The startups fixing Europe’s AI power problem

Earlier this year, I wrote about the AI power crisis and argued that coordination now matters more than new capacity. The numbers have not improved. Research from the International Energy Agency (IEA) shows that global data centre electricity demand is projected both to roughly double, from 485 TWh in 2025 to around 950 TWh by 2030, and to grow more than four times faster than total electricity demand from all other sectors.

Much of the capacity Europe needs already exists but sits unused most of the time. Software that orchestrates existing assets unlocks that headroom years before new lines and substations can.

Following the first piece, my colleague Anna Trendewicz and I looked into who is actually building this software in Europe. We found a small group of companies working across the four efficiency layers: grid, facility, compute and software.

Grid: making existing capacity visible

Amsterdam’s Sympower and Dublin-based GridBeyond have each raised more than €70 million. Both turn data centres and industrial energy users into flexible grid assets by aggregating their demand and bidding it into balancing markets across multiple European countries.

Munich’s Entrix ranks among Germany’s more established battery storage optimisers and trades flexible capacity across day-ahead, intraday and balancing markets.

Finland’s Capalo AI runs a virtual power plant that had contracted more than 1 GW of battery storage by the end of 2025 and has since then expanded from the Nordics and Baltics into Poland and Bulgaria.

London’s Piclo operates a flexibility marketplace that connects distributed energy assets with grid operators, and Vienna’s enspired automates AI-driven trading of flexible assets across European power spot markets (Note: Future Energy Ventures is an investor in both Piclo and enspired).

None of these companies started out as a data centre specialist. Most built their platforms for industrial and renewable flexibility, which shows how new the AI-specific version of this problem still is.

Darmstadt-based etalytics, a spin-off from TU Darmstadt, optimises the cooling, heating and power systems inside a facility. Its etaONE platform combines AI with physics-based digital twins and model predictive control, and operators can either let it adjust set points automatically or approve its recommendation manually. Data centre customers include Equinix, Digital Realty, NTT and Telehouse.

NTT Global Data Centers cut cooling energy by 19% with the software, and Telehouse Germany cut cooling electricity by 10.4% at its Frankfurt campus, a site it already considered well optimised. Across its customer base, etalytics reports savings up to 50%.

In 2025, M12, Microsoft’s venture fund, led an €8 million extension that brought its Series A to €16 million, and etalytics opened its U.S. business this year.

For AI operators this matters beyond the energy bill. A site with a fixed grid connection can shift every kilowatt it saves on cooling to its IT load, which turns efficiency directly into sellable rack capacity without a new site or new interconnection.

Rotterdam’s Gradyent works one step beyond the fence line. Its real-time digital twin optimises district heating networks for energy companies in more than 35 European cities, and it raised €28 million in a Series B in 2025. The platform can route a data centre’s waste heat into a city’s heating grid, where it warms homes instead of being vented.

Paris-based FlexAI launched in 2024 with a €28.5 million Seed round led by Alpha Intelligence Capital, Elaia and Heartcore, Bpifrance participating. Its software places AI training and inference workload across heterogeneous hardware from different vendors and clouds, so customers are not tied to a single GPU generation.

Bristol-based YellowDog schedules workloads across on-premises, hybrid and multi-cloud fleets in a similar way.

Model efficiency cuts energy demand before a workload reaches the rack. San Sebastián-based Multiverse Computing compresses large language models by up to 95%, using quantum-inspired tensor networks. It raised €189 million in 2025 and announced a Series C of up to €500 million in July 2026.

Pruna AI, based in Munich and Paris, and backed by EQT Ventures, has open-sourced an optimisation engine that combines pruning, quantisation and other compression methods to make models smaller and cheaper to run.

In the US, incumbents are already buying this layer. Nvidia acquired OctoAI in September 2024, and Red Hat agreed to buy Neural Magic two months later. Both deals show how early this sub-layer is and how quickly it is consolidating.

The interface: where the value sits

Almost none of the companies work across layers. A flexibility aggregator does not manage GPU orchestration, and a workload schedule does not see grid prices. etalytics comes closest: beyond cooling, its platform shaves peak loads and schedules on-site storage and generation against market prices, which links facility operations to the grid. No one yet connects that to compute.

The most valuable position in this stack sits at the seams, where one system makes grid, facility and compute decisions together instead of separate vendors reporting to separate budget owners. No European company holds that position today. For founders, it is the clearest and most defensible opportunity still open.

According to Stanford’s AI Index Report, the inference cost of running a model at GPT-3.5 level fell more than 280-fold between November 2022 and October 2024, mostly through efficient models and software. Yet most AI infrastructure capital still flows into physical buildout.

In the U.S., generation can theoretically at least be revived or expanded. In Europe, where grid connection takes years longer than construction, capital that ignores the software layer is betting that the connection queue clears on schedule.

The companies above show that Europe is already building the underlying technology. From an investor’s perspective, the window to back the layer that gets more out of Europe’s existing infrastructure is still open.