Source: By EmDee via Wikimedia Commons
Europe has spent years writing rules for artificial intelligence (AI). Its next task is different: it must find enough compute. Compute refers to the infrastructure that allows AI systems to be trained, tested and deployed. It includes data centres, specialised chips, cloud platforms, electricity grids, cooling systems, land and skilled operators. This material side of AI is well established. Yet European debates still tend to foreground regulation, rights and risk.
That balance is changing, with the EU’s AI agenda now extending beyond the AI Act and into the machinery that makes AI possible. Large models need scalable cloud infrastructure. AI companies need access to accelerators. Public institutions need trusted services for sensitive data. Data centres need power, water, permits and grid connections. The politics of AI is becoming a politics of infrastructure.
This is where compute sovereignty comes in. The term describes a state’s ability to secure meaningful control over the computational resources that underpin AI. That control has several layers. It depends on whether compute is located within a country or region, who owns the data centres and who controls the chips and accelerators inside them.
For the EU, this creates a hard policy problem. Europe can regulate AI. It can fund supercomputers and launch industrial strategies. It can define sovereign cloud through procurement rules and legal safeguards. Yet its cloud market remains shaped by hyperscalers and its data-centre expansion faces energy, water and grid constraints.
The central question is simple. Can Europe build the infrastructure for sovereign AI, and avoid deepening dependence on foreign clouds and energy-intensive data centres?
Beyond the AI Act
The AI Act gave Europe a template for how artificial intelligence should be governed. The AI Continent Action Plan gives it a different task. It seeks to make Europe a place where AI can be built, trained and deployed at scale.
This marks a shift in emphasis. For years, the EU’s comparative strength laid in market rules, standards and rights-based governance. The new agenda adds industrial capacity to that playbook. European AI factories are designed to give companies, researchers and public bodies access to high-performance computing, while AI gigafactories aim at the larger infrastructure needed to train and run frontier models. The planned Cloud and AI Development Act sits at the centre of this strategy, since cloud capacity links the scientific ambitions of high-performance computing to the commercial reality of AI deployment.
The Commission now faces a practical problem. Large models need cloud infrastructure across their lifecycle. Training demands high-performance computing. Deployment creates demand for scalable servers. Companies outside the small group that own such infrastructure depend on cloud providers. Public administrations handling sensitive data need highly secure EU-based cloud and AI computing capacity.
The planned Cloud and AI Development Act responds to these bottlenecks by tying Europe’s AI ambitions to three linked tasks. It would fund research and innovation, improve the conditions for investment in sustainable data centres and secure EU-based cloud and AI computing capacity for critical use cases. The data-centre pillar carries the most visible political promise. Within five to seven years, the EU wants to triple its data-centre capacity.
This turns digital sovereignty into an investment question. More compute requires more sites, permits, capital, electricity, water and grid connections. It also requires chips, accelerators and supply chains. The European Chips Act is important in this regard, because data centres become strategic assets only when Europe can secure the hardware that powers them.
The policy shift is clear. Europe still wants to govern AI through law, and now it also wants to shape the infrastructure beneath it, and out of this a central tension arises. The EU is trying to build sovereign AI capacity in a market where cloud power, chip supply and data-centre ownership remain highly geographically concentrated.
Three layers of control
Compute sovereignty gives a name to Europe’s AI problem. It moves the debate from models and rules to the infrastructure that makes them usable. A state can write AI law in parliament. It enforces that law more effectively when the machines that train and run AI systems sit within its jurisdiction.
The concept has three layers. The first is territorial. A data centre located in Europe gives EU and national authorities a legal hook over the physical infrastructure, while also giving companies and researchers closer access to compute. The second is corporate. A European site owned by a foreign cloud provider offers less control than infrastructure owned by a European provider. The third is industrial. Even a European-owned data centre depends on accelerators, especially graphically processing units (GPUs) and other specialised chips, whose design, fabrication and distribution sit in concentrated global supply chains.
Source: By Berkut via Wikimedia Commons
Each layer solves one problem and creates another. More data centres on European soil can improve supply security, but they also raise demand for energy, water and land. European ownership can reduce exposure to foreign law, but it also requires a provider ecosystem able to match the scale and reliability of the hyperscalers. Chip supply however brings the hardest limit. Even a data centre under European ownership still depends on hardware markets shaped by a small number of firms and production sites.
Compute gives governments something to grip. Advanced AI runs on chips that pass through concentrated supply chains, then sit in high-performance clusters inside data centres with grid connections, cooling systems and power contracts. These are practical points of leverage. Regulators can count large clusters, set reporting duties, attach conditions to procurement and shape where new data centres connect to the grid.
The EU in general has begun to turn sovereignty from a slogan into a procurement test. The Commission’s Cloud Sovereignty Framework asks who controls the provider, which law applies, where data and AI services are processed, who can operate the service, how exposed the supply chain is, how interoperable the technology stack remains and how sustainable the infrastructure is. Its highest assurance level requires technology and operations under complete EU control and subject only to EU law.
That framework shows the current developments direction. European AI sovereignty will be judged less by speeches about autonomy and more by infrastructure choices. Where is the compute? Who owns it? Which chips power it? Which law reaches it? Which grid sustains it? These questions set up the next part of the argument. Europe’s current answer to these questions still points toward foreign cloud providers.
Foreign Clouds, European Ambitions
Europe’s compute agenda runs into a market shaped by three American firms. A briefing by the European Parliament states that the EU has insufficient data-centre capacity, while the United States has roughly twice Europe’s share of global data-centre capabilities. The same briefing notes that three US-based companies account for 65% of the EU cloud services market.
That concentration matters because cloud computing is more than rented storage. It is the layer through which companies train models and manage data. For Europe, it means that a strategy built around sovereign AI begins from a market where the strongest suppliers are headquartered elsewhere.
This is the sovereignty paradox. The EU wants more EU-based cloud and AI computing capacity. It also wants a more competitive European cloud industry. Yet the firms best placed to finance, build and operate large AI infrastructure are often the same hyperscalers whose dominance creates the dependency problem.
Cloud dependence then becomes sticky. The OECD points to switching barriers, interoperability limits, complex migration processes and egress fees that can make leaving a provider costly.
The EU has already tried to loosen this grip. The Data Act includes rules on switching between cloud providers and safeguards against unlawful international access to non-personal data. The planned Cloud and AI Development Act would go further by creating better conditions for European cloud capacity. The Commission’s Cloud Sovereignty Framework also gives public buyers a more precise vocabulary for sovereignty. It asks whether services are anchored in the EU legal and industrial ecosystem, whether data and AI services stay under EU control, whether operations can continue without foreign control, and whether the supply chain remains exposed to non-EU dependencies.
These tools move the debate from aspiration to procurement. A sovereign cloud becomes something that can be scored, bought and audited. The framework’s highest assurance level requires technology and operations under complete EU control. That is a demanding standard. It shows how far the EU’s cloud ambitions stretch, and how large the gap remains between legal sovereignty, market capacity and industrial scale. More data centres on European soil can improve access to compute. They can also strengthen the position of foreign providers if ownership, operation and service ecosystems remain external. The question is less whether Europe can attract AI infrastructure. It is whether Europe can shape who controls it, who benefits from it and which dependencies come with it.
Procurement gives Brussels one lever. The Commission’s sovereign cloud procurement shows how demand from EU institutions can create a market for trusted providers. The Cloud Sovereignty Framework turns that preference into criteria. It looks at legal and jurisdictional sovereignty, data and AI sovereignty, operational sovereignty, supply-chain sovereignty, technology sovereignty and environmental sustainability. At its highest level, it points to technology and operations under complete EU control.
The Act should also keep sustainability inside the sovereignty debate. The EU already has a common rating scheme for data centres, requiring large operators to report indicators such as total energy consumption, water input, potable water input and waste heat reused.
The grid joins the race
Source: BalticServers.com via Wikimedia Commons
Europe’s race for compute sovereignty will be won as much in energy ministries as in digital departments. AI data centres are industrial loads with software attached. The IEA estimates that global data-centre electricity consumption reached around 415 TWh in 2024 and could more than double to around 945 TWh by 2030. In advanced economies, this growth arrives after years of relatively flat power demand. That makes AI infrastructure a test of grid planning as well as digital strategy.
Large data centres need firm power, grid connections, backup systems and cooling. The IEA warns that around 20% of planned data-centre projects could face delays if grid bottlenecks persist. A policy built around faster deployment also needs a view of where grids have spare capacity, where renewable power can be matched with demand, and where new loads would crowd out electrification, industry or households.
Water adds another constraint. AI infrastructure uses water directly for cooling and indirectly through electricity generation. The distinction between water withdrawal and water consumption matters because a data centre can draw water from local systems even when part of it returns later. The EU has started to make this visible. Its common Union rating scheme for data centres requires large operators to report indicators such as total energy consumption, total water input, potable water input and waste heat reused. That reporting base can turn a vague sustainability debate into measurable policy.
The next step is to connect disclosure with discipline. EU policy should move from reporting requirements toward binding sustainability targets that can contain rebound effects in the data-centre sector. This is the point at which sustainability becomes part of sovereignty. A cloud system that depends on strained grids, scarce water or opaque energy contracts creates another form of exposure. Control over infrastructure means little if that infrastructure produces new local vulnerabilities.
Chips sharpen the same lesson. More European data centres will need more accelerators, servers and advanced semiconductors. The chip map is diverse, with TSMC in Taiwan dominating advanced foundry capacity, South Korea’s Samsung and SK Hynix being central to memory chips. Japan remains important in silicon wafers and wafer-fabrication inputs, and ASML gives Europe a rare chokepoint in lithography. AI chip design, fabrication and distribution remain concentrated across a small number of firms and production sites. The Commission’s Cloud Sovereignty Framework recognises this problem through supply-chain sovereignty and technology sovereignty. Compute sovereignty ends up reaching far beyond the data-centre fence.
The danger is a larger European data-centre footprint paired with limited European control. The measure of success should be different. New capacity should expand access to compute, strengthen European providers, reduce cloud lock-in, protect critical public-sector uses and fit within Europe’s energy transition. Compute sovereignty is no longer a slogan about autonomy. It is a practical test of whether Europe can align industrial policy, digital regulation and physical infrastructure in one strategy.
Questions
- Should the EU prioritise faster AI infrastructure growth, or should new data centres only be approved when their energy, water and grid impacts are clearly manageable?
- Can Europe build a competitive cloud and AI ecosystem while keeping its digital market open to global technology firms?
- How much control over AI infrastructure does Europe need before digital sovereignty becomes more than a regulatory ambition?
Readings
image sources
- Belgique_-_Bruxelles_-_Schuman_-_Berlaymont_-_01: Wikimedia commons | CC BY-NC-SA 4.0 International
- 6600GT_GPU: By Berkut, via Wikimedia Commons | CC BY-SA 3.0 Unported
- BalticServers_data_center: Baltic Servers.com via wikimedia commons | CC BY-SA 3.0 Unported
