GpuaaS is reinforcing the illusion of European AI sovereignty

At a glance:

• Europe relies on Nvidia GPUs for roughly 85% of its AI compute, a figure unlikely to fall below 75% before 2026.
• The EU has earmarked €20 billion for up to five AI gigafactories, yet US hyperscalers control about 70% of European cloud‑GPU capacity.
• Average GPU utilisation in Kubernetes clusters sits at only 5%, driving over‑provisioning and a false sense of scarcity.

What the report says

The European Union is pouring billions into AI development, from sovereign‑cloud contracts to the AI Continent Action Plan. The ambition is clear: build a continent‑wide stack that can train large language models and agentic AI systems. Yet the underlying hardware reality remains stark – the AI GPU market is dominated by non‑European designers, chiefly Nvidia, whose chips are fabricated in Taiwan’s TSMC foundries. Analysts estimate Nvidia holds about 85 % of the AI GPU segment today, a share projected to drift toward 75 % by 2026 as AMD and custom silicon gain ground.

This concentration matters because GPU‑as‑a‑service (GPUaaS) is the primary way European firms access the massive parallel compute needed for modern AI. By renting GPUs by the minute from cloud providers, organisations avoid the capital expense of owning hardware, but they also cede control over pricing, capacity allocation, and the strategic direction of the underlying compute.

European investment and infrastructure

The European Commission’s AI Continent Action Plan includes €20 billion for up to five AI gigafactories, part of a broader €200 billion InvestAI ambition. Earlier sovereign‑cloud contracts totalling €180 million were awarded to providers such as Scaleway, StackIT and Post Telecom. By 2026 Europe will operate 14 supercomputers and 19 AI Factories under the EuroHPC Joint Undertaking, backed by roughly €10 billion of public funding.

National champions are also stepping up. French cloud leader OVHcloud, Deutsche Telekom and T‑Systems launched a Munich‑based Industrial AI Cloud in early 2026, deploying 10 000 Nvidia Blackwell GPUs. French startup Mistral announced a €1 billion CapEx plan for 2026, including a €1.2 billion Swedish data centre and a Paris facility powered by 13 800 Nvidia chips.

Dominance of US hyperscalers

Despite these sovereign initiatives, the GPUaaS market in Europe is still dominated by US hyperscalers – Amazon Web Services, Microsoft Azure and Google Cloud. Together they account for roughly 70 % of European cloud‑infrastructure revenue and supply the bulk of GPU capacity on the continent. Their combined AI‑infrastructure CapEx is projected at $725 billion in 2026, a figure that dwarfs the €12 billion (about $12.6 billion) Europe expects to spend on sovereign cloud that same year.

The result is a structural dependency: European AI developers can access cutting‑edge GPUs, but only through platforms whose pricing, allocation policies and geopolitical exposure are set outside EU jurisdiction. Any shift in US export controls or trade policy could instantly constrain European AI projects.

Economic and strategic implications

Control over compute translates directly into economic capture. Hyperscalers not only charge for raw GPU minutes but also capture the margin on value‑added services built atop that hardware. European users therefore face externally set pricing and capacity constraints, limiting the ability to develop home‑grown AI products at scale.

Beyond the balance sheet, this dependency raises security concerns. Even when data is stored locally, the underlying infrastructure is governed by US law, exposing European AI workloads to extraterritorial subpoenas or policy‑driven access restrictions. A group of 18 European Parliament lawmakers warned that the AI gigafactory plan could deepen reliance on a single supplier, undermining the very sovereignty the initiative seeks to protect.

Utilisation inefficiencies and the illusion of scarcity

Industry studies show that GPU utilisation in Kubernetes clusters averages just 5 %. Seventy‑one percent of enterprises cite inefficient GPU use as a major barrier to scaling AI workloads. The typical response is over‑provisioning – buying more GPU capacity than needed to guarantee access – which reinforces the perception of scarcity while actually lowering overall return on investment.

Because GPUaaS does not optimise workload placement, many organisations end up with idle accelerators sitting in racks. This under‑utilisation erodes the economic case for large‑scale sovereign investments: even if Europe builds its own AI stack, inefficient allocation could turn billions of euros into underperforming assets.

Path forward for Europe

Europe’s strength lies not in replicating the full US‑scale hardware stack, but in leveraging strategic levers where it can exert control. Building sovereign compute capacity still yields tangible benefits: it cultivates domestic engineering talent, offers regulatory leverage over data flows, and creates a platform for European‑origin models.

Future competitiveness may depend more on orchestration layers – such as emerging capabilities from Mistral AI – and specialised equipment like ASML’s extreme‑ultraviolet lithography machines, the only tool capable of producing advanced chips at scale. Parallel experiments with cost‑efficient “neocloud” providers such as Nscale and Nebius could also help Europe extract more value from existing GPU resources without deepening vendor lock‑in.

In short, GPUaaS expands access but does not shift control. Europe must acknowledge this structural constraint and focus its resources on areas where genuine autonomy is achievable – from regulatory frameworks that shape market dynamics to niche hardware and software stacks that reduce reliance on a single foreign supplier.