Designing a site for modern AI clusters

"Data center space" and "AI-ready capacity" sound interchangeable. They are not. A site designed for legacy enterprise racks and a site designed for a modern accelerator cluster diverge on the three variables that decide whether the hardware can actually run: power density, thermal, and networking. Getting these wrong does not produce a slightly worse facility: it produces one the cluster cannot occupy.

Power density: the rack is the new unit

Legacy facilities were planned around rack averages that today's accelerators blow past. Modern AI racks draw far more than the enterprise norm, and the distribution, busway, and protection have to be designed for that sustained draw rather than a comfortable average. A site that cannot deliver high per-rack density forces the cluster to spread out, stranding floor space and connectivity.

Thermal: planning around the heat, not the room

Dense accelerators concentrate heat to a degree that air alone increasingly struggles to remove economically. AI-ready design plans cooling around the actual heat load: hot/cold aisle discipline at minimum, and direct-liquid-ready provisioning where the workload warrants it. The decision that matters is made early: a site that is plumbed and structured for liquid from the start can adopt it; one that is not faces a retrofit.

Networking: bandwidth and latency are part of the building

Training and large-scale inference are as much a networking problem as a compute problem. The site has to support high-bandwidth, low-latency fabric internally, and carrier-diverse, high-capacity paths externally for hybrid and burst patterns. Fiber that is already in the ground, municipal dark fiber with Tier 1 peering, turns this from a bespoke build into a provisioning step.

Why "we run this ourselves" matters

These are not checklist items you can copy from a generic spec. They are decisions informed by operating the hardware. A developer that runs AI infrastructure itself designs for the failure modes and the thermal reality it has actually seen, which is the difference between a building that looks AI-ready on paper and one a cluster can move into.

Generic real estate can be leased. AI-ready capacity has to be designed for the workload from the substation to the rack.