Rack power density has historically crept upward in small increments — a few kilowatts per generation, absorbed gradually by incremental cooling upgrades. AI workloads have broken that pattern. Current-generation GPU platforms, such as NVIDIA's GB200 NVL72, draw in the order of 120–132 kW per rack, with next-generation architectures expected to exceed 200 kW. That is not an incremental change; it is an order-of-magnitude shift that touches nearly every engineering discipline involved in data center design.
Electrical Distribution Has to Be Rethought From the Switchboard Up
Distribution architecture designed around 10–15 kW racks does not scale linearly to 120-plus kW. Bus duct sizing, busway capacity, power distribution unit configuration, and breaker coordination all require re-engineering, not simple duplication. AI training workloads also impose highly synchronised, transient load patterns across thousands of accelerators — clusters can swing significant load within seconds as training jobs start, pause, or checkpoint, which has real implications for UPS sizing and switchgear response characteristics.
Cooling Moves From a Building System to a Rack-Level Requirement
Air cooling's practical ceiling sits at roughly 15 kW per rack in well-optimised hot-aisle/cold-aisle configurations, extending to perhaps 40–70 kW with rear-door heat exchangers. Beyond that, direct-to-chip liquid cooling becomes the default design assumption, not an option layered on afterward. This affects floor plate layout, the routing of primary and secondary coolant loops, coolant distribution unit placement, and the building's overall mechanical plant sizing.
- Cold plate systems require careful integration with server and rack OEM specifications, since coolant connections are no longer purely a facilities decision
- Facility water temperature setpoints increasingly target 35–45°C "warm water" operation, enabling dry coolers rather than energy-intensive mechanical chillers for much of the year
- Leak detection, fluid quality monitoring, and maintenance procedures require operational training that differs meaningfully from traditional air-cooled facilities
Structural and Spatial Assumptions Change Too
Liquid-cooled, high-density racks are typically heavier and require piping infrastructure that did not exist in earlier facility generations. Floor loading calculations, raised-floor or slab-on-grade decisions, and the routing of primary cooling loops from rooftop or yard-level heat rejection equipment down to the data hall all need to be considered at the earliest design stages — retrofitting structural capacity later is materially more expensive than designing for it from the outset.
The building that can house a 15 kW rack and the building that can house a 140 kW rack are, in almost every engineering respect, different buildings.
Commissioning and Validation Are Becoming Their Own Discipline
High-density, liquid-cooled facilities also demand a more rigorous commissioning process than earlier generations of data centers. Coolant loops must be flushed, pressure-tested, and validated for water quality before IT equipment is exposed to them. Coolant distribution units require integration testing with both the facility's primary loop and the rack-level secondary loop. Leak detection systems need to be commissioned and validated under realistic operating conditions, not just on paper. Skipping or compressing these steps to meet a delivery deadline is one of the more common — and costly — failure modes in high-density facility delivery, since a coolant leak in a 130 kW rack carries far higher consequence than an equivalent failure in a 10 kW air-cooled environment.
This has implications for project scheduling: commissioning for high-density, liquid-cooled facilities should be planned as a distinct, adequately resourced phase of the programme, not compressed into the final weeks before handover as has often been the practice for conventional air-cooled data centers.
Workforce and Operational Skills Have to Keep Pace
Engineering design is only half of the high-density compute challenge. Facility operations teams need different skills to maintain liquid cooling systems safely and effectively than they needed for purely air-cooled environments — understanding of fluid dynamics, coolant chemistry, and leak response procedures, for example. Developers and operators who treat this as a training and staffing question from the outset, rather than discovering the skills gap after a facility is already operational, tend to achieve materially better reliability outcomes in the first years of operation.
Designing for the Generation After Next
Because GPU power draw has climbed sharply across recent hardware generations, facilities designed only for today's density risk premature obsolescence. The more resilient approach is to design core infrastructure — electrical risers, structural capacity, piping headers — with headroom for at least one additional hardware generation, while keeping final fit-out modular enough to adapt as actual deployed density becomes clear. This balances near-term capital discipline against the real risk of designing a facility that cannot accommodate the hardware its tenants want to deploy within a few years of opening.
DATAPERT's engineering and design teams work through exactly this set of trade-offs with clients across data center development programmes, from early feasibility through detailed design. Learn more about our technical advisory capabilities or start a project.
