Summary
- CloudGrid Energy and Policloud plan to deploy 280 distributed AI computing units across 16 European sites by the end of 2027.
- The network is expected to provide 29,000 GPUs, two million vCPUs, and 35MW of energy capacity across France, Germany, Italy, Spain, and Sweden.
- The model places modular compute near renewable generation, reducing dependence on conventional hyperscale siting while raising questions about operations, connectivity, and scale.
CloudGrid Energy and Policloud have signed a €580m framework agreement to deploy 280 distributed AI computing units across Europe, tying modular compute infrastructure directly to renewable energy and electric-vehicle charging sites.
The agreement covers 16 secured locations in France, Germany, Italy, Spain, and Sweden, with full deployment targeted by the end of 2027. Once complete, the network is expected to provide access to 29,000 GPUs, two million vCPUs, and 35MW of energy capacity. CloudGrid said the first unit is already operational at Bonne Voisine in the Aube region of France.
The project is set out in the companies’ joint release, which says the sites will mainly include renewable energy facilities, including solar plants, wind farms, and biomass plants, as well as EV charging infrastructure. Urbasolar, part of the Axpo Group, is listed as a deployment partner and will provide access to solar plant sites while handling installation, permitting, and grid-connection processes.
The project does not try to copy the hyperscale campus model. Each Policloud unit is described as a high-performance computing module connected to the Poligrid distributed network, integrating GPUs, CPUs, storage, water-free cooling, and high-speed connectivity. The companies say units can be deployed within a few months and scaled gradually, with local maintenance and a smaller land footprint than traditional data centres.
That shifts the constraint map. AI demand is pulling more capacity towards high-density campuses, but those campuses increasingly collide with grid queues, planning pressure, local environmental scrutiny, and long lead times for substations and transmission upgrades. Smaller compute modules sited beside generation do not remove power risk, but they can reduce exposure to congested demand centres where new grid capacity is already rationed.
Distributed compute follows the electrons
The physical logic is simple enough: where AI inference and some HPC workloads can run away from the largest metro clusters, compute can move closer to available energy. Renewable generation sites with established grid connections, controlled land, and long-term power agreements become candidates for digital infrastructure rather than only energy export points.
The operational model will be harder than the siting logic. Distributed computing still needs fibre, security, remote monitoring, spares, field engineering, orchestration software, and customer confidence. A 280-unit estate across five countries will need consistent asset management, cyber and physical security, fault response, replacement logistics, and power-quality management.
Cooling is part of the commercial claim. Policloud’s units are described as using water-free cooling, which gives the deployment a clearer environmental position in regions where data centre water use is under scrutiny. Water-free does not mean heat-free, and each site still needs to reject thermal load into the local environment. The suitability of that approach will vary by climate, density, neighbouring land use, and operating profile.
The 35MW figure is modest beside the largest European campus proposals, but the agreement tests a different approach to AI infrastructure. Capacity would be built in smaller increments across multiple energy-linked sites rather than concentrated in a single power-hungry campus. That could appeal to policymakers seeking sovereign compute capacity without relying entirely on a few billion-euro locations.
The buildability test
The strongest part of the proposal is the connection between compute deployment and sites with green power, controlled land, and existing grid routes. Those are now gating items for many large schemes. If CloudGrid can move through permitting and energisation by working with renewable infrastructure partners, distributed AI capacity becomes a practical development model rather than a branding exercise.
Modular infrastructure still has to earn confidence. Customers will want evidence on uptime, latency, data security, incident response, maintenance, energy sourcing, and workload orchestration across a distributed network. Investors will want to see whether the model can scale without creating a high-friction operating estate of small facilities with complex service requirements.
The agreement also sits beside the EU’s AI gigafactory plans. Gigafactories aim at very large clusters of frontier model training capacity. The CloudGrid-Policloud model appears better suited to decentralised AI, inference, and distributed enterprise workloads. The two approaches may become complementary, with training concentrated in the largest facilities and inference moving closer to energy, users, or available grid capacity.
The immediate test is execution. A €580m framework agreement and 16 sites create a credible pipeline, but the sector will judge the project by installed units, energised capacity, customer take-up, service availability, and cross-border operations. If the deployment holds together, it will strengthen the case that European AI infrastructure can be built around powered sites rather than waiting for every megawatt to pass through congested data centre hubs.
