Europe's AI infrastructure build-out is no longer a theoretical conversation about future demand. It’s a live test of whether the continent can actually engineer, power, and finance AI factories at the pace the demand requires.

This is the second piece in our two-part coverage of the inaugural Data Center World Europe, held October 13-15 in Vienna, Austria.

On the event’s day two keynote panel Cosmin Georgescu and Jonathan Berney approached that challenge from different vantage points. Georgescu, CEO and Founder of ClusterPower, builds the AI factories. Berney, Founding Partner at Accelerated Infrastructure Capital, has to find a way to finance them.

Power Is the Foundation, Not the Afterthought

Georgescu's ClusterPower became one of Europe's first live AI factories back in 2022, training large language models for Bay Area customers before most of the industry had settled on what an "AI factory" even meant. Describing the broader concept, he outlined a five-layer stack: applications, models, AI infrastructure, chips, and, underpinning all of it, energy. "If you don't have the megawatts, you are actually not able to produce AI," he said.

That principle now defines ClusterPower's 800 MW Romanian campus, split between grid offtake and on-site generation. The site draws on five separate power sources, including high and medium voltage grid connections, on-site generation, diesel backup, and battery storage, giving it what Georgescu described as deep redundancy.

ClusterPower also operates as a licensed Romanian energy producer, feeding power back into the national grid and providing balancing via their generation and storage infrastructure. "We are helping the energy back," he said, positioning the company as a net contributor to grid stability rather than a drain on it, a direct response to recent blackouts in Spain, Portugal, and at Heathrow Airport in the UK.

Berney has a name for the operators who aren’t actually ready: braggawatts, a term he credits himself and a colleague with coining, for capacity claims that are more promise than power. "There are so many people who want to jump on this bandwagon and say I've got 100 or 200 megawatts, and then you dig into it and what they've got is a promise of it in eight years' time," he said. "It's not real."

Berney sees the strains in the grids themselves. Many AI hotspots are "creaking" under demand, pushing operators toward new models entirely, including behind-the-grid generation, fuel cells, and self-supply — the same self-sufficiency Georgescu has already build into Romania. any people who want to jump on this bandwagon and say I've got 100 or 200 megawatts, and then you dig into it and what they've got is a promise of it in eight years' time," he said. "It's not real."

Financing a Market Without a Track Record

Power sets one hard limit on how fast this can scale. Financing sets another: Berney estimated that an 800 MW facility represents roughly $34 billion in investment and noted that one megawatt of GPU capacity alone reaching $40 million, a steep speculative bet for any uncommitted capacity.

Traditional data center financing relied on investment-grade hyperscaler tenants, 15-year contracts, and lengthy write-down periods. That model is breaking down. A new class of offtaker, the "neoclouds" such as CoreWeave and Nebius, is driving massive demand but lacks the credit ratings banks are used to.

Berney's response has been to finance AI factories as fully integrated packages, combining GPUs and data center infrastructure into a single deal structure designed to pay back within five years rather than fifteen. "The equity markets are there," he said. "The debt available is much more restricted, much more structured, and much more reticent to come to the party."

Georgescu sees the same shift from the asset side: returns now vary dramatically by hardware generation. Newer GPU generations can break even within a year, while older or lower-margin infrastructure layers require contracts of ten years or more. "It's a totally different risk profile. It's a totally different capital stack," he said.

From 15 kW Racks to Dedicated Buildings

That generational churn is also rewriting the physical floor plans. Georgescu recalled starting in 2021 and 2022 with 15 to 30 kW racks; today's deployments run at 120 to 125 kW, with 250 kW expected next year and 1 MW racks already under discussion. Air cooling, standard just a few years ago, is now "impossible" at these densities.

This shift has also ended the idea of flexible, multi-tenant facilities housing mixed hardware generations. Even a specialised operator like ClusterPower finds it difficult to run multiple GPU generations within the same environment, Georgescu said, let alone a generic provider managing it for multiple tenants.

"It would be very difficult for a provider to build a single environment that lets all types of machines run at the same time," he said. "That's why the usual way the market and the industry work today is dedicated halls, dedicated floors, or even dedicated buildings for a particular type of machine."

ClusterPower now designs each new facility around a specific anticipated hardware generation, built in close coordination with Nvidia ahead of public hardware launches, rather than building speculative capacity and hoping demand follows.

Sovereignty Is About Control, Not Geography

Both panelists pushed back on definitions of sovereign AI tied purely to data center location. For Berney, sovereignty is fundamentally about control: who holds encryption keys, administrative access, and the software control plane. He pointed to Stargate Norway, a 230 MW facility backed by Nscale and Aker with Microsoft stepping in to take GPU capacity after OpenAI exited its original agreement, as proof that European sovereign AI infrastructure was both possible and in demand.

Still, he was candid about the limits of that control. "The GPUs are still produced by Nvidia, a US company, or AMD, a US company. The biggest off takers are still Microsoft, Meta, etc. We're getting halfway there from a European and Asian perspective."

Georgescu was more pointed about Europe's regulatory lag. Sovereign AI’s momentum started in the U.S., he said, leaving other geographies uncertain about real demand. True sovereignty, in his view, depends less on encryption or physical location and more on whether AI capability actually benefits the local economy. "What is the local demand? How can you transfer the AI factory intelligence back to the companies, to the startups, to the institutions, to universities, to the government?" Berney agreed the bigger obstacle is uncertainty, not ambition: "Sovereign AI is tripping over itself with the level of regulation that's being discussed and needed and thought about, but not actually implemented."

What Would Actually Prove This Can Scale

Pressed on what evidence should convince a skeptical buyer that a 12-month delivery promise is credible, Berney pointed to a completed project: the Arc Data Centre in Weybridge, delivered by BladeRoom in 12 months from cleared site to a complete 19 MW facility. He noted the site already had power and infrastructure in place, which shortened the timeline, but set an aggressive new benchmark. "This time next year, our question is going to be different. It's going to be six months."

Georgescu's own standard for separating real capacity claims from inflated ones was blunter still. "The only way to assess a site is to believe that everybody is lying... I'm looking to energy proof when assessing if this is real or not."

Asked what would convince them the model could scale across all of Europe, Berney again pointed to Norway. "It proved absolutely that the ability to be able to deliver AI factories in Europe was not only a possibility, but actually a reality." Georgescu put it more simply: "If you did it once, you can do it more."

Data Center World Europe showed that every assumption defining this industry is being rewritten at once: grid-only power, fifteen-year leases, flexible multi-tenant buildings, location-based sovereignty. None of this is theoretical anymore. Investors and operators like ClusterPower and AIC are proving that these new deal structures are the new baseline, and any operator or investor still planning around old assumptions risks falling behind.