AI Camps in Germany: Live, Construction, and Planning Separated
Munich live, Lübbenau in billion-dollar construction, ESC in Brandenburg: How cloud teams are honestly planning AI factory capacity.
Munich is already delivering AI capacity, Lübbenau is planning on a multi-billion scale, Brandenburg hosts the AWS Sovereign Region, Bavaria is promoting Gigafactory plans: Germany is in the midst of a wave of AI Factories, campuses and hyperscale buildings. For cloud teams, what matters is not the announcement slides-it’s what goes live when, and what remains tied to grid capacity, permits and utilization.
Key Takeaways
- Live and pipeline collide. Telekom/NVIDIA’s Munich facility is up and running; Lübbenau and other mega-projects are under construction or in planning with long lead times.
- Capacity ≠ availability. GPUs, megawatts and “sovereign” labels only help if grid tie-ins, cooling and contracts are in place-see failed site projects.
- Operator question: Which workloads need local AI-factory capacity, which stay flexible in public cloud-and how do you measure utilization before you commit?
Related:Sovereign cloud fails at the grid, not the code / When GPUs eat the SaaS budget
The map: what’s live, what’s growing, what’s planned
Munich, Tucherpark – Industrial AI Cloud by Telekom/T-Systems with NVIDIA. Officially live since February 2026. Built with NVIDIA and data-centre partner Polarise. Announced: roughly 10,000 NVIDIA Blackwell GPUs (including DGX B200 and RTX-PRO servers), up to 0.5 exaFLOPS, operated in Germany with focus on industry, research and the public sector. According to Telekom, the factory was already more than one-third booked with first customers. This is the most tangible “AI Factory live” case in the DACH region-no mere slide deck.
Lübbenau, Lausitz – Schwarz Digits data centre. Schwarz Digits reports an €11 billion investment in a data centre on a former power-plant site, modular in construction phases, target completion end-2027, with capacity for up to 100,000 GPUs and green-power hook-up plus waste-heat reuse. This is hyperscale/AI-campus scale, not a regional colocation box. For the region it’s both industrial policy and energy-transition rolled into one.
Brandenburg – AWS European Sovereign Cloud. The first ESC region sits in Brandenburg and has been generally available since January 2026. It couples “sovereign” hyperscaler capacity to the same physical grid-constrained space where other large projects also compete for power and permits.
Schweinfurt / Blue Swan and further bids. Bavaria has positioned itself with a concept for an AI Gigafactory (including up to 100,000 AI chips in the narrative), but at the time of writing such projects are still applications, permitting or subsidy logic-not the same as “go-live next quarter.” PR blurs the line. Practitioners keep three buckets: live, under construction with a date, bid/plan.
Why the wave is coming now
Three drivers are converging:
- Demand for local AI capacity – industry wants training and inference workloads with data residency and low latency, not just API calls routed to the U.S.
- Policy and incentives – AI Factory initiatives, IPCEI-style narratives, regional and EU competition for “sovereign” compute blocks.
- Hyperscalers and industrial groups invest in parallel – AWS ESC, Telekom/NVIDIA, retailer IT (Schwarz Digits) and applicant gigafactories. This is not a single player race but a capacity sprint.
At the same time, the physical bottleneck remains: power, cooling, waste heat, zoning, skilled labor. A campus on paper without transformers and processes is marketing. The data-center paradox lesson from Babenhausen applies here too.
- First real AI-Factory capacity in Germany (Munich) as reference for SLA and onboarding
- Major projects (Lübbenau et al.) signal more GPU supply from mid/late decade
- Sovereign-cloud regions couple hyperscalers to the same local market
- Many announcements are still construction or application status
- Grid and permits can derail timelines
- Utilization and price per GPU-hour decide the business case
What cloud and platform teams should do now
1. Inventory demand classes. Which workloads truly need “GPU in Germany with operating model X” – and which only need cheap inference somewhere in the EU?
2. PoC against live capacity, not renderings. Munich and similar offerings can be tested today. Pipeline projects with 2027 horizons belong in scenario planning, not in Q3 production commitments.
3. Multi-source strategy. One factory is a single point of failure. ESC, European colos, Telekom Cloud and EU public-cloud regions are options – with clear failover criteria.
4. Fold site due-diligence into the plan. When you buy “local” capacity, you’re indirectly buying grid, cooling and politics. The seven site questions from the data-center paradox article apply 1:1.
5. FinOps before the hype-commit. GPU hours, utilization, idle rates – otherwise the campus story eats the same budget that is currently cutting SaaS licenses.
// Definition
What is an AI Factory / AI-Campus capacity? A bundled offering of GPU/AI-server infrastructure, often with an operator, located in one jurisdiction and focused on training/inference for industry or research – as opposed to a plain public-cloud region. “Campus” and “Factory” are marketing terms; technically, MW, GPU count, network, SLA and operating model are what matter.
Verdict
Germany is finally building physical AI capacity. Munich proves it can be done. Lübbenau and others are setting the benchmark for the coming years. AWS ESC shows that even hyperscalers are joining the location race. The strategic trap would be to confuse announcements with available capacity. Those who plan to produce in 2026 should book what’s live and measurably scalable. Those planning for 2028 need scenarios that include grid access, permits, and exit strategies-not just a map dotted with logos.
Frequently Asked Questions
What is an AI Factory?
A site or offering that bundles AI compute capacity (typically GPUs), operator, and often sovereignty or industry focus. The term isn’t standardized-what matters are GPU count, performance, location, SLA, and operating model.
Is the Telekom/NVIDIA cloud in Munich already usable?
Yes. The Industrial AI Cloud went live in February 2026. Telekom reports roughly 10,000 Blackwell GPUs and early industrial customers with measurable utilization.
What is Schwarz Digits planning in Lübbenau?
A large data center in the Lausitz region with an announced €11 billion investment, modular construction, a target completion of late 2027, and capacity for up to 100,000 GPUs-sited on a former power plant with energy and waste-heat focus.
Are all campus projects already approved and financed?
No. Some are live, some under construction, some still applications or political positioning (e.g., gigafactory concepts). Always verify status and sources before baking capacity into your architecture.
How should teams procure capacity?
Define demand tiers, run PoCs against live offers, plan multi-source, assess location and network risks, and lock in GPU FinOps (utilization, idle, commit) before long-term commitments.
Editor’s Reading Picks
- Sovereign cloud fails on connectivity, not on code
- When GPUs devour the SaaS budget
- Sovereign cloud extends far beyond server location
More from the MBF Media Network
MyBusinessFutureWhen a German AI model actually pays offDigital ChiefsSovereign AI: responsibility stays in-houseSecurityTodayWhat is KRITIS? Operators, obligations and thresholdsImage source: AI-generated (July 2026)
