Sovereign AI for Europe – The Supply Chain Problem

The European debate on “sovereign AI” is often reduced to regulation, data protection, and societal acceptance. What is often overlooked: Sovereignty in Artificial Intelligence is not only determined by algorithms or models but crucially by the supply chain of the underlying hardware. Without chips, without GPUs, without the necessary infrastructure, any vision of European AI sovereignty is nothing more than an academic exercise. In this post, I aim to highlight the real bottlenecks blocking Europe on this path and simultaneously identify the remaining opportunities for action. This will not be a romantic plea for autarky, but a sober analysis of dependencies, market mechanisms, and industrial policy options.

The Bottleneck Begins with the Machines

To manufacture GPUs, cutting-edge semiconductor fabrication technologies are required. These, in turn, rely on machines for Extreme Ultraviolet Lithography (EUV). The fact is: These machines come exclusively from one company – ASML in the Netherlands. Without ASML, there are no modern GPUs, no 3nm or 5nm chips, no accelerators for Large Language Models. The dependency is extreme: There is exactly one company worldwide that builds these machines. Every GPU used for AI training fundamentally relies on the delivery capability of a European company.

Sounds good for Europe? At first glance, yes, because a part of the value creation is in Europe. On second glance, however, it also means enormous vulnerability. ASML delivers globally, with the USA and Asia as main customers. The production capacity of ASML machines is limited, and demand has exceeded supply for years. Europe may hold the key to the door – but many buyers are queuing.

The Capacity Limit for GPUs

Even if the machines exist, their limited number restricts global manufacturing capacity. Every production line for GPUs depends on these highly complex systems. The result: Only a handful of foundries (TSMC, Samsung, Intel) can manufacture GPUs at this level. For Europe, this means: Access to GPUs is determined not only by political questions but simply by production capacity. Those who book early get the deliveries. Those who wait miss out.

NVIDIA and the Monopoly through CUDA

Hardware is one thing. Software is another. The critical lever for NVIDIA’s success lies not only in their GPUs but in CUDA – the proprietary programming interface that makes the parallel use of GPUs for Machine Learning efficient. CUDA is optimized, high-performance, and above all: patent-protected. This means: Even if another company builds GPUs, their software ecosystem lags far behind NVIDIA’s. AMD (AMD ROCm), Intel OneAPI, or smaller players have hardware alternatives, but for current AI tasks (Transformer models, LLMs, generative models), they are simply less performant or not as widely supported.

For startups, research, and companies, this means: It’s hard to bypass NVIDIA. Those who want sovereign AI are currently at the mercy of a US company.

Export Controls: Politics as an Additional Bottleneck

Another problem: Export restrictions. In 2023, the US government temporarily imposed export bans on NVIDIA cards to China and other markets. Europe was not affected, but the message was clear: If Washington decides that strategic interests take precedence, the supply line for GPUs can be interrupted or restricted at any time. The restrictions have since been partially lifted, but the risk remains. Europe builds its AI infrastructure on sand as long as it does not establish its own capability in hardware procurement.

Restrictions through Licensing Terms

Another less noticed detail: Even if one gets NVIDIA GPUs, they cannot be used arbitrarily. Gaming GPUs, which are cheaper and sometimes sufficiently powerful, are not allowed to be used in data centers per licensing terms (EULA). NVIDIA strictly differentiates between consumer cards and enterprise hardware. Those who want to build a business model in the AI sector must rely on enterprise cards like A100 or H100 – and these are in completely different spheres in terms of price and availability. For European startups or medium-sized companies, this is a de facto access restriction.

Power and Data Centers: Europe’s Structural Deficit

An underestimated problem: Even if the hardware were available, Europe lacks the data center and power capacity for GPUs on a large scale. AI training on H100 or A100 scaling requires data centers designed for power density, cooling, and energy supply. Europe lags significantly behind here. While hyperscalers in the USA and China are building gigantic GPU farms, Europe is discussing electricity prices, grid expansion, and site approvals. Without massively accelerated infrastructure policy, Europe remains dependent on external Clouds.

Pre-orders and Market Mechanisms

The biggest practical obstacle: Orders for GPUs are awarded years in advance. Meta, Microsoft, Google, and Apple secure billions worth of NVIDIA cards through long-term contracts. Those who want to order a GPU fleet as a European company today must expect to be delivered only in two to three years – and only if they can muster enough capital for billion-dollar pre-orders. For smaller players, startups, or public institutions, this is an insurmountable obstacle. The market is effectively pre-reserved.

Europe Between Aspiration and Reality

The bitter truth: Europe will not build a fully sovereign competitor to NVIDIA, OpenAI, or the hyperscalers in the short term. Both the production capacities and market power are lacking. Instead, it will come down to compromises: solutions based on foreign technology but combined with European governance, transparency, and infrastructure. Sovereignty in this context does not mean autarky but manageable dependency.

Europe must therefore do two things in parallel:

1. Accept the realistic framework – without hardware dependency from the USA and Asia, it won’t work in the medium term.
2. Build strategic flanks – in software, in open source, in European clouds, in specific niches of the value chain.

Opportunities for Action in Europe

Instead of clinging to an illusion of complete autarky, Europe should deploy its resources strategically:

• Strengthening ASML: Europe holds one of the biggest trumps with ASML. Investments in their supply chains, suppliers, and capacity are central.
• Alternative Software Stacks: Projects like OpenCL, ROCm (AMD), or OneAPI (Intel) deserve support, even if they lag behind CUDA.
• Open Source AI Models: Europe can lead in open, transparent models that do not depend on proprietary Cloud APIs.
• Cloud Consortia: Joint GPU infrastructures through alliances of European providers can help reduce availability barriers.
• Energy Infrastructure: No AI without electricity. Location policy must prioritize data centers with cheap, sustainable energy.

Conclusion

Sovereign AI in Europe is not a pipe dream, but neither is it a project for romantic autarky ideas. The reality is: We remain dependent – on ASML, on NVIDIA, on TSMC, on geopolitical decisions in Washington and Taipei. The art lies in strategically managing these dependencies rather than ignoring them.

Europe must learn that technological sovereignty is not black and white. It arises from a balance of cooperation and independence. Those who wait for the continent to one day build its own complete GPU and AI industry from scratch will fail. But those who cleverly use the existing strengths – from ASML to Open Source to the European Cloud landscape – can at least achieve a sovereign negotiating position.

The question is not: Can Europe build its own NVIDIA? – but: How do we manage to remain capable of action despite dependencies?