AI Colocation UK: What AI Workloads Require
AI infrastructure places fundamentally different demands on colocation facilities compared to standard server deployments. This guide explains what to look for and how to assess whether a UK facility can support your AI workload.
Why AI workloads are different
A standard enterprise server rack typically draws 5 to 15 kilowatts. A rack populated with NVIDIA H100 or comparable GPU servers can draw 30 to 80 kW or more. This order of magnitude difference in power density has profound implications for the facility infrastructure required: power distribution, cooling, physical structure, and fire suppression all need to be specified for high-density deployments.
Beyond power density, AI training workloads generate substantial network traffic. Moving large datasets from storage to GPU, distributing training across multiple nodes, and checkpointing model weights all require high-bandwidth, low-latency network connectivity. A facility well suited to traditional enterprise workloads may not have the network infrastructure to support large-scale AI training efficiently.
Power requirements for AI colocation
The most important technical specification for an AI-ready facility is the maximum sustainable power density per rack, expressed in kW. This is not the peak rating, which some facilities overstate; it is the power level the facility can sustain continuously without thermal or electrical overload. Ask for this figure in writing, and ask how it was determined.
For GPU racks drawing 30 to 50 kW, you will typically need dual A+B power feeds per rack (for N+1 or 2N redundancy), PDUs rated for 32A or 63A per phase, and a facility power design that allows adjacent racks to draw high power simultaneously rather than relying on statistical averaging. Many standard facilities are designed on the assumption that not all racks will draw their contracted maximum at the same time; AI facilities cannot make this assumption.
Cooling requirements
Cooling is the most significant constraint for AI colocation. Below approximately 20 kW per rack, standard air cooling is generally adequate. From 20 to 30 kW, supplemental in-row cooling and hot aisle containment can typically manage the thermal load. Above 30 kW, standard air cooling becomes marginal or insufficient; above 40 kW, liquid cooling or direct-to-chip cooling is generally required.
Liquid cooling options include rear-door heat exchangers (which cool air as it exits the rack without direct liquid contact with hardware), direct-to-chip cooling (where liquid cooling plates are attached directly to processors), and full liquid immersion. Each has different facility requirements and hardware compatibility constraints.
Use the free AI Readiness and GPU Colocation Checker to assess the cooling requirement for your specific workload.
What to look for in a UK AI colocation facility
When evaluating UK facilities for AI workloads, ask specifically for: maximum sustainable power density per rack (not peak rating), cooling technology available at your target density, whether liquid cooling is available and at what cost premium, confirmation of diverse fibre entry routes for training data movement, availability of high-bandwidth interconnects (100G or 400G), and whether the facility holds NVIDIA DGX Ready certification or has existing GPU tenants as references.
Frequently asked questions
What is AI colocation?
AI colocation refers to placing GPU servers and AI infrastructure in a third-party data centre facility. Unlike standard server colocation, AI workloads require high power density per rack (often 30 to 80 kW or more), specialist cooling (liquid or direct-to-chip), and high-bandwidth network connectivity for data movement during training.
Which UK data centres are AI ready?
A growing number of UK facilities are upgrading their power and cooling infrastructure to support AI workloads. The most capable facilities are typically in London (Docklands, Slough), but regional facilities in Manchester, Leeds and elsewhere are investing in high-density capability. When assessing AI readiness, ask specifically about maximum sustainable power density per rack, cooling technology, and whether the facility holds NVIDIA DGX Ready certification.
How much does AI colocation cost in the UK?
AI colocation in the UK is typically priced on a per-kW basis rather than per-rack, reflecting the high power intensity. London pricing ranges from approximately £150 to £250 per kW per month for AI-ready facilities, with premium pricing for liquid cooling capability and dedicated high-density zones. Regional pricing is typically 30 to 50 percent lower. Use the free AI Readiness Checker to assess your requirements.
What is NVIDIA DGX Ready certification?
NVIDIA DGX Ready is a programme that certifies data centre facilities as capable of supporting NVIDIA DGX AI systems. Certified facilities have been assessed for power, cooling, network and physical infrastructure requirements. DGX Ready certification is not mandatory for AI workloads but provides confidence that the facility has been independently verified for GPU compute suitability.
Looking for AI-ready colocation in the UK?
Use the free AI Readiness Checker to assess your requirements, or get in touch with Cagelab for help identifying facilities that match your GPU workload specification.