The density shift reshaping AI infrastructure

From kilowatts to hundreds of kilowatts

Training and inference clusters have compressed a decade of density evolution into three years. A campus designed for 8 to 12 kW per rack cannot host current-generation accelerators without a fundamental redesign of thermal and electrical distribution.

Three consequences for site selection

Land footprint falls, power footprint rises. The same 100 MW now fits on a materially smaller parcel. Sites previously dismissed as too small return to the shortlist provided the connection is firm.

Water strategy becomes a permitting variable. Liquid cooling shifts the conversation from consumption volumes to closed-loop design and heat reuse commitments — increasingly a condition of local political support.

Latency tolerance widens. Training workloads accept distances that colocation never would, which opens secondary and tertiary markets where grid headroom still exists.

What this means for the pipeline

Every position should be re-tested against a high-density reference design. Sites that fail on structural loading, cooling corridor or substation proximity are not marginal — they are unsuitable, and should be recycled early rather than carried.