Summary
- OpenAI and Anthropic are reported to be discussing 20–30MW data centre capacity deals.
- The reported European discussions include the Nordics and, for Anthropic, the UK.
- The smaller deployments are understood to be aimed at inference rather than large training clusters.
OpenAI and Anthropic are reportedly seeking data centre capacity in the 20MW to 30MW range, including discussions in the UK and Nordic markets.
DatacenterDynamics reported, citing CNBC, that the two AI companies are holding talks over smaller deployments likely to support inference workloads rather than the very large clusters associated with training frontier models.
OpenAI is reported to have discussed opportunities with providers in the US and Nordics, while Anthropic has held similar conversations in the UK and Nordics.
No individual operator, site, country, transaction value, delivery date, or power arrangement has been confirmed, and neither company has announced a signed European capacity agreement arising from the reported talks.
OpenAI told CNBC that it is building a diversified compute portfolio and evaluates infrastructure according to requirements including performance, reliability, timing, and cost, while declining to comment on specific commercial discussions.
Inference changes the geography of AI capacity
The 20MW to 30MW scale is material in conventional data centre terms but far smaller than the multi-hundred-megawatt and gigawatt-scale projects increasingly associated with AI training.
That distinction can change where capacity is useful. Training clusters benefit from extreme concentration of accelerators and high-speed interconnects, whereas inference demand is driven by running deployed models for users and applications. That can create a case for a more distributed infrastructure footprint, depending on latency, network costs, customer location, resilience requirements, and the economics of available power.
The Nordic markets have attracted substantial digital-infrastructure investment because of electricity availability, renewable generation, climate, and land. The UK offers a large customer and connectivity market but faces well-documented constraints around power availability and connection times in several established data centre clusters.
A 20MW to 30MW requirement therefore still presents a significant physical delivery challenge. Operators need energised capacity, suitable cooling for accelerator-heavy racks, network connectivity, electrical redundancy, and a credible commissioning timetable. Sites designed around traditional enterprise or cloud density may also require mechanical and electrical upgrades before they can accommodate AI hardware.
The reported discussions suggest that the next phase of AI capacity demand may not consist exclusively of giant dedicated campuses. Large training environments can coexist with smaller regional deployments used to serve models closer to customers and distribute operational risk.
For European colocation operators, that could broaden the addressable market for AI-ready wholesale capacity. It may also sharpen competition for the relatively small pool of sites that can combine 20MW-plus blocks of power with high-density cooling and delivery times measured in months rather than years.
The discussions remain unconfirmed commercial negotiations rather than committed projects. Until contracts, sites, or development partners are disclosed, the reported megawatt figures should be treated as an indication of demand being tested in the market rather than booked European capacity.
