Microsoft: Germany Is EU’s Top Cyberattack Target
Eva Mickler
Among EU countries, Germany was the country where Microsoft customers were most frequently affected ...
Read Article
8 min read
Deutsche Telekom and NVIDIA announced the Industrial AI Cloud in November 2025 and brought it into service in Munich on 4 February 2026. More than 1,000 DGX-B200 systems, up to 10,000 Blackwell GPUs, the Munich site and the claim of sovereign AI for industry and the mid-market. It is the biggest hardware push Telekom has ever undertaken. For DACH CIOs, a different question matters: how much of this is location policy, and how much is a platform on which productive industrial workloads actually run.
Correction of 10 October 2026: An earlier version stated that the Industrial AI Cloud had launched at the end of April 2026. It also included an estimated roadmap and Wolfspeed as anchor customer. In fact, the platform was announced in November 2025 and has been running since February 2026. Data, schedule and customer information have been corrected.
5 Min. read
Key Takeaways
- Launch of the Industrial AI Cloud in Munich: announced on 4 November 2025, in operation since 4 February 2026. More than 1,000 DGX-B200 systems and RTX PRO servers with up to 10,000 Blackwell GPUs, operated by Deutsche Telekom on NVIDIA hardware (source: NVIDIA blog dated 4 November 2025; Telekom press release).
- Named users include Agile Robots (anchor customer), Perplexity, Noxtua and Wandelbots. SAP and Siemens are contributing their own platforms, and Mercedes-Benz and BMW are, according to Siemens, set to run simulations on it. The platform is a compute layer plus the NVIDIA software stack. Industry-specific models, MLOps and data integration remain the responsibility of the users.
- Sovereignty refers to location, operator and audit rights. The hardware architecture, accelerator stack and supply chain stay NVIDIA‑centric.
- Timeline: announced in November 2025, in operation since February 2026 and bookable via T-Systems. Public pricing, funding schemes for the mid-market and an expansion plan have not yet been published.
- CIO question: Which workload class justifies moving from a hyperscaler or on‑prem solution to a sovereign GPU cluster and what exit clauses are included in the contract.
What is the Industrial AI Cloud? A GPU cloud platform from Deutsche Telekom built on NVIDIA Blackwell hardware in Munich, in operation since 4 February 2026. More than 1,000 DGX B200 systems with up to 10,000 Blackwell GPUs provide compute for foundation model inference, industrial digital twins and real-time AI on production data. Sold as AI-as-a-Service via T-Systems, with contracting under German law; the first named anchor customer is Agile Robots.
What Telekom and NVIDIA actually built in Munich
The Industrial AI Cloud is Telekom’s largest single AI compute investment since the start of the AI cycle, a project worth around one billion euros. More than 1,000 NVIDIA DGX B200 systems and RTX PRO Servers bring up to 10,000 Blackwell GPUs into a converted former HypoVereinsbank data centre in Munich’s Tucherpark. NVIDIA provides the hardware architecture, the software stack (CUDA, NIM, NeMo, Triton) and the cloud operating model. Telekom supplies the data centre, power, connectivity and the contract wrap under German jurisdiction, with sales handled via T-Systems. Agile Robots, Perplexity, Noxtua and Wandelbots are named as the first users.
The official positioning is a sovereign AI platform for German and European industry. That is a statement about location and operation. The data stay in a German data centre, the operation is carried out by a German corporation, and the compliance clause references German and EU law. What is not sovereign: the accelerators are Blackwell, the software platform is NVIDIA, the model family is delivered via NIM micro‑services, and the strategic dependency on NVIDIA remains unchanged. Anyone reading the sovereignty claim as an architectural promise is interpreting the wording beyond what the press release actually says.
From the multi‑level view of the Hannover Messe analysis a clear program obligation emerges. Compute sovereignty in Munich can be combined with software sovereignty (an abstraction layer over the NVIDIA stack) and data sovereignty (own classification, own audit trail). The Telekom platform provides only the lowest layer. Those who purchase the middle and upper layers without scrutiny start a lock‑in program that, in five years, will be as hard to unwind as today’s switch between the major hyperscalers.
From launch to production operation: what has been documented since February 2026
Three months elapsed between announcement and commissioning. According to T-Systems, the platform has been available to customers since February 2026. The documented status at a glance:
Industrial AI Cloud: timeline
Nov 2025 | Announcement in Berlin by Deutsche Telekom and NVIDIA, volume of around one billion euros, privately financed.
4 Feb 2026 | Commissioning in Munich’s Tucherpark, in the converted former HypoVereinsbank data centre with partner Polarise.
Since Feb. 2026 | Available as AI-as-a-Service via T-Systems. The target group expressly also includes the mid-market, start-ups and public authorities; prices are available on request.
Open | The expansion plan, funding lines for the mid-market and a second site remain officially unconfirmed.
Between availability and productive workload, however, work remains on the customer side. First, companies must move their data pipelines onto the cluster before productive inference can run. In any industrial group that means at least three months of engineering work, often more. Second, NVIDIA’s NeMo and NIM stack needs its own model tuning per use case. Third, terms are agreed in individual discussions with T-Systems; there is no public price list. Any mid-market CIO planning productive workloads should factor in this lead time.
Sovereignty promise vs. platform maturity: where the trade‑off really lies
The strategic question is not whether the Munich cluster is a good industrial location. It is, when you look at power availability, connectivity and political backing. The question is which workload class actually justifies migration to a sovereign GPU cluster and which is better kept on hyperscalers or on‑premises.
Sovereignty promise vs. platform maturity (as of 10/2026)
| Pro Industrial AI Cloud Munich | Contra (as of today) |
|---|---|
| Location sovereignty (German law, T‑Systems operation, audit according to BSI basic protection) | Architecture lock‑in: NVIDIA hardware, NVIDIA software, NIM micro‑services as standard |
| Compute volume sufficient for foundation‑model inference and industrial digital twins | Terms on request only; public tariffs are missing |
| Political backing from the federal government and the EU as a sovereign‑AI anchor | MLOps layer maturity (CI/CD, model lifecycle) not yet publicly documented |
| Named users such as Agile Robots, Perplexity and Noxtua provide referenceable use cases | Exit clauses and data mobility in the contracts are the real negotiation points |
The pro side is clear for three workload classes: foundation‑model inference with data sovereignty, industrial digital twins with machine telemetry, and real‑time inference on production‑close data that must stay on‑site. For classic training workloads of smaller models, for standard MLOps and for administrative AI use cases, the hyperscaler remains the more economical option for now. The differentiation belongs in the written architecture decision, not in a blanket migration.
What really has to happen in the DACH‑CIO calendar by the end of 2026
The Munich cluster is a strategic option, and turning it into a programme is up to each company. Anyone who wants to pull the option should complete three concrete steps by the end of 2026. First, a workload classification that assigns every production‑grade AI application a data class, a compute requirement and a sovereignty requirement. Second, an architecture diagram that decides the three layers – compute, software and data – separately, with an explicit lock‑in risk per layer. Third, a negotiating position with T‑Systems that covers exit clauses, data mobility and tariff paths before the first production workload is migrated.
From the Google‑Cloud‑Next analysis you can also infer that the competition is not standing still. Google positions the TPU‑8i and agent‑inference pods for exactly the workload class the Munich cluster is targeting. Anyone who reads the Industrial AI Cloud as the only sovereign option ignores the fact that hyperscaler‑sovereignty constructs (German data centres, BSI‑certified operating models) are maturing in parallel. The negotiating position with Deutsche Telekom improves when the hyperscaler alternative is openly on the table.
The Industrial AI Cloud München is location policy with a serious compute substrate. It is unlikely to become the dominant AI platform for DACH industry in 2026. It is a valid option for the workload classes it genuinely serves. Anyone who structures the programme cleanly now will hold the stronger negotiating position in 2027, whether with Telekom, with the hyperscalers or with a possible second sovereign site.
Frequently Asked Questions
What distinguishes the Industrial AI Cloud from a hyperscaler region in Germany?
The platform is operated by Deutsche Telekom, the hardware is located in Munich, and contractual sovereignty falls under German law. Hyperscaler regions in Frankfurt or Berlin offer BSI‑certified operating models but are legally tied to US parent companies. The hardware stack is the same; the difference lies in location and operator.
When can mid‑size companies start running production workloads on the platform?
The platform has been available to customers since February 2026, expressly including small and medium-sized enterprises. T-Systems quotes terms on request. For the migration of production workloads, companies should plan for several months of engineering lead time.
Which workloads truly belong on the Industrial AI Cloud?
Foundation‑model inference with a high data class, industrial digital twins with machine telemetry and real‑time inference on production‑proximate data that must not leave the plant site. Standard training and administrative AI use‑cases often remain cheaper on hyperscalers.
How high is the lock‑in risk with NVIDIA?
High if the NVIDIA software stack (CUDA, NIM, NeMo, Triton) is accepted unchecked as the default. Manageable if an abstraction layer with container inference, portable model formats and an independent model lifecycle is introduced.
What must absolutely be included in the contract with T‑Systems?
Exit clauses with data mobility in standard formats, service‑level guarantees for GPU availability, tariff paths for scaling and audit rights. The negotiating position improves when hyperscaler‑sovereignty options are openly on the table in parallel.
More from the MBF Media Network
- Sovereign AI after Hannover Messe 2026: Architecture sovereignty as a multi‑level program
- TPU 8i and Agent‑Inference Pods: What Google Cloud Next 2026 means for the next board decision
- Meta Muse Spark: What CIOs can learn from the strategic shift for their own AI architecture
Source cover image: Pexels / panumas nikhomkhai (px:17489151)
Read more
- Technical Debt: Why the Board Must Act Now
- AI as a Digitalization Engine: What is Really Driving the German Economy Now
- Agentic AI without an owner: Who is liable when the AI agent makes a mistake
Translated from the German original using artificial intelligence. The German version is authoritative.