad, but because investors have integrated hyperscaler dependency as the norm, and any deviation from the norm reads as inefficiency. That's the trap inside the trap.

I think of the parallel with energy, because it seems exact. We developed a structural dependency on Russian gas over several decades, knowing, actually knowing, that this dependency created geopolitical vulnerability. We waited for the crisis to discover that the price was unbearable. In the tech domain, the crisis hasn't arrived yet. But the signals are identical. And we are, in several important ways, making the same mistake again with AI, building deep dependencies on infrastructure we don't control and don't fully understand, at prices that are subsidised and temporary.

Maybe we're already too far behind to do anything other than manage the dependency. That's possible. But I don't think we've seriously tried the alternative. And the alternative has a name: it's called using the leverage we actually have.

Europe has the largest regulated market in the world. 450 million people. A rule of law that companies across the globe respect precisely because it's the price of access to that market. We are not using this as a lever. We are giving it away.

The path I find most convincing, offered as a working hypothesis rather than a prescription, is conditional access. If you want to sell to European institutions, European banks, European healthcare systems, you meet European standards. Your data stays in European jurisdiction. Your infrastructure is auditable under European law. That is not protectionism. It's what the United States does with its own procurement rules, what China does with its digital market, what any serious economic bloc does when it decides its strategic interests are worth protecting. Europe has been the exception, the only major economy that opened its most sensitive infrastructure to foreign providers without reciprocal conditions.

This has a real cost, and it deserves naming. Conditional access means higher short-term friction for European buyers. It means some procurement processes get slower. It means some companies build on more constrained infrastructure. These costs were the reason European organisations chose Azure in the first place. The argument isn't that the cost doesn't exist. It's that continuing to deepen a dependency with geopolitical consequences carries a much larger cost, one that arrives later and is paid by everyone.

China made a different calculation, deliberately, over two decades, and Alibaba Cloud now dominates at home the way AWS dominates here. Russia didn't choose. Sanctions chose for it. But the result is the same: a domestic infrastructure that exists, imperfect, but sovereign. Neither path is available to Europe, or desirable. But both prove that the outcome wasn't a law of nature. It was the absence of a decision.

We have built something equivalent before. Airbus didn't emerge from market forces. It emerged from a deliberate European political decision, a decade-long plan, a preference for the European supplier until it was strong enough to compete without one. Today it challenges Boeing. Nobody calls that protectionism. They call it industrial strategy. The question is why we apply that logic to aircraft and not to digital infrastructure, when the dependency consequences are arguably more severe.

There is a structural difficulty in this path, and I want to name it directly because I think it's the most important thing in this article. China plans on twenty-year horizons because there are no meaningful elections to interrupt them. The United States benefits from the long time horizon of private companies: Amazon, Microsoft and Google operate on decade-scale strategies that no four-year government cycle can match. Democratic Europe changes Commission every five years. Majorities shift. Priorities rotate. The risk isn't that the CADA fails loudly. It's that it gets quietly diluted, one lobby meeting at a time, until it produces something that looks like Gaia-X.

So no, I'm not optimistic. I think the language has shifted, the institutional movements are more concrete than anything that came before, and the political urgency is real for the first time. I also think the window is narrowing every quarter, and that the dependencies compound while we deliberate. The question isn't whether Europe could build its own cloud giants, its own AI infrastructure, its own sovereign digital stack. It could. The question is whether it will decide to, before the crisis decides for it.