Civo has named the first of 40 planned UK edge datacentres, and CNBC reports Anthropic and OpenAI are sounding out 20-30 MW deals in the UK and Nordics. Both point the same way: inference capacity is moving to smaller, powered sites that can be brought online quickly.
What was announced
- Civo announced on 22 September 2026 the first of 40 planned UK edge datacentres, in Hertfordshire, with a combined long-term target of one gigawatt across the network.
- The Hertfordshire site is due to open by March 2027 at 8 MW, rising to 38 MW on the same plot. Civo says five further sites totalling 150 MW are secured in Lancashire, Greater Manchester, London and Oxfordshire, with the rest planned over three years.
- Every Civo site is designed for direct-to-chip liquid cooling around NVIDIA Vera Rubin NVL72 racks, which combine 72 Rubin GPUs and 36 Vera CPUs with 288 GB of HBM4 per GPU. Civo is taking Vera Rubin orders now, with early-access pricing from 10 dollars per hour, alongside existing B300, B200, H100 and H200 capacity.
- Civo's own survey of 1,000 UK IT leaders found 73 percent rate AI sovereignty a strategic priority this year, up from 61 percent last year.
- CNBC reported on 18 September 2026 that Anthropic has sounded out 20-30 MW capacity deals in the UK and the Nordics, and that OpenAI has explored deployments of that size in the Nordics, alongside their multi-hundred-megawatt and gigawatt agreements.
- Structure Research told CNBC the appeal is speed to usable capacity: a few megawatts at an existing powered site can beat waiting for one large block, because many inference workloads can serve requests across separate smaller clusters.
- JLL figures cited by CNBC put inference at 9 percent of global datacentre workloads in 2025 against 14 percent for training, with inference expected to overtake training in 2027 and reach 37 percent by 2030.
The ETON view
Two signals in one week say the same thing. Training still wants gigawatt campuses, but serving models does not, and the buyers with the deepest pockets are now willing to take capacity in 20-30 MW slices if it is powered and available sooner. In the UK, where grid connections rather than GPUs set the timetable, that makes the existing and mid-sized site more valuable, not less.
For hosting providers and regional operators this is an opening, but only if the hall can take the load. Vera Rubin NVL72 is a direct-to-chip liquid-cooled rack, and Civo is designing every site around it from day one. Most UK colocation still runs at 6-12 kW per rack on air. Before chasing an inference tenant, be honest about which halls can host rack-scale liquid-cooled systems and which are better suited to air-cooled 8-GPU servers built on H100, H200, L40S or RTX PRO class cards, where power per rack stays manageable.
On the buying side, smaller sites change the procurement shape. Instead of one large order, you get a run of repeatable pods that need the same servers, NICs, optics, drives and spares delivered to several locations on a predictable lead time. That is where vendor neutrality matters: mixing current-generation GPU nodes with proven previous-generation servers for control plane, storage and CPU inference keeps each site's cost down, and a buyback route for kit displaced by each upgrade keeps the lifecycle honest. We can quote UK availability on specific configurations and hold spares against a multi-site rollout.
Related infrastructure
Category: Market · Vendor: NVIDIA · Technology: Vera Rubin NVL72, direct-to-chip liquid cooling, edge inference · Last verified: 24 Sep 2026 · ~3 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
