An AI data centre starts with a question: what work will the systems actually perform? GPU configuration, power supply and cooling follow from that answer. Planning these areas separately risks expensive bottlenecks, even when the building still has plenty of rack space.
Why rack count alone tells you little
Six racks can house compact business IT or a high-density GPU cluster. The footprint may be similar, but power, heat removal and operating requirements differ considerably. A feasibility study therefore needs more than a rack count: proposed systems, load profiles, expansion targets and availability requirements.
NVIDIA specifies approximately 120 kW for its DGX GB racks, with liquid cooling through cold plates on CPUs and GPUs. Other components continue to use air cooling. This is a specific manufacturer example of high power density, not a standard consumption figure for every AI rack. Source: NVIDIA hardware documentation.
Local knowledge search with limited concurrency can start much smaller. A system serving many simultaneous users, large models or intensive processing needs a different design. Start with the use profile and measurable requirements for response time, throughput and quality.
Distinguish IT power from the required electrical connection
IT power describes the demand of computing, storage and networking systems within a defined measurement boundary. The whole facility also needs energy for cooling, power conversion and other infrastructure. “120 kW per rack” therefore does not mean a 120 kW grid connection is sufficient for the complete site.
A simplified example illustrates the difference: six racks at 120 kW each give 720 kW of IT power. Applying an assumed total-supply factor of 1.25 gives 900 kW. This is a preliminary planning assumption, not a guaranteed PUE or a completed connection design. Reserves, start-up behaviour, permissible load steps, redundancy and other consumers require separate assessment.
Installed power and average energy consumption are also different quantities. A connection must support the intended operating conditions, while a business case needs a realistic load profile over time. Assuming continuous maximum load can distort costs just as much as an overly optimistic utilisation assumption.
Direct-to-chip cooling is a system, not a single component
With direct-to-chip cooling, cold plates absorb heat close to the processors. However, planning data centres requires consideration of the entire path: from the chip through manifolds and coolant distribution to building services and, ultimately, heat rejection or useful heat delivery. Each interface needs defined operating limits and responsibilities.
Items to establish include supply and return temperatures, flow rates, pressures, water quality, material compatibility and permissible contamination. A cooling distribution unit, or CDU, can separate the IT circuit from the building circuit. Whether and how to provide that separation must suit the chosen system.
Leak detection, isolation and maintenance also belong in the design. A component failure should not be the first time you discover that replacing it requires an entire area to shut down. Access and replaceability are as much part of a professional solution as nominal cooling capacity.
Account for the remaining air-cooling load
Liquid-cooled CPUs and GPUs do not automatically eliminate room air-cooling requirements. Depending on the hardware, power supplies, switches, storage and other components still release heat into the room. Include this residual load in the airflow and room design.
In practice, airflow must suit the rack layout, hot air must not mix uncontrollably with supply air, and cooling must function during maintenance and partial operation. Simply adding up liquid-cooling capacities does not answer these questions.
Design redundancy around the operating objective
AI use cases do not all require the same availability. An internal assistant that can tolerate a short interruption needs a different operating strategy from a time-critical production function. Describe the objective as a clear scenario: which task must continue through which failure?
This provides a basis for considering UPS, backup supply, pumps, distribution, network paths and data storage together. Redundant components only help if an unintended common point of failure does not defeat them. Recovery, orderly shutdown and incident responsibilities are equally important.
Consider growth and heat reuse early
A modular layout can make expansion easier, but spare floor positions alone do not provide expandability. Connection capacity, cable and pipe routes, cooling circuits, structural loads and maintenance space must also support the next planned step. Make these reserves visible in the options assessment and evaluate their cost.
Temperature, availability over time and a suitable recipient determine the potential for heat reuse. Nearby heat demand is a starting point, not yet evidence of a viable arrangement. Distribution routes, transfer equipment and seasonal differences need assessment. This turns a general idea into a technically testable concept.
What a sound preliminary design should deliver
Preliminary design should produce a coherent picture: prioritised use cases, load and expansion profiles, defined measurement boundaries, a supply schematic, a cooling concept and key site conditions. Assumptions and outstanding evidence must be clear so later decisions can build on them.
- IT: models, memory requirements, concurrency, networking and performance targets.
- Infrastructure: electrical power, liquid and air cooling, reserves and redundancy.
- Site: installation, structural capacity, service routes, grid connection and heat rejection.
- Operations: monitoring, maintenance access, alarms, responsibilities and recovery.
High Knowledge brings these areas together in the feasibility study and in DC-Planning. Blue DC A | Smart AI describes our approach to local AI based on compact data centre infrastructure. The key is a design in which the application, site and operations fit together.
