On-Prem

Systems

CoreWeave drops £1bn in UK datacenters – but don't expect the latest Nvidia magic just yet

Rent-a-GPU outfit's latest datacenters are packed to the brim with H200s


As the UK government reaffirms its aspirations to become an AI superpower, CoreWeave says two new GPU bit barns packed to the brim with Nvidia accelerators are ready for business.

Located in Crawley and the London Docklands and developed in collaboration with colocation providers Digital Realty and Global Switch, respectively, the two sites are part of a £1 billion($1.2 billion USD) investment by the New Jersey-based datacenter operator.

"The UK is an important market for CoreWeave, with our European headquarters here and further operational plans for the country," Mike Mattacola, chief business officer at CoreWeave, said in a canned statement. "We are pleased to be partnering with Digital Realty and Global Switch to deliver the next generation of AI infrastructure in the UK."

But before anyone gets too excited, these sites aren't outfitted with Nvidia's latest and greatest Blackwell chips. Those parts only began making their way to market in Q4 2024. Instead, the facilities are using accelerators based on Nvidia's older Hopper architecture, which first made its debut in late 2022.

CoreWeave opted for Nvidia's H200 and 400 Gb/s Quantum-2 InfiniBand networking. Introduced in 2023, the H200 is among Nvidia's most advanced GPUs of the Hopper generation, with up to 141 GB of speedy HBM3e memory good for 4.8 TB/s of memory bandwidth and nearly 4 petaFLOPS of sparse FP8 performance apiece.

However, with more than a terabyte of high bandwidth memory — eight GPUs — per server, CoreWeave's UK datacenters should be better equipped to support ever growing frontier models, which have stretched existing systems to their limits.

For instance, a single H200 system is now able to run Meta's largest model Llama 3.1 405B at full 16-bit resolution. Previously, the model would have had to be distributed across two nodes or quantized to 8-bits to run on an H100-based system.

Just how many of these GPUs are housed at the Crawley and London Docklands sites hasn't been disclosed. The Register reached out to CoreWeave for comment; we'll let you know if we can get any specifics. However, past deployments have typically involved clusters of 10,000 or more accelerators.

The two datacenters are the last in a string of large-scale deployments over the past two years and reflect CoreWeave's ambition to break into new markets. Last spring, CoreWeave expanded its operations across the pond, announcing London as the location of its European headquarters alongside an initial investment of £1 billion. In October, during the International Investment Summit, CoreWeave inflated that investment to £1.75 billion ($2.1 billion USD).

These efforts have been supported by a flood of venture capital and debt financing totaling billions of dollars amid the AI boom. CoreWeave currently operates 28 datacenters around the globe with plans to bring 10 additional bit barns on line by the end of 2025.

While the datacenter builder is breaking into the UK and European markets, the British government is doubling down on its efforts to establish itself as an AI superpower.

On Monday, the nation outlined the AI Opportunities Action Plan, which will see the government adopt 50 recommendations made by a venture capitalist to use AI to boost productivity by 1.5 percent a year. The economic impact of the strategy is expected to generate £47 billion ($57 billion) a year.

Aspirations aside, supporting the additional AI infrastructure necessary to see these plans to fruition may prove troublesome for the UK's National Grid, as its CEO has previously warned. ®

Send us news
2 Comments

Intel has officially missed the boat for AI in the datacenter

But it still has a chance at the network edge and the PC

Where does Microsoft's NPU obsession leave Nvidia's AI PC ambitions?

While Microsoft pushes AI PC experiences, Nvidia is busy wooing developers

Nvidia deprecates CUDA support for aging architectures

Maxwell, Pascal and Volta, oh my! But fear not, driver support is still safe

Intel knocked off global chip revenue top spot after rotten 2024

Missed the AI processor boat, split with CEO savior, lost #1 seat to Samsung

IBM seeks $3.5B in cost savings for 2025, discretionary spend to be clipped

Workforce rebalancing? Yes, but on the plus side, the next 12 months are all about AI, AI, and more AI

DeepSeek means companies need to consider AI investment more carefully

But Chinese startup shakeup doesn't herald 'drastic drop' in need for infrastructure buildout, say analysts

South Carolina's abandoned nuclear reactors positioned to fuel the AI datacenter boom

VC Summer units 2 and 3, abandoned in 2017, are looking for a buyer; owners say tech industry needs are a perfect fit

Who's afraid of DeepSeek's impact on AI hardware sales? Not AMD CEO Lisa Su

Predicts more efficient ML architectures will drive adoption, see Instinct sales accelerate, shares dive

What happens when we can’t just build bigger AI datacenters anymore?

We stitch together enormous supercomputers from other smaller supercomputers of course

Ubuntu upgrade had our old Nvidia GPU begging for a downgrade

If you need legacy drivers, you might want to keep your older OS version – possibly indefinitely

Mental toll: Scale AI, Outlier sued by humans paid to steer AI away from our darkest depths

Who guards the guardrail makers? Not the bosses who hire them, it's alleged

US AI shares battered, bruised, and holding after yesterday's DeepSeek beating

Nvidia says its chips are still needed, OpenAI says it'll keep buying them en masse, but shares are still down