What are Civo's 40 UK edge data centres?

When people hear that Civo is building a gigawatt of AI capacity in the UK, they usually picture one thing. A vast campus in a field somewhere. Hundreds of thousands of GPUs under one roof. The kind of site that gets announced with a ministerial visit and a five-year timeline.

We are doing the opposite.

We are building 40 data centres, spread across the country, each one close to the people and organizations that will use it. On 22 September we unveiled the first one, in Hertfordshire.

That is a deliberate choice, and it comes down to where AI is going next. This post explains what the network is, why we built it this way, and what it means if you are trying to run AI in Britain.

Watch on YouTube

Civo unveils first of 40 UK edge data centres

Civo's edge network is 40 liquid-cooled data centres across the UK, adding up to a combined one gigawatt of high-performance capacity. Every site is built for sovereign AI compute and designed around NVIDIA Vera Rubin. The remaining sites will come online over the next three years.

Why the edge, and not one big campus?

The first wave of AI was about training. Training is patient. You can send a job to a giant cluster on another continent, wait, and collect the model at the end. That is why so much of the world's compute ended up concentrated in a handful of enormous sites, a long way from most of the people using it.

Inference is different. Inference is the model actually doing its job, every time someone asks it something. An agent approving a transaction. A robot on a production line deciding where to move. A vehicle that cannot pause mid-junction to wait for an answer from Virginia.

And inference is now most of the work. Deloitte predicts inference will account for roughly two-thirds of all AI compute in 2026, and expects most of it to keep running in data centers rather than on devices. So the real decision is not whether AI needs data centers. It is how far away they are.

Agents, robots, and vehicles cannot wait on a round trip across the Atlantic. They need compute near the device making the request. If inference is the future, and we think it is, the infrastructure has to follow it. Not one campus. A network.

What is inside each site? 

Every site is built for direct-to-chip liquid cooling. Coolant runs across a plate in direct contact with the silicon. That is not a nice-to-have. As GPU density climbs with each generation, air simply cannot carry the heat away fast enough. Daniel Bizo, research director at Uptime Intelligence, says that above 50 kW per rack, air cooling becomes “increasingly impractical” and direct liquid cooling becomes mandatory. Yet Uptime's 2025 survey found 82% of operators had no rack running above 30 kW, while a current NVL72-class system can need roughly 140 to 160 kW per rack.

Liquid cooling has quietly become the real constraint on AI infrastructure. A building designed around air will struggle with what comes next, and if a facility cannot cool the hardware, it does not matter how many GPUs you have ordered.

So none of this is built on assumption. Our in-house design team runs full computational fluid dynamics modeling on every site before construction starts. We know how a facility behaves at full load before a single module ships. And when the next generation of silicon arrives, the modules scale, not the building.

The first hardware those modules are designed around is NVIDIA Vera Rubin, the successor to Blackwell, with up to five times the inference performance. A single Vera Rubin NVL72 rack brings together 72 Rubin GPUs and 36 Vera CPUs, with 288GB of HBM4 memory per GPU. Civo will be among the first cloud providers in the UK to offer it. We are taking orders now, with early-access pricing from $10.00 per hour.

You do not have to wait for Vera Rubin to start. NVIDIA B300, B200, H200 and H100 capacity is available on Civo today, on the same open, interoperable stack we built in-house.

There is one more thing these sites produce. Heat. Rather than waste it, the design allows it to be recovered and fed into district heating networks for homes, factories and greenhouses near each site. A data center that gives something back to the town it sits in.

Why does distributed mean sovereign?

There is a second reason we chose 40 sites over one, and it matters as much as latency.

Concentration is a risk. Put a country's AI capacity in one place, or at the end of someone else's supply chain, and you have created a single point of failure. The UK government already treats this seriously. Data centers were designated Critical National Infrastructure in 2024, alongside energy and water.

The risk is not theoretical. In our Digital Sovereignty Revolution 2026 research, 39% of UK IT leaders had experienced an outage originating from a US hyperscaler in the past year. A national network spreads that exposure. When geopolitical pressure reaches infrastructure, and it increasingly does, capacity that is distributed and under British control is capacity that stays available.

That is what sovereignty means to us. Not a UK region of a foreign cloud. Not a contract clause. Compute that is built here, owned here and governed here, close to the organizations that depend on it.

It also gives international organizations a secure, compliant UK home for their services and data, rather than a UK address on someone else's platform.

Britain has the ideas, now it has somewhere to run them

Britain has never been short of people with ideas. Researchers who wait months for compute time. Enterprises putting agents into production. Smaller companies priced out of building at this scale.

What they have lacked is British compute at the scale those ideas demand, close enough to actually use.

That is what the 40 sites are for.