Asking "what rack density should this facility be designed for" is, increasingly, the wrong question. A more useful framing treats power density as a roadmap spanning several plausible scenarios — 30 kW, 50 kW, 100 kW and beyond per rack — and designs core infrastructure capable of accommodating progression through that roadmap over the facility's operating life, rather than committing to a single point estimate that hardware evolution may quickly outpace.
Why a Single-Density Design Assumption Is Risky
GPU power draw has climbed sharply across recent hardware generations — from roughly 400 watts per chip for early accelerator generations to 1.4 kilowatts or more for current Blackwell-generation hardware — and there is little reason to expect this trajectory to plateau in the near term given continued AI compute demand growth. A facility designed rigidly around today's 120-140 kW rack density, with no credible pathway to accommodate the next generation's likely higher density, risks becoming a less competitive asset for AI tenants within just a few hardware cycles.
What a Genuine Roadmap Approach Looks Like
- 30 kW scenario: Suitable for enterprise and lighter AI inference workloads, achievable with advanced air cooling or modest rear-door heat exchanger assistance
- 50-70 kW scenario: Requires direct-to-chip liquid cooling as a baseline, with electrical distribution and structural capacity designed accordingly
- 100 kW-plus scenario: Represents current frontier AI training density, requiring full liquid cooling infrastructure, reinforced structural capacity, and electrical architecture designed for highly dynamic, synchronised load patterns
- 200 kW-plus scenario: An increasingly plausible near-future density, potentially requiring immersion cooling or advanced two-phase direct-to-chip systems for at least the most demanding zones of a facility
Designing core infrastructure — primary electrical risers, structural capacity, piping headers — with headroom for progression through this roadmap, while phasing actual fit-out to match real, confirmed tenant requirements, balances near-term capital discipline against the genuine risk of premature obsolescence.
A facility's power density roadmap is not a prediction exercise — it is an insurance policy against the near-certainty that today's density assumption will look conservative within a few hardware generations.
Zoning Within a Single Facility
Many AI-ready facilities are now designed with deliberately differentiated zones — areas fitted out for current frontier density to serve the most demanding tenants immediately, alongside areas designed with structural and electrical headroom but initially fitted out at a lower density for tenants with more modest current requirements. This zoned approach allows a single facility to serve a broader range of tenant requirements simultaneously, while preserving a credible upgrade pathway for lower-density zones as demand evolves.
Capital Discipline Without Sacrificing Flexibility
The roadmap approach is not a licence for unlimited speculative over-engineering — every increment of headroom carries real capital cost, and the right level of investment in future flexibility depends on a facility's specific tenant strategy, financing structure, and risk tolerance. The discipline lies in making this trade-off deliberately and explicitly, rather than defaulting either to rigid design around today's density or to indiscriminate over-building that erodes near-term project economics.
DATAPERT's engineering teams build power density roadmaps into data center development programmes from the earliest design stages. Explore our technical advisory capabilities or start a project to discuss a power density strategy for your next facility.
