How European Startups Are Quietly Solving AI's Biggest Bottleneck

Jan de Vries · 2026-09-21

Earlier this year, I wrote about the AI power crisis and argued that coordination now matters more than new capacity. Well, the numbers haven't gotten any better. It's a real problem. Research from the International Energy Agency (IEA) shows that global data center electricity demand is projected to roughly double by 2030, and it's growing more than four times faster than total electricity demand from all other sectors combined. Think about that for a second. Here's the twist, though. Much of the capacity Europe desperately needs already exists. It's just sitting there, unused, most of the time. The real unlock isn't just building more power lines—it's building smarter software that orchestrates the assets we already have. That's the headroom we can tap into years before new substations are even a possibility. After that first piece, my colleague Anna and I decided to dig in. We wanted to find out who's actually building this critical software in Europe right now. What we found was a small but fascinating group of companies, each tackling the problem from a different angle. They're working across what I see as four key efficiency layers: the grid, the facility, the compute itself, and the software that ties it all together. ### The Grid: Making Hidden Capacity Visible This is about seeing and using what's already on the network. Companies like Amsterdam's Sympower and Dublin-based GridBeyond (both have raised over $75 million) are turning data centers and big industrial users into flexible grid assets. They aggregate demand and bid it into energy balancing markets across Europe. It's like creating a virtual battery from a bunch of buildings. - Munich's Entrix is a seasoned player in Germany, optimizing battery storage and trading that flexible power. - Finland's Capalo AI runs a massive virtual power plant, having contracted over 1 gigawatt of battery storage. They've expanded from the Nordics into Eastern Europe. - London's Piclo operates a marketplace connecting energy assets to grid operators, while Vienna's enspired uses AI to automate trading on European power markets. What's really telling? None of these companies started as data center specialists. They built their tech for industrial and renewable energy flexibility first. That just shows how new—and urgent—the AI-specific version of this power crunch really is. ### The Facility: Optimizing the Physics Layer This is where the rubber meets the road, inside the data center walls. Darmstadt-based etalytics, a university spin-off, is a prime example. Their platform uses AI and physics-based digital twins to optimize cooling, heating, and power systems. Big names like Equinix and NTT use it. The results speak for themselves. NTT Global Data Centers cut cooling energy by 19%. Telehouse Germany managed a 10.4% reduction at a site they thought was already optimized. Across the board, etalytics reports savings of up to 50% for some customers. Here's why this is a game-changer for AI: a data center with a fixed power connection can take every kilowatt saved on cooling and shift it directly to powering more servers. Efficiency suddenly turns into sellable rack space, no new construction required. In 2025, Microsoft's venture fund, M12, led an $8.6 million extension to their funding round, and they've now opened for business in the U.S. And then there's Rotterdam's Gradyent, thinking even bigger. Their digital twin optimizes district heating networks in over 35 European cities. Their big idea? Routing a data center's waste heat into a city's heating grid to warm homes, instead of just venting it into the air. They raised about $30 million in 2025 to make it happen. ### The Compute Layer: Smarter Workload Management This is the recovery layer—getting the most out of the expensive hardware we've already got. Paris-based FlexAI launched with a $30.7 million Seed round. Their software places AI workloads across different types of hardware from various vendors and clouds. The goal? Freeing customers from being locked into a single generation of expensive GPUs. Bristol-based YellowDog takes a similar approach, intelligently scheduling workloads across on-premise, hybrid, and multi-cloud setups. It's all about maximizing utilization and avoiding expensive idle time. Look, the scale of the challenge is daunting. But the emerging solution isn't just about throwing more megawatts at the problem. It's a story of clever software, smart coordination, and startups finding new ways to use what we've already built. It's a quieter, smarter path forward, and it's being built right now across Europe.