AI compute is becoming the most consequential infrastructure of the century — and today it is concentrated in the hands of a few U.S. hyperscalers. Castellan exists to correct that imbalance. We are a Swiss company that develops, builds and leases data centres for AI in Europe and North America — as a powered shell or fully fitted, so our customers only bring their GPUs.
The United States hosts about three-quarters of the world's GPU-cluster performance (Epoch AI, 2025), and three U.S. providers — Amazon, Microsoft and Google — hold 63% of cloud infrastructure services spending (Synergy Research Group, Q3 2025). For most workloads, most of the time, that is merely convenient. For governments, defence and research institutions, banks, insurers, and healthcare systems — organisations whose mandates do not permit their data, model weights, or inference traffic to fall under foreign jurisdiction — it is a structural dependency with no easy exit.
The exposure is legal before it is technical. Under instruments such as the U.S. CLOUD Act, providers subject to American jurisdiction can be compelled to disclose data regardless of where the server physically stands. A data centre in Frankfurt operated by a U.S. entity is, in every way that matters to a European ministry or regulator, not European infrastructure. We set out the full argument in our analysis of why sovereign compute is Europe's defining infrastructure question.
European regulation is moving in the opposite direction of the market's concentration. The EU AI Act, GDPR, DORA, and a growing body of sector-specific rules increasingly demand demonstrable control over where models are trained, where data resides, and who can be compelled to hand it over. Workloads cannot always travel. Data sovereignty is not a slogan — it is becoming a procurement requirement.
And the constraint that decides who can respond is not silicon — it is power. Global data-centre electricity demand is set to more than double by 2030, to around 945 TWh (IEA, 2025), and the EU's InvestAI initiative aims to mobilise €200 billion for investment in AI (European Commission, 2025). Whoever controls powered land, grid interconnection, and energy supply controls the pace of the build-out. That is where Castellan starts.
"Sovereign" has become a label attached to almost anything with a European flag on the website. Castellan treats it as an engineering and legal specification: four layers, each independently verifiable — and each of which must hold for the whole to mean anything.
The same logic extends to the compute layer. Customers who need GPU capacity rather than a building can contract AI Compute with Castellan: capacity from vetted third-party hardware owners, selected to match their jurisdictional requirements. We are your contracting party; we do not own or operate GPU fleets ourselves.
Our build-out follows a deliberately simple sequence: secure power, sign the lease, build, commission, repeat — each phase building on the experience of the last. The pipeline reaches a minimum of 800 MW of AI capacity by 2032, with access to more than 5 GW of power through our network to scale further as demand requires.
We lease every campus in one of two equal models, both described under Powered Shell & Turnkey: a powered shell — land, power, building shell and core infrastructure — for customers who fit out to their own specification, or a turnkey data centre, fully fitted and ready for IT load, so customers only bring their GPUs. Both come on 15–20 year triple-net leases: cost transparency, control over your facility, and capacity on a schedule you can plan around. We normally build on our own balance sheet; organisations that want to own their facility can engage the same team through Design & Build.
If your organisation needs AI capacity that stands up to scrutiny — from your board, your regulator or your government — the conversation starts here.