Germany’s sovereign cloud push is advancing with Deutsche Telekom and SAP winning a BMDS AI platform tender built on the “Deutschland-Stack”, while Thales expands its Google Cloud-powered sovereign model in Germany alongside its French PREMI3NS infrastructure. STL Partners has a good overview of the EU gigafactory push, alongside.
In sum – what to know:
Deutschland-Stack – Deutsche Telekom and SAP have won a €250m BMDS tender to build a sovereign PaaS for public administration after Google Cloud withdrew its protest; the platform runs on the Deutschland-Stack.
Sovereign redundancy – Thales is building a German sovereign cloud to mirror its French S3NS/PREMI3NS setup, both with (and separate of) Google Cloud; it claims full EU-based control and geo-redundancy.
Telco-based inference – Analyst house STL has reviewed the EU’s €20bn AI gigafactory scheme to build five large GPU clusters; telcos are often engaged as minority partners, but their opportunity is with inference services.
A couple of developments around sovereign cloud in Germany – as Deutsche Telekom (DT) has won a joint-tender with SAP from the Federal Ministry for Digitalization and State Modernization (BMDS) to build a local cloud system for government AI applications, and French aerospace and defence group Thales has said it will build a sovereign cloud in the country to go complement and also back-up its PREMI3NS facilities in France.
Apart from the location and the logic, there is a common thread here in the shape of US-based Google Cloud – which withdrew legal objections about how the BMDS tender was evaluated and awarded, to clear approval for the DT/SAP win, and which is also directly involved in the Thales proposition, providing the underlying infrastructure and platform technology for a new Thales legal entity to manage on sovereign EU terms.
Deutsche Telekom and SAP
DT and SAP were the “top-ranked bidder” on the BMDS tender, said the German telco. The deal, reportedly worth €250 million, is to provide PaaS services for AI applications on a “high-performance, secure, and sovereign” cloud platform, which will be shared by federal government, states, and municipalities. Confirmation of the decision by BMDS comes after Google Cloud and German IT consulting firm Adesso withdrew their “procurement complaints.
Adesso and Google Cloud had submitted their own Thales-style consortium-bid for the BMDS contract. Reports in Germany say they challenged the tender after being excluded from the procurement process for alleged formal errors in its bid documents. Their complaint triggered a review procedure before Germany’s procurement chamber, temporarily delaying the contract for Deutsche Telekom and SAP. Its withdrawal removed the legal block.
The BMDS platform is based on the so-called Deutschland-Stack, which underpins DT’s whole sovereignty push – combining its Munich-based AI compute factory, home to 10,000 Nvidia GPUs, with an management / orchestration layer from T-Systems, handling sovereign cloud operations, governance, and services. SAP and Siemens have been involved from the start, supplying enterprise AI software and industrial AI workloads, respectively.
The BMDS PaaS environment is deployed as a dedicated sovereign instance on the stack, tightly integrated with the underlying infrastructure, but also functionally separate. A statement said: “The platform is designed as a central hub for the public administration sector: scalable, expandable, and interoperable with existing specialized procedures. What begins today as infrastructure will be tomorrow’s development environment for AI-powered public services.”
It added: “The goal is for public authorities to build on common technical standards and platforms in the future instead of developing numerous standalone solutions. This is intended to make public administration more modern, secure, and efficient.” The first application is n AI solution, called KIPITZ, to support public sector staff with document processing, knowledge management, translations, text summarization, and faster planning and approval procedures.
Digital Minister Karsten Wildberger said: “We are implementing a strategic decision: high-performance digitalization for the federal government, states, and municipalities will run on infrastructure that we control ourselves – secure, scalable, and interoperable within Europe. Together with our partners, we are setting the benchmark by which digital sovereignty will be measured in the future. Now it is time to bring digital solutions into public administration at scale.”
Thales and Google Cloud
Meanwhile, the Thales project in Germany will follow its French model, where S3NS, a majority-owned Thales joint-venture subsidiary with Google Cloud, runs a kind of sealed EU-version of Google Cloud’s infrastructure. Its PREMI3NS facilities, distributed across three data centres, all managed by S3NS, launched in October (2025) and secured SecNumCloud 3.2 certification from France’s National Cybersecurity Agency (ANSSI) in December.
It was the first, and remains the only, market player to secure the ANSSI qualification for its infrastructure (IaaS), container (CaaS), and platform (PaaS) as-a-service layers simultaneously, in a single decision. The SecNumCloud 3.2 designation is a security and digital sovereignty qualification standard, created and enforced by ANSSI, and a component of the French government’s Cloud at the Centre doctrine.
This mandate requires public administrations and critical infrastructure (OIVs) – including such as healthcare providers, defence groups, and utility companies – to use qualified platforms to host sensitive data. Thales’ tie-up with Google Cloud in Germany will follow the same pattern: dedicated infrastructure, managed and operated by a new German entity, which Thales will fully own and control. It will be available generally by the end of 2026.
The new German operating unit, as yet unnamed, will be “legally and operationally independent from Google Cloud”, and managed by local staff – “ensuring that no third party, including non-European entities, can access the data within it”, it said. The combination of its cloud sites in France and Germany is an “industry-first” as a “pan-European, geo-redundant” sovereign cloud offering a “unique cross-border disaster recovery solution in Europe for Europe”.
The German solution is available in “preview” now. It is being designed to achieve C5 and C3A frameworks from Germany’s Federal Office for Information Security (BSI) – as a security-based evaluation framework (rather than a sovereignty certificate, about legal isolation etc), and a newer ‘cloud autonomy’ assessment (still not a full SecNumCloud-style certificate with legal and operational constraints; but Germany’s top sovereignty measure).
Christoph Ruffner, chief executive and country director for Thales in Germany, said: “This is a direct response to private and public sector German organizations wanting access to Google Cloud’s technology under full German control… We are delivering a solution that guarantees sensitive workloads remain protected from any extraterritorial reach while meeting the unique security and compliance requirements of our customers.”
Marianne Janik, vice president for EMEA North at Google Cloud, said: “[This is] a significant milestone in our commitment to digital sovereignty in Germany and Europe. By combining the power and scale of Google Cloud with Thales’ deep expertise in cybersecurity and local operational control, we are enabling German organizations even in the most sensitive sectors to innovate with confidence, while meeting specific legal and operational safeguards.”
The press note includes quotes from a host of Germany-based customers, including: AOK Niedersachsen, Deutsche Börse, University Hospital Schleswig-Holstein, Bechtle Clouds, Deloitte Germany, and SoftwareOne Deutschland.
Gigafactories and telcos
Separately, analyst house STL Partners has a new report about the EU’s €20 billion AI ‘gigafactory’ initiative, which will use public and private capital to fund up to five large-scale AI compute facilities to support frontier model training with more than 100,000 GPUs per site. The programme is the centrepiece of its InvestAI initiative, announced at the AI Action Summit in Paris in February last year (2025). The STL report provides context for the discussion above.
RCR wrote last week about Telefónica’s consortium bid for funds for a Spanish gigafactory. The narrative is, of course, about weaning the bloc of US (plus some Chinese) tech, whether hyperscalers – like Google Cloud, now repeating its French trick with Thales in Germany; or Microsoft and AWS, with similar initiatives – or frontier AI model providers. STL namechecks OpenAI, Anthropic, Google DeepMind and DeepSeek with regards to the latter cohort.
It writes: “Europe has world-class AI research, a growing open-source ecosystem, and one credible frontier AI player in Mistral, [but it lacks] the infrastructure to train and scale models. EU companies rely on US hyperscalers, [and] non-EU commercial priorities, supply constraints, and policy decisions. The policy addresses a strategic gap: ensuring Europe has the compute capacity and industrial base to participate meaningfully in frontier AI.”
It has a useful-looking map as well (see above) that shows the proposed location of all the bloc’s AI factories, whether giga-sized or standard – basically everywhere except Portugal, Turkey, Estonia, bits of Scandinavia (billed as AI factory partner-countries). It also provides a scale comparison (see below), gigafactory versus common-or-garden AI factory, including measures for site numbers, GPU counts, funding sources, funding models, and purposes.
STL says the European Commission (EC) received 76 submissions to its initial call for a show of hands in June last year, covering 60 sites in 16 member states. Proposed investments totalled more than €230 billion. Telefónica said last week it expected the final call for proposals to close in June, with selection decisions to follow later in the year.
It cites several other consortia: “Deutsche Telekom and Brookfield in Germany, Scaleway’s AION consortium in France, a Portuguese public-private bid involving Altice Portugal and NOS, Nokia’s Finnish proposal around the existing Lumi supercomputer, and Romania’s planned multi-site Black Sea AI Gigafactory at Cernavodă and Doiceşti.”
The firm also notes that commercial AI gigafactories are being built “on European soil”: OpenAI’s Stargate Norway facility in Norway, also with Nscale and Aker; Microsoft’s Start Campus site in Portugal, with Nscale again; Deutsche Telekom’s AI factory in Munich, discussed above. It runs through the value prop, as included at the bottom.
It writes: “The EU’s AI gigafactory programme… will not close the frontier compute gap with the US on its own: €20 billion is substantial by EE public-policy standards, but modest compared with the scale of AI infrastructure investment being deployed by US hyperscalers and frontier model companies. Even so, the programme matters.
“Frontier-scale AI compute is coming to Europe through a mix of public-backed initiatives and private-sector builds. The EU’s role will be to influence where that capacity is built, who can access it, and how much of the surrounding value chain remains under European control or influence.”
Gigafactories will be sited where significant power is available
“AI gigafactories need 200–500 MW of contracted, low-carbon power and a permitting pathway that can deliver capacity in three to five years. That rules out FLAP-D metros and directs the build-out toward Member States with grid headroom, hydropower or nuclear capacity e.g. the Nordics, Spain and parts of central Europe. Operators with land-banks, substation relationships and grid-adjacent fibre in those geographies will anchor the consortia that win sites.”
Long-haul interconnect moves from accessory to critical path
“Frontier training spans multiple sites, and the resulting models still have to reach end users under EU data-residency rules. That makes deterministic, high-capacity optical links, between gigafactory clusters, hyperscaler regions and edge inference nodes, non-negotiable rather than nice to have. The opportunity for fibre operators is unusually predictable: gigafactory geography points to Nordic-to-FLAP-D and Iberian-to-FLAP-D corridors as the primary new long-haul routes, allowing capacity planning ahead of demand rather than chasing it.”
Inference is the most accessible telco play, not the gigafactory itself
“A handful of European telcos are in gigafactory consortia (e.g. Deutsche Telekom with Brookfield in Germany, Altice Portugal and NOS in Portugal, Scaleway via the AION consortium in France), but they enter as minority partners alongside industrial, data-centre and energy infrastructure leads.Telcos will likely not own or operate the frontier training layer at scale. The more accessible value sits one tier down, in serving trained models to enterprises and public-sector users under EU data-residency, security, latency and compliance rules.
“Telcos bring a credible starting point with distributed infrastructure, enterprise relationships, security credentials, network APIs and edge compute a credible starting hand. However, this is the most contested market in European digital infrastructure: hyperscalers’ sovereign-cloud offerings, GPU clouds (Nscale, CoreWeave), specialist sovereign players (OVH, Scaleway, T-Systems) and Mistral itself are all chasing the same workloads. Telcos that wait for the gigafactory programme to bed in will find the inference customer base already spoken for.”
