The Plug Can Be Pulled: A German Cluster Confronts AI as Infrastructure
By Nico Stino
Three months ago, the mayors of two small towns in the Lippetal region of North Rhine-Westphalia visited Professor Roman Dumitrescu, head of the it’s OWL technology cluster, with what seemed like a simple question. An American investor was preparing to spend somewhere between five and twelve billion euros on a new data centre near the A2 motorway, next to an RWE power plant. The mayors wanted to know how much AI computing capacity local industry would actually need by 2030. It was, Dumitrescu thought, a fair question — and also one nobody could answer. He asked around the cluster’s leadership for the following weeks. Nobody had a figure. Most had never been asked.
That gap, more than any single statistic, became the spine of Dumitrescu’s recent address to the it’s OWL cluster, delivered in the sweltering heat of a regional gathering in Ostwestfalen-Lippe. His central claim is blunt: within five years, German industry either becomes AI-enabled or it largely ceases to be competitive industry at all. Anything short of that, he argues, leaves companies dependent on computing infrastructure, models, and platforms controlled by others — a position he compares to calling the local utility to cancel one’s own electricity connection.
From chatbots to co-workers
Dumitrescu sketched a timeline of the current AI wave that will be familiar to close observers but is worth restating for its compression. ChatGPT’s late-2022 debut produced novelty use — greeting cards, jokes, little else. By 2023, professionals began quietly testing chatbots for real work. Late 2024 into 2025 brought the rise of early agentic systems capable of completing bounded tasks with minimal supervision. Now, he argues, the frontier has moved again: multi-agent systems that function less like tools and more like virtual colleagues, complete with defined skill profiles and task assignments, operating alongside a human who retains final judgement — a “human in the loop.”
The productivity data he cites, drawn from Anthropic’s own published research, is striking: engineering task performance has climbed steadily across difficulty tiers, with the sharpest gains concentrated in the past several months. Code output per engineer, he noted, has moved from roughly 1–1.5 times baseline productivity a year ago to something closer to eightfold today — a shift he says explains recent layoffs at firms like DeepL, which cut around 250 developers as software efficiency gains outpaced headcount needs, even after only modest workforce reductions.
The Fable episode
The talk’s most pointed illustration concerned Anthropic’s own product history. According to Dumitrescu’s account, Anthropic developed an internal model of unusual capability — one apparently proficient not just at patching software vulnerabilities but at discovering and exploiting them — and declined to release it publicly in that form. Additional safety measures were layered on, and the resulting product, Fable 5, launched with certain capabilities deliberately restricted. Within days, he said, a small team — he believes linked to Amazon — found a simple prompt that circumvented those restrictions. The U.S. government subsequently gave Anthropic a narrow window to patch the issue or restrict access for foreign nationals worldwide, including foreign employees of Anthropic itself and non-Americans residing in the United States — an enforcement problem Dumitrescu called nearly unworkable, since nationality isn’t something these systems track. He noted Anthropic has reportedly since considered building in some form of verification.
The episode briefly cut off European access to the model. Dumitrescu is careful not to overstate the novelty of what this exposed: the dependency was not created by the incident, he argues, but merely made visible. Europe’s compute buildout remains a fraction of what the United States and, by some accounts, China are deploying — Germany’s entire data-centre expansion through 2030, he noted pointedly, is roughly comparable in scale to a single facility Meta is building in New Orleans.
An engineering revolution, not a software one
Where Dumitrescu diverges from much of the AI commentary aimed at industry is in framing the moment specifically as an engineering revolution rather than a software one. For a region like OWL, whose economic base is manufacturing, mechanical and process engineering, and production know-how, he argues the relevant frontier is not chatbots answering questions but AI systems intersecting the physical world in two directions: humanoid robotics on one hand, and AI systems attempting to automate engineering work itself on the other.
He singled out Prometheus, a well-funded new venture — reportedly valued near $41 billion, with Jeff Bezos as CEO — pursuing what it calls an “Artificial General Engineer.” Dumitrescu is skeptical the approach will be straightforward: CAD and physical-world engineering tasks, he noted, tend to look deceptively simple and prove far more complex than they appear, a mismatch he expects such ventures will eventually confront. He also pointed to more unsettling data-collection efforts already underway — from camera-based productivity monitoring in workplaces to Meta’s reported capture of employee mouse movements and keystrokes as training data for physical-world AI systems, developments he called useful for building models but, on a personal level, “a little unsettling.”
The case for building, not deliberating
Dumitrescu’s closing argument is less about technology than posture. He lists what he sees as the region’s structural advantages — deep industrial complexity, genuine understanding of physical-world constraints, high latent productivity gains still uncaptured in industrial processes, strong technological appetite among the German public, and an already well-networked cluster structure. Against these, he sets four risks: underestimating the pace of change, over-indexing on regulatory caution before even establishing whether rules actually apply, remaining passive technology consumers dependent on providers who — as the Fable episode demonstrated — do not fully control their own products either, and simply thinking too small.
His prescription is deliberately unglamorous: rather than spend six months drafting strategy documents, the cluster intends to build working AI infrastructure now and reassess in six months. “The question isn’t who has the highest compute or the best benchmark score,” he said. “It’s about being sovereign, reliable, and genuinely effective.” His closing line doubled as his talk’s title in spirit: not let’s write something, but let’s build something.
About Techvox
Techvox is the English-language technology and AI desk of Swissvox, an independent Swiss media platform funded solely by its readers — free of party affiliation, advertisers, or state support. Techvox covers technology and AI developments with a particular focus on Switzerland and Asia. Readers who want to support independent tech journalism can do so through Techvox’s membership model.