Why a German town of 16,500 matters for personal AI agents
Europe's most credible answer to the brutal economics of personal AI agents is a family-owned company in a Bavarian town of 16,500 — and even it is getting squeezed.
A collage of circuit boards interwoven with aerial photographs of water treatment facilities, agricultural lands, industrial sites and mining machinery.Image: Sinem Görücü / Better Images of AI / CC BY 4.0
In March, NVIDIA chief and AI god Jensen Huang anointed OpenClaw as "definitely the next ChatGPT." He might be right. What he didn't mention is that the economics of running the next ChatGPT are brutal — and Europe's most credible answer to that problem is a family-owned company in a Bavarian town of 16,500.
That company is Hetzner, a cloud-hosting provider headquartered in Gunzenhausen, 45 kilometres south-west of Nuremberg. It is, by some margin, the most globally significant thing to come out of Gunzenhausen since the Romans left their border wall there some 1,600 years ago.
First, the cost problem. OpenClaw is an AI bot that lives in your messaging app and does tasks like scheduling meetings and checking the weather. It also burns tokens aggressively — users routinely report spending $20–50 in their first few hours of testing; one blogger spent $3,600 in a single month stress-testing the tool. Most of those dollars don't go to OpenClaw. They go to whoever runs the servers underneath — overwhelmingly Amazon, Microsoft and Google.
Hetzner sells the alternative. Its virtual private servers — slivers of a bigger machine — start at €3.49 a month. Independent benchmarks put AWS and Azure at four to ten times more expensive for equivalent compute, with one analysis finding Hetzner delivers roughly 14× the value-per-compute-unit of AWS. That has made it a favourite of the tinkerer class — developers running their own agents without haemorrhaging money — and the closest thing Europe has to a working sovereign cloud.
Which matters, because Europe is trying to build one.
The €25 billion problem
The proposed EU Cloud and AI Development Act is meant to grow European compute capacity. The numbers it's working against are stark. Amazon, Microsoft and Google control more than 70 percent of the European cloud market. The US has at least four times Europe's data-centre capacity. One recent analysis estimated that building a European cloud provider at just 5 percent of AWS's scale would require roughly $25 billion.
For context: the most-cited European contender is Evroc, a Swedish start-up that raised $55 million. The next is StackIt, built by the Schwarz Group — the company behind Lidl. Yes, a supermarket. Schwarz has committed €11 billion, which sounds impressive until you note that a single hyperscale data-centre can cost $2–4 billion to build. €11 billion buys you maybe three or four. AWS has more than 100.
Hetzner sits in the gap between these ambitions and reality. It owns its data centres in Germany and Finland, posted revenues of roughly €447 million in its last reported figures, and — unlike Evroc and StackIt — has been running at scale for nearly three decades. While Europe's would-be challengers are still raising money or breaking ground, Hetzner is already serving customers.
The squeeze
Then last month, the squeeze. Hetzner announced price hikes of 30–50 percent effective 1 April 2026, citing a 500 percent increase in RAM and SSD costs since September 2025 — driven by AI data-centre demand from the very hyperscalers it positions itself against. The irony is exact: the boom that makes a cheap sovereign cloud so valuable is also the boom inflating the cost of building one.
For Hetzner's customers, the increase still leaves it far cheaper than the American incumbents. But it punctures the idea that Europe can simply price its way to independence. The components — the memory, the storage, the accelerators — are sold into a global market that the hyperscalers dominate as buyers. A small Bavarian firm has no leverage over that market. Neither, yet, does the European Union.
The lesson Brussels keeps relearning is that sovereignty is a stack, not a single layer. You can own the data centre and still rent the chips. You can subsidise the cloud and still depend on foreign memory. Hetzner has gone further up that stack than almost anyone in Europe — and even it is exposed at the bottom of it.
For the executives reading this, the practical point is not that Hetzner will save you money, though for many workloads it will. It is that the question "where does our compute actually come from?" now has strategic weight. The firms that can answer it — and that have an alternative to the three American clouds — will have optionality the rest do not.
This is the first in a series on Europe's compute stack. Next: the chip layer, and why the most important European AI company may be one that makes no AI at all.
Continue reading
This investigation is for Second Chapter members. Join the executives, founders and policymakers reading our reporting from inside the room.
From €250/year. Become a member to read in full.
The conversation
14 comments · 3 from our teamSecond Chapter is a multi-directional resource. Comment, challenge our reporting, or suggest an angle for the next piece in this series — our journalists are reading.
The next piece in this series looks at the chip layer — the part of the stack even Hetzner doesn't control. If you've made a recent build-vs-rent decision on European compute, good or bad, I'd genuinely like to hear it. Reply here or message me directly.
We moved a chunk of inference workloads off the big three to a European provider last year, largely on cost. The Hetzner price hike is a useful reminder that "sovereign" doesn't mean "insulated" — the memory market is global. Would value a piece on hedging that exposure.
Noted, Annika — that's a strong suggestion. Flagging the hedging angle to the Advisory Group for the next theme cycle.
The €25bn-for-5%-of-AWS figure is the whole story, honestly. Until that gap closes, European "sovereignty" at the cloud layer is aspirational. The interesting bets are one layer down — silicon and memory. Hope the series goes there.