AI-Ready Infrastructure: What It Actually Means for UK Data Centre Operators

Why AI workloads are fundamentally different

A standard enterprise server rack draws between 5 and 15 kilowatts of power. A rack of NVIDIA H100 GPU servers draws between 30 and 80 kilowatts, and in dense configurations can exceed 100 kilowatts. This is not an incremental difference in power consumption; it is an order-of-magnitude shift that changes every fundamental aspect of facility design. The power infrastructure, the cooling systems, the structural floor loading, and the connectivity architecture required to support AI workloads differ from standard colocation in ways that make AI capability a genuinely significant infrastructure investment.

When operators describe their facilities as AI-ready without explaining what that means in concrete terms, they lose credibility with the buyers who matter most. AI infrastructure buyers are technically sophisticated. They know the difference between a facility that has accommodated a couple of GPU servers opportunistically and one that has invested properly in the power density, cooling capacity, and connectivity to serve GPU clusters at scale. Read our AI colocation UK guide for the full buyer-side view of what these decisions involve. For the definitive technical specification reference, the NVIDIA H100 GPU specification gives the exact power and cooling figures buyers use when evaluating facilities.

Power density requirements in detail

Power density is the primary technical differentiator between AI-capable and standard colocation. The thresholds are broadly as follows. Below 20 kilowatts per rack, standard air cooling works reliably and most existing colocation facilities can accommodate the workload without significant additional investment. Between 20 and 40 kilowatts per rack, in-row cooling or rear-door heat exchangers become necessary to manage heat effectively. Above 40 kilowatts per rack, direct liquid cooling to the chip level is required for thermal management and energy efficiency.

For operators, the practical implication is that accommodating AI workloads is not simply a matter of accepting higher power draws. Each threshold requires different cooling infrastructure, different power delivery architectures, and different structural capacity. A facility that can reliably supply 40 kilowatts per rack to a customer with liquid cooling infrastructure is a materially different product from one that offers 10 kilowatts per rack with standard air cooling. Buyers who know what they need ask very specific questions about these thresholds. Read the high-density colocation UK guide for a detailed breakdown of how operators approach this segment and what investment is required at each density level.

Cooling infrastructure for AI racks

Cooling is where the operational complexity of AI-ready infrastructure concentrates. In-row cooling units are deployed within the data hall between or alongside racks, drawing hot exhaust air directly from rack rears before it mixes with the wider hall environment. This approach is effective up to around 30 to 40 kilowatts per rack, depending on the specific equipment and airflow management in place. Rear-door heat exchangers attach directly to the back of racks and use chilled water loops to absorb heat at the source, enabling higher densities without requiring room-level cooling redesigns.

For the highest density AI deployments, direct liquid cooling delivers coolant directly to processors, memory, and power electronics via rear-door or manifold systems that remove the bulk of heat without relying on air at all. This approach is increasingly required for NVIDIA H100 and H200 configurations and is becoming the baseline for any facility genuinely pursuing the premium AI infrastructure segment. Operators who have made this investment should communicate it clearly and specifically. Use the AI readiness checker tool to assess how a specific facility compares against AI workload requirements. Uptime Institute provides authoritative guidance on cooling approaches at uptimeinstitute.com/resources.

Network requirements for AI workloads

AI training requires high-bandwidth, low-latency interconnects between GPU nodes for data movement and gradient synchronisation. The bottleneck for large training runs is frequently the network, not the compute. Training clusters use InfiniBand or high-density Ethernet fabrics operating at 100 gigabits per second or higher, with single-digit microsecond latency requirements that are meaningfully different from standard enterprise network specifications.

AI inference workloads have different but equally demanding requirements. Real-time inference services, particularly those serving latency-sensitive applications, need network architectures that minimise latency between compute and the applications consuming model outputs. For operators, supporting AI training clusters requires facilities that can accommodate InfiniBand switch infrastructure, high-density patch connectivity, and the power density to support the interconnect hardware itself. See the NVIDIA InfiniBand networking documentation for the specific interconnect requirements that buyers bring to facility evaluations.

The NVIDIA DGX-Ready certification

The NVIDIA DGX-Ready Data Centre programme certifies that a colocation facility meets the power density, cooling, network, and physical infrastructure requirements to host NVIDIA DGX systems in production. It is not a simple self-declaration; NVIDIA assesses facilities against a defined technical specification before granting certification. Equinix and Digital Realty hold DGX-Ready certifications for several of their UK and European facilities. Most mid-market UK operators do not.

The commercial significance of this is that enterprise AI buyers, particularly large organisations deploying substantial GPU infrastructure, frequently use the DGX-Ready list as a first-pass qualification filter. A facility that is not on the list may be capable of supporting the workload, but it faces the additional friction of having to demonstrate this during the sales process rather than beginning from a position of certified credibility. Operators who can legitimately pursue DGX-Ready certification should treat it as a commercial priority. Those who cannot should be precise about what density they can support rather than using vague AI-ready language. See the NVIDIA DGX-Ready programme for the current list and certification criteria.

What operators need to invest to serve AI demand

An honest assessment of the investment required: power upgrades to increase available kilowatt capacity per rack, which typically involves working with the grid operator and potentially upgrading substation and distribution infrastructure; cooling infrastructure investment that may range from rear-door heat exchangers as a relatively accessible upgrade to full direct liquid cooling as a more substantial facility redesign; and connectivity infrastructure including high-density fibre, potentially InfiniBand switching, and the physical space for the associated equipment.

Not all existing sites can be upgraded economically to serve premium AI density requirements. Older facilities with structural or power constraints may find the investment unviable. But for operators with sites that can be upgraded, the commercial opportunity is significant. AI workloads command premium pricing, they generate longer contract terms as buyers commit to multi-year GPU deployments, and the market segment is growing faster than any other area of enterprise colocation demand.

How to communicate AI readiness to buyers

Even operators with partial AI capability, those who can support 30 to 40 kilowatts per rack with in-row cooling but not full liquid cooling for 80-kilowatt configurations, need to update their website copy and sales materials to use the language buyers are searching for. "AI-ready infrastructure", "GPU colocation", "high-density power" and "liquid cooling capacity" are the specific phrases that appear in buyer search queries and evaluation criteria.

Operators who have made meaningful investments in AI capability and are not communicating this clearly on their websites are systematically missing enquiries from the buyers they have positioned themselves to serve. Read the GPU colocation UK guide for the buyer-side view of what these searches are looking for and what a credible answer looks like. Then run the free visibility audit to see where your current search presence stands against the specific AI colocation terms that buyers are using.