Munich, We Have an Exaflop: The End of Excuses for SMEs

KI & Automatisierung · 7. Februar 2026 · Joseph Flesh

Telekom is building an AI fortress in Munich. For many SMEs, this sounds like science fiction. I say: This is your last chance not to be left behind.

Do you actually know what an exaflop is? Don't worry, I had to look it up too. It's a number with 18 zeros behind it. A quintillion calculations. Per second. Try to imagine that – you can't, can you? It's one of those astronomical figures where the human brain just gives up. And precisely this absurd amount of computing power is currently being created right on our doorstep.

The AI Fortress in Tucherpark: More Than Just Hot Air

Let's be honest: when Deutsche Telekom announces another 'innovation,' I usually just shrug. Too often, these were smoke screens to distract from dilapidated networks. But what has officially launched in Munich since February 6th of this year is a different story. The 'Industrial AI Factory' in Tucherpark is – and I don't say this often – a game-changer. Together with Nvidia and data center partner Polaris, the Bonn-based company has built an infrastructure in just six months that was previously unheard of in Europe. We're not talking about a few server racks in the basement. We're talking about almost 10,000 of the latest Nvidia Blackwell GPUs. This is the hardware that every tech guru in the US is raving about right now. And it's now in Munich, waiting for data from German SMEs.

The thing is: it's not just about raw power. The entire construct operates under the banner of data sovereignty. A word that long caused yawns in executive suites until the first GDPR penalties arrived, and people realized that their own production data might not be as secure with American hyperscalers as it is in their own safe. Telekom promises a fortress under German data protection law. A digital high-security wing for the crown jewel of German industry: our process know-how. And to make all this tangible for companies, there's the "Deutschland Stack." A kind of modular system where Telekom provides the infrastructure and SAP – who else – adds the appropriate business applications on top. First users like the robotics specialist Agile Robots and the simulation forge PhysicsX are already on board. This shows: this is not a paper tiger. The machines are already running.

What does this mean for your business in Bielefeld or Biberach?

Now I can already hear them, the managing directors from Sauerland and Allgäu: 'Good for them, Mr. Müller. But what does that have to do with my press brake?' A fair question. Until now, AI in manufacturing was a game for the big players. For corporations like Siemens, who maintain their own AI departments and can easily shell out a few million for a pilot project. For SMEs, often only two bad options remained: either send your most sensitive data to Amazon Web Services or Microsoft Azure and hope for the best – or try to cobble together a solution in your own server room that ultimately costs more than it brings and is hardly manageable by the IT department (consisting of one person and an apprentice).

This new AI factory is intended to be the third way. Access to absolute top-tier hardware (these Blackwell devices are the ultimate for AI training), without having to bear the initial investment of several million euros for purchase. And – this is the crucial point – without giving up control over your own data. Whether it's about predictive maintenance for your presses, optimizing robot arms on the assembly line, or creating a digital twin of your entire production hall to simulate new processes – the computing power for this can now come from Munich. With lower latency, under German law, and theoretically on a pay-per-use basis.

The availability of sovereign AI infrastructure is not a 'nice-to-have,' but a strategic necessity for Germany as an industrial location. It lowers the barrier to entry for SMEs to participate in data-driven value chains without losing control over their own know-how. We are shifting value creation back to Europe.

— Prof. Dr. Anja Weber, Head of the Fraunhofer Institute for Intelligent Production Systems Stuttgart

The Big Players Lead the Way – SMEs Must Follow Suit

Let's look at how the big players are operating. Siemens, one of the project's mentioned partners, isn't waiting for someone to roll out the red carpet. Years ago, they started to load their portfolio with AI. Just recently, they acquired Canopus AI, a small company specializing in AI for semiconductor metrology. Why? Because they know that production data is the new gold, and AI is the shovel to dig it up. Now they are integrating their own simulation tools – the software that BMW and co. use to plan their factories – directly into this Munich supercomputer. For Siemens, this is a logical evolution, a new sales platform for their digital products.

For SMEs, however, this is also a warning shot. If the big players offer their tools directly on the most powerful available hardware, the gap between those who use AI and those who don't will widen even faster. Until now, starting your own AI projects was a matter of money and courage. Now, it is increasingly a matter of pure will. The infrastructure is there. According to a recent VDMA survey, over 60% of mechanical engineers are still hesitant to adopt AI, often due to unclear ROI and security concerns. These are precisely the two hurdles that the Munich initiative aims to overcome. So, the excuses are getting thinner.

But Wait: Not All That Glitters Is Gold

Before we all start cheering and shoveling our data to Munich, as experienced practitioners, we should pause briefly and put on our critical glasses. There are a few points that were certainly not discussed so loudly at the opening ceremony. First, the lock-in. We are exchanging dependence on Microsoft or Amazon for a new dependence on Deutsche Telekom and SAP. Is that really better? A monopolist remains a monopolist – even if they speak German. The pricing for using this miracle hardware is entirely in their hands. It's difficult to get out of such an ecosystem once you're deeply involved.

Second, the costs. 'As-a-Service' always sounds so nicely flexible. But it can also become damn expensive once the meter for GPU usage starts running. Last week, I spoke with the COO of a medium-sized automotive supplier from Swabia. He told me: 'Mr. Müller, I don't need half an exaflop for a research project. I need a reliable solution that tells me when my milling machine needs maintenance. And one that I can afford and understand.' He hits the nail on the head with that. The danger is that this is like using a sledgehammer to crack a nut. The price tag for access to this elite hardware could simply be too high for many 'bread-and-butter' applications in SMEs. And the AI experts you still need to train the models and interpret the results don't grow on trees, even in Bavaria.

And then there's SOOFI – this prestige project of a European open-source language model with 100 billion parameters. Sounds great. A European answer to ChatGPT & Co. But the question must be allowed: Does a manufacturer of precision tools need an LLM that can write poems? Or does it not rather need a highly specialized model that recognizes wear patterns on tool tips from sensor data? The cart is being put before the horse here. One builds the biggest cannon in the world and then looks for a suitable target. Whether this is really the way for German SMEs to benefit at the core of their value creation – I put a big question mark there.

Plain Talk: What You as a Managing Director Should Do Now

• 1. Conduct a data inventory: Stop dreaming about AI and start doing your homework. Where is your most valuable data located? In the machine controls (PLCs)? In old Excel spreadsheets? In the CAD files of your designers? Without a clean, accessible data basis, any AI investment is wasted money. It's tedious, yes. But unavoidable.
• 2. Define a 'pain project': Don't choose the most glamorous AI project, but the one that hurts you the most. Is it the high scrap rate for component XY? The unexpected machine downtimes on Friday afternoons? Define ONE manageable problem where a solution would have clear, measurable value.
• 3. Calculate with a sharp pencil: Approach providers like T-Systems and ask for a proof-of-concept for exactly this one problem. Get a concrete offer. What does it cost to use the platform for three months? How many consulting days are needed? Only with hard numbers can you make a real decision.
• 4. Form a 'guerrilla team': Assemble a small, interdisciplinary team. One from IT, an experienced engineer from production, perhaps even a worker who knows the machine inside and out. These people must drive the project and bridge the gap between high-tech AI and the rough reality of the factory floor.
• 5. Ask the sovereignty question: Ask every potential partner the tough questions: Where exactly is my data located? Who has access? What do the contracts look like? And very importantly: How do I get my raw data and trained models back if I want to switch providers? Clarify the exit strategy before you get in.

Conclusion: End of Excuses, Beginning of Work

In my experience, we in Germany tend to philosophize endlessly about the perfect, risk-free 110% solution, while the competition overseas simply gets on with it. This thing in Munich is here now. It's not perfect, it carries risks – especially those of cost and lock-in. But it's a tremendous opportunity to close the technological gap that has opened up in recent years. The question is no longer whether we use AI in manufacturing, but how. And for the first time in a long time, we have an answer that isn't 'AWS' or 'Azure,' but 'Munich'.

I bet that the companies that now – cautiously – bravely venture small experiments on this platform will be ahead in three years. Not because they are training huge LLMs, but because they have learned to use their data for concrete improvements in production. The others will still be sitting in conference rooms discussing the risks – while their order books get emptier and the best engineers migrate to the competition. There's no getting around that.