Europe is spending heavily on artificial intelligence, with billions of euros flowing into GPU cloud services and supercomputer projects. GPU-as-a-service (GPUaaS) has become a key enabler, allowing organizations to rent processing power by the minute. Yet despite this expansion, the continent remains structurally dependent on non-European chip makers, primarily NVIDIA and AMD, whose GPUs are manufactured in Asia. This creates an illusion of sovereignty that masks persistent vulnerabilities.
The semiconductor industry is undergoing a boom driven by AI workloads. According to Deloitte, global semiconductor sales could reach $975 billion by 2026, with generative AI chips contributing roughly $500 billion. GPUs are essential for training large language models and agentic systems, but the supply chain is concentrated in the hands of a few US hyperscalers—Amazon, Google, Microsoft—which dominate European cloud infrastructure revenue by about 70%. European initiatives such as the AI Continent Action Plan, backed by €20 billion for AI gigafactories and a total €200 billion investment ambition, aim to build domestic capacity. However, these efforts do not alter the fundamental control of compute resources.
The European Commission has also funded sovereign cloud contracts worth €180 million, awarded to providers like Scaleway, StackIT, and Post Telecom. The EuroHPC Joint Undertaking operates 14 supercomputers and 19 AI Factories with roughly €10 billion in combined funding. French cloud leader OVHcloud and Deutsche Telekom's T Cloud Public, running 10,000 NVIDIA Blackwell GPUs, are part of this push. Yet at the base lies a persistent constraint: dependence on external suppliers for the chips themselves. NVIDIA holds approximately 85% of the AI GPU market, a figure expected to drift to 75% by 2026 as AMD and custom silicon scale, but the core dependency remains.
The economic implications are stark. Hyperscalers capture disproportionate value by intermediating access to scarce GPU resources, leaving European users exposed to externally set pricing and capacity allocation. The combined CapEx of Google, Amazon, Meta, and Microsoft for AI infrastructure in 2026 is projected at $725 billion, a 77% increase from 2025. This figure exceeds the GDP of many European countries and underscores the scale gap. European sovereign cloud spending is forecast at only $12.6 billion in 2026, an order of magnitude smaller. French AI champion Mistral has committed €1 billion in CapEx for 2026, including a Swedish data centre and a Paris facility powered by 13,800 NVIDIA chips, but such investments still rely on non-European hardware.
The Scarcity Illusion
Industry studies show that GPU utilisation in Kubernetes clusters averages just 5%, and 71% of enterprises cite inefficiency as a major barrier to scaling AI. Many organisations overprovision compute to secure access, reinforcing the perception of scarcity while reducing overall system efficiency. GPUaaS improves access but does not inherently optimise how compute is used. For Europe, this creates an additional risk: inefficient allocation could turn large-scale investments into underperforming assets.
Building sovereign compute capacity creates tangible optionality: it develops domestic engineering talent, establishes regulatory leverage over data flows, enables the development of European models, and builds institutional capacity to absorb sovereign hardware once European chip alternatives mature. The EuroHPC JU supercomputer network already enables research communities in smaller member states that previously lacked access to frontier computing. However, the debate remains focused on access and capacity, not control over allocation. As long as compute resources are distributed through external platforms, Europe's ability to shape how AI is deployed remains limited.
GPUaaS should be understood as a short-term enabler, not a structural solution. Long-term competitiveness will depend on securing strategic positions of leverage, whether through orchestration layers like Mistral AI, specialised infrastructure like ASML's lithography systems, or regulatory frameworks that shape market dynamics. Exploring more cost-efficient models, such as neocloud providers like Nscale and Nebius, may offer pragmatic ways to build on Europe's current position.
Europe is not falling behind because it lacks access to compute, but because it does not control it. While EU and member state efforts are moving towards a more competitive posture, the current trajectory risks deepening dependency rather than reducing it. Capacity building and sovereignty are not the same thing. The real question is whether Europe can convert managed dependence into genuine strategic autonomy. Compute dependency is not the only issue; data sovereignty and political autonomy are increasingly intertwined. Sovereignty will not arrive with the next gigafactory, but it might when Europe's own chips finally do.