GPU Colocation UK
GPU servers require fundamentally different facility specifications from standard compute. This guide covers what to look for when placing GPU infrastructure in a UK colocation facility.
GPU server power requirements
Modern GPU servers draw substantially more power than standard CPU servers. An NVIDIA H100 PCIe card draws approximately 350W; the SXM5 version draws up to 700W. A fully populated 8-GPU server can draw 3,000 to 5,000W. A rack of eight such servers reaches 24 to 40 kW, which is at or beyond the capability of most standard air-cooled facilities.
Before selecting a UK colocation provider for GPU workloads, calculate your expected power draw per rack using the Rack and Power Requirements Builder, then use that figure to filter facilities based on their maximum sustainable density.
Cooling for GPU racks
The heat generated by GPU racks must be removed effectively to prevent thermal throttling, which reduces performance, and component damage. At densities below 20 kW per rack, well-designed air cooling systems with hot-aisle containment can manage GPU heat loads. Between 20 and 30 kW, in-row cooling supplementing standard CRAC/CRAH units is typically needed. Above 30 kW, rear-door heat exchangers or direct-to-chip cooling are required at most UK facilities.
Network requirements for GPU colocation
GPU workloads vary significantly in their network requirements. Inference workloads serving requests in real time require low-latency external connectivity but relatively modest bandwidth. Training workloads, particularly distributed training across multiple GPU nodes, require both high-speed node-to-node interconnect (InfiniBand NDR or 400G Ethernet within the facility) and high external bandwidth for data ingestion and model distribution.
When evaluating UK facilities for GPU training, ask about the internal switching infrastructure, whether InfiniBand is available, and what the uplink bandwidth options are from your rack.
Colocation vs cloud for GPU workloads
For sustained, predictable GPU workloads running at high utilisation, colocation is typically more cost-effective than cloud GPU instances over a three-year period. Cloud GPU pricing (AWS p4de, Azure NDv4, Google A3 etc.) carries a significant premium over the underlying hardware cost, reflecting the provider's infrastructure, management and margin. At consistent utilisation above roughly 60 to 70 percent, owned hardware in colocation usually wins on cost.
The trade-off is lead time and capital: GPU servers have multi-month delivery lead times, and the capital commitment is substantial. Cloud offers immediate availability with no capital commitment, which makes it better suited to development, proof-of-concept and variable workloads.
Frequently asked questions
Can standard UK colocation facilities host GPU servers?
Standard facilities can host GPU servers at lower densities (below 20 kW per rack) without modification. At higher densities typical of modern GPU racks, most standard facilities reach their power or cooling limits. Always confirm the maximum sustainable density before signing a contract.
What network bandwidth do GPU servers need in colocation?
For GPU inference workloads, standard 1G or 10G connectivity is often adequate. For distributed training across multiple GPU nodes, you need high-speed interconnect between nodes (typically InfiniBand or 100G Ethernet) and high uplink bandwidth for data ingestion. Confirm the facility's internal interconnect and uplink capacity.
How do I compare GPU colocation pricing in the UK?
GPU colocation is typically priced per kW per month. Compare on a total cost basis including power, connectivity, any premium for high-density or liquid cooling capability, remote hands rates, and minimum contract term. Use the Colocation vs Cloud Calculator to compare against cloud GPU pricing over three years.
What is the difference between GPU colocation and GPU cloud?
GPU cloud (AWS, Azure, Google, CoreWeave, Lambda Labs etc.) provides GPU compute as a managed service with no hardware ownership and pay-per-use pricing. GPU colocation means you own the GPU hardware and place it in a facility that provides power, cooling and connectivity. Colocation is typically more cost-effective for sustained, predictable GPU workloads; cloud is more cost-effective for short-term, bursty or development use.
Planning a GPU deployment in a UK data centre?
Use the AI Readiness Checker to assess what your workload needs, or contact Cagelab for help identifying facilities with the right specification.