TL;DR
CommonAI has chosen Kao Data’s campus in Harlow to house the UK’s Scaling Inference Lab, Data Centre Magazine reported on 21 October. Paid for by ARIA, the Advanced Research and Invention Agency, the lab will install at least five high-density GPU clusters across two technology suites, with all of them due to be running by early 2027 for UK startups to test new AI hardware and software at rack scale.
What the lab is for
The Scaling Inference Lab is meant to be neutral ground where emerging AI technologies can be tried out and measured against commercial systems. Startups get shared access to complete racks, so they can plug in their own technology and see how it performs on current workloads rather than on a small test bench.
Each cluster tests a different idea about AI infrastructure, such as pipeline parallelism, orchestration across mixed hardware, and optical circuit switching at nanosecond speeds. The hardware is deliberately varied: AMD Instinct MI350X accelerators sit beside a range of Nvidia hardware, from H200 NVL cards to HGX B300 systems, in servers from Supermicro, MSI, Fujitsu and Dell. New experimental clusters are planned every six months, running on certified renewable power.
The named technology partners so far are Oriole Networks, Callosum and Quettaflop AI.
Funding and location
ARIA put £50 million behind the lab when it launched it this year, and in June the government added more than £20 million to build a national capability in new AI compute systems.
CommonAI’s partnerships lead, David Stringer, said Harlow’s position between London and Cambridge made it a good base for welcoming “startups, investors and policymakers”. Spencer Lamb, Kao Data’s chief executive, linked the project to rising demand for “sovereign, homegrown AI infrastructure”.
Looking forward
Callosum, one of the partners, featured in our report on its seed round. In our view, the six-month build cycle is the detail to watch. If it holds, the lab could give UK founders a regular, publicly funded way to benchmark new chips and interconnects against current Nvidia and AMD systems. The first test of that will be whether all five clusters are live on schedule in early 2027.