DT and Nvidia jointly funded the Industrial AI Cloud, Europe’s most ambitious sovereign AI facility
In early February 2026, Deutsche Telekom’s CEO Tim Höttges stood inside a gutted bank vault in Munich’s Tucherpark district and declared that Europe can do AI. Behind him sat 10,000 Nvidia Blackwell GPUs, wired together with 75 kilometers of fiber optic cable, cooled by river water from the nearby Eisbach, and already running workloads for Siemens, Perplexity, and a handful of robotics startups. The facility had taken six months to build. It cost one billion euros. And it carried a very specific label: Sovereign.
The Industrial AI Cloud is the most ambitious sovereign AI facility Europe has built. DT and Nvidia jointly funded it. T-Systems operates it. SAP provides the enterprise software. Siemens plugs in its simulation tools for digital twins. A European research consortium called SOOFI is using it to train a 100-billion-parameter open-source language model entirely on European soil. Over a third of its capacity was spoken for at launch, with industrial customers paying commercial rates. No government subsidy required.
All of which raises a question that European policymakers, enterprise buyers, and technology strategists should care about. Is it actually sovereign? And what would we even mean by that?
What sovereign actually means
The word has been stretched to the point of uselessness. Hyperscaler marketing teams use it. National governments use it. EU commissioners use it. Each means something different. The result has been a discourse where sovereignty is easy to claim and impossible to verify.
That changed in November 2025. At the Summit on European Digital Sovereignty in Berlin, all EU member states signed what became known as the Berlin Declaration. It contained the first formal political definition. Digital sovereignty is the ability of the EU and its Member States to act autonomously and to freely choose their own solutions, while reaping the benefits of collaboration with global partners when possible.
Two things matter here. First, the definition centers on the ability to choose, not the obligation to go it alone. Sovereignty is not autarky. The Declaration explicitly preserves space for global partnerships. Second, the accompanying Franco-German commitments demanded sovereignty across the entire value chain. Not just at the data layer, where GDPR already provides strong protections, but down through the infrastructure, compute, and silicon layers where Europe’s dependencies are deepest. Across the entire value chain is a very high bar.
The European Commission’s Cloud Sovereignty Framework, issued in September 2025, added operational specifics. Data residency requirements. Operational control by EU-based entities. Jurisdictional insulation from extraterritorial legislation, particularly the U.S. CLOUD Act. Software portability to prevent vendor lock-in.
Taken together, these policy instruments suggest a five-layer test for any facility claiming sovereignty.
Data sovereignty. Data stays within the jurisdiction, governed by its laws, processed by its residents, and insulated from foreign legal access.
Operational sovereignty. The facility is run by entities under the jurisdiction’s legal control.
Platform sovereignty. Software and cloud services come from entities under the jurisdiction’s legal and operational control.
Compute sovereignty. Processing hardware can be sourced, maintained, upgraded, and if necessary replaced through supply chains the jurisdiction can influence or control.
Strategic sovereignty. The jurisdiction retains the ability to make autonomous decisions about the facility’s future without requiring permission from foreign entities.
No facility will score perfectly on all five. The Berlin Declaration acknowledges as much by framing sovereignty as a capacity rather than an absolute state. But the framework sharpens the question. Where does a facility deliver genuine autonomy? Where does it create dependencies? And are those dependencies managed and diversified, or concentrated and single-threaded?
That last distinction is the one that matters most. Perfect sovereignty is not the standard. Managed dependency is. The real question is whether operators control their critical decisions, can influence their key relationships, or simply depend on external parties for outcomes they cannot shape.
Running the test
DT’s facility opened on February 4, 2026. Nearly 10,000 Nvidia Blackwell GPUs. Half an exaFLOP of compute. One billion euros invested. Over a third of capacity utilized at launch by paying customers including Agile Robots, PhysicsX, Siemens, EY, and Perplexity.
That commercial traction is itself a sovereignty signal. A facility that survives on government life support is not truly independent. DT proved that European industrial customers will pay for sovereign AI compute at commercial rates. Economic viability is a prerequisite for sustained autonomy.
Here is how the facility performs against each layer.
Data sovereignty: Strong.
DT operates the facility under German data protection law, GDPR, and NIS2. Data never leaves German soil. T-Systems, a German subsidiary, manages everything with EU-based staff.
This matters because the U.S. CLOUD Act gives American law enforcement the legal authority to demand data held by U.S.-controlled companies regardless of where that data physically sits. By operating through T-Systems rather than a U.S. hyperscaler, DT removes this vector entirely. No American entity has administrative or legal access to customer data or workloads.
Data sovereignty is the most mature layer of Europe’s digital independence, built on years of GDPR enforcement and regulatory iteration. It is also where the gap between European and American approaches is widest. DT’s facility delivers the real thing.
Operational sovereignty: Strong.
T-Systems manages infrastructure. Polarise, a European partner, built the data center in a repurposed former Hypovereinsbank facility. Cooling runs on Eisbach river water. Power comes entirely from renewables. Waste heat will supply the surrounding Tucherpark district. The entire operational chain, from facility management to incident response, runs through European entities under European jurisdiction.
Platform sovereignty: Strong, with a coupling risk.
This is where DT made its most deliberate architectural choices. The Deutschland Stack combines T-Systems infrastructure and T Cloud with SAP’s Business Technology Platform. Siemens contributes its simulation portfolio for digital twins. SOOFI will train a European language model entirely within the facility.
The platform layer is almost entirely European. Customers develop on SAP BTP, run Siemens simulation tools, and interact with T Cloud, all provided by entities under German and EU legal control. This is a genuine full-stack European alternative to AWS, Azure, or Google Cloud.
But every workload on this platform ultimately executes through Nvidia’s CUDA-X libraries and related frameworks. The platform is European. The runtime beneath it is not. This coupling means platform sovereignty is not fully independent of the compute layer below it. If Nvidia materially changed its CUDA licensing terms or restricted access to key libraries, the platform layer would feel the effects even though no platform-level decision had changed.
DT delivers strong platform sovereignty today. Its durability depends on the stability of what sits beneath it.
Compute sovereignty: The critical gap.
Every GPU in the facility is an Nvidia Blackwell chip. Designed in Santa Clara. Manufactured by TSMC in Taiwan. Governed by Nvidia’s proprietary CUDA software ecosystem. Subject to U.S. Export Administration Regulations.
This creates a triple dependency:
Chip design. Nvidia is the sole supplier. Europe has no indigenous GPU capable of commercial AI workloads at this scale. SiPearl’s Rhea1, developed under the European Processor Initiative, is an Arm-based CPU just reaching the sampling stage, designed for supercomputing rather than GPU-class AI. The EU-funded DARE project has 240 million euros for RISC-V chiplets but will not produce commercial hardware for years.
Chip manufacturing. Even SiPearl’s European-designed processor is fabricated by TSMC in Taiwan. If Europe designed its own GPU tomorrow, it would still depend on a non-European foundry to build it. The European Chips Act’s goal of 20 percent global semiconductor production by 2030 remains far from reality.
Software ecosystem. CUDA is a proprietary parallel computing platform with over two decades of developer tooling and optimized libraries. The degree of lock-in varies by workload. Pure AI training and inference are becoming more portable through PyTorch, JAX, and AMD’s maturing ROCm platform. But industrial simulation workloads, the kind Siemens and PhysicsX run on this facility, are deeply integrated with CUDA and would take years of engineering to migrate.
The probability that the United States would actually restrict GPU exports to Germany is low. Germany is a NATO ally, deeply embedded in Western defense supply chains, and allied pushback killed the January 2025 AI Diffusion Rule that would have imposed chip quotas on partner nations. But sovereignty is about structural capability, not current probability. The legal and administrative machinery to restrict GPU supply to any country now exists within the U.S. export control framework. Europe’s AI infrastructure operates inside a system where a foreign government possesses the ability to constrain supply, even if exercising it against a close ally remains unlikely.
DT does not deliver compute sovereignty. The gap is structural. DT cannot close it through better execution or smarter procurement. Only long-term European industrial policy can address it.
Strategic sovereignty: Constrained by compute.
DT can autonomously decide to expand capacity, onboard customers, develop new services, or shift commercial strategy. Those decisions are fully within European control.
But DT cannot autonomously replace its GPU vendor. It cannot diversify its hardware supply chain to a European alternative that does not yet exist. It cannot upgrade to next-generation compute without Nvidia’s cooperation.
DT’s own roadmap reflects this. The next phase, an AI Gigafactory developed with Nvidia and Brookfield Asset Management and competing for EU funding through the 20 billion euro AI gigafactories program, deepens rather than diversifies the Nvidia relationship.
The most honest way to map this is through three tiers. DT controls its data, operations, and platform decisions. DT influences its Nvidia relationship through purchasing volume and the leverage of being a flagship European customer. But DT depends on Nvidia for silicon design, on TSMC for fabrication, and on the U.S. government for export authorization. The line between what you control, what you influence, and what you depend on is the truest measure of any sovereignty claim.
What this means
DT’s AI factory delivers genuine sovereignty at the data, operational, and platform layers, and structural dependency at the compute layer. This is not a failure unique to DT. It describes every AI facility in Europe and most facilities globally outside the United States and China.
The question is not whether DT should have done better. Given available technology, it built the strongest sovereign AI facility Europe could have built in the time and budget available. The question is what sovereignty can reasonably mean in a world where GPU supply is concentrated among a handful of American and nascent Chinese suppliers.
Three things follow:
Sovereignty is layered, not binary. Europe can achieve near-complete sovereignty at the upper layers of the stack while still depending on non-European compute hardware. For a German manufacturer running digital twins on DT’s facility instead of AWS, the difference is significant and material. Data stays in Germany. Operations are managed by German staff. CLOUD Act risk is eliminated. The American GPU underneath does not erase those protections. It qualifies them. Europe should claim the sovereignty it has achieved and be precise about where dependencies remain, rather than overstating one or dismissing the other.
Compute sovereignty is a decade-long industrial project, not a procurement decision. DT cannot close the gap by switching suppliers. There is no European GPU to switch to. The path forward runs through three parallel tracks. Continued investment in European chip design through SiPearl, DARE, and startups like Axelera AI, accepting that these are five-to-ten year programs. Diversification of GPU sourcing as AMD, Intel, and emerging vendors mature. And investment in open software frameworks like SYCL, OpenCL, and Triton that reduce CUDA lock-in and make workloads portable across hardware. Software portability is the prerequisite for hardware diversification. None of these delivers compute sovereignty soon. Together, over a decade, they could move Europe from concentrated dependency to managed diversification. That is more honest and more achievable than promising full autonomy.
The real test is what happens under stress. Sovereignty reveals itself under pressure, not on a normal Tuesday. What happens when the GPU supply chain breaks, when export controls shift, when a vendor reprices, when a geopolitical crisis threatens Taiwan’s foundries? DT’s facility has genuine sovereignty at the layers where European entities make decisions and genuine vulnerability at the layer where they do not. The ability to operate the data center, serve customers, and protect data is sovereign. The ability to replenish, upgrade, and scale the compute hardware is not.
Conclusion
Deutsche Telekom’s Industrial AI Cloud is the most substantive sovereign AI facility Europe has built. It delivers protections at the data, operational, and platform layers that no U.S. hyperscaler can structurally match. It proves commercial demand for sovereign AI compute. And it positions DT for the EU-funded gigafactory program that will shape the next phase of European AI infrastructure.
But it also draws a sharp line between sovereignty achieved and sovereignty that remains a dependency. The compute layer, built entirely on American silicon running American software manufactured on a Taiwanese foundry, represents a concentrated reliance that Europe cannot resolve through procurement, regulation, or operational excellence alone.
The goal should not be perfect autonomy. It should be managed dependency. Minimize concentration at critical layers. Build diversification paths before they are needed. Invest in the open software and indigenous hardware that give Europe options rather than obligations.
The Berlin Declaration defined sovereignty as the ability to act autonomously and freely choose. DT’s AI factory shows that Europe can act autonomously in how it manages, operates, and governs AI infrastructure. What it cannot yet do is freely choose who supplies the silicon that makes it run.
Closing that gap is the work of the next decade. Defining it honestly is the work of today.
