AI Infrastructure

GPU Roadmaps and Facility Flexibility

Illustration showing successive GPU hardware generations and their increasing power draw inside a flexible data center design

Few hardware categories have seen power consumption climb as steeply, across as few generations, as AI accelerators. NVIDIA's A100 drew around 400 watts per chip at launch; the H100 pushed that to roughly 700 watts; the B200 reached around 1,000 watts; and current Blackwell-generation chips can reach up to 1.4 kilowatts per unit. A facility designed with fixed assumptions tied to any single generation in that sequence risks becoming a poor fit for the next one within just a few years.

Why This Trajectory Is a Genuine Design Problem, Not Just a Procurement One

Each step up in per-chip power draw cascades through a facility's entire electrical and mechanical design — distribution capacity, cooling capacity, structural loading for denser racks, and even network fabric requirements as higher-performance accelerators demand correspondingly higher-bandwidth interconnects. A facility's core infrastructure — electrical risers, primary cooling loops, structural capacity — is also the most expensive and disruptive part of the building to retrofit after construction, making it the part of the design where getting the flexibility question right matters most.

What Genuinely Flexible Design Looks Like

  • Electrical infrastructure sized with headroom beyond current requirements, allowing capacity upgrades without requiring a full rebuild of primary distribution systems
  • Cooling architecture designed around a liquid cooling baseline capable of scaling to higher densities, rather than air cooling with no credible upgrade pathway
  • Structural design that anticipates heavier, liquid-cooled racks even in areas initially fitted out for lighter equipment, avoiding the need for structural reinforcement later
  • Modular fit-out strategies that allow specific zones of a facility to be upgraded incrementally as tenant requirements evolve, rather than requiring uniform specification across the entire building from day one
The goal is not to correctly predict which specific GPU will be deployed in five years — it is to build a facility that does not care which one it turns out to be.

The Cost of Flexibility Has to Be Weighed Honestly

Designing in this kind of headroom is not free — over-sizing electrical and structural capacity beyond near-term requirements involves real upfront capital cost, and an overly conservative approach to flexibility can erode a project's near-term financial returns just as surely as an inflexible design risks future obsolescence. The right balance depends on a facility's specific tenant profile and intended use case: a facility built speculatively for an unknown future tenant base may warrant more aggressive flexibility investment than one built to a confirmed, long-term anchor tenant with a clearly understood technology roadmap.

Facility Flexibility as a Competitive Differentiator

As GPU power draw continues its steep generational climb, facilities that can credibly demonstrate flexibility to accommodate future hardware generations are likely to become increasingly differentiated and valuable relative to facilities locked into a single generation's design assumptions. This is a meaningful consideration for both developers planning new facilities and investors evaluating existing assets.

DATAPERT's engineering and design teams build this kind of forward-looking flexibility into data center development programmes from the earliest design stages. Explore our technical advisory capabilities or start a project to discuss flexible facility design.

Share LinkedIn X Email

Build the Infrastructure Behind
Tomorrow's Digital Economy

Whether you are planning a hyperscale campus or an AI-ready data center, DATAPERT can support the journey from concept to delivery.