Physics / mechanism
AI-grade datacentre real estate is purpose-built infrastructure engineered around the electrical, thermal, and structural demands of GPU/accelerator clusters rather than general compute. Key parameters: power density (AI racks run 30–120 kW/rack vs. 5–15 kW legacy), power-usage effectiveness (PUE target ≤1.2 for competitive facilities), cooling architecture (direct liquid cooling, rear-door heat exchangers, immersion), fibre dark-fibre proximity, and grid interconnect capacity. Site selection is now gated by utility agreements and substation buildout timelines, not floorspace. Current state: hyperscalers (Microsoft, Google, AWS) are committing 10–20 GW forward capacity; co-location players (Equinix, Digital Realty, CoreWeave) are repositioning legacy stock and breaking ground on greenfield AI campuses.
Competitive landscape
Competing supply vectors: hyperscaler owned-and-operated campuses capture roughly 60% of AI compute deployment, removing co-lo demand. Modular datacentre containers (Schneider Electric, Vertiv) compress build time to 6–9 months vs. 24–36 months for purpose-built. Edge micro-datacentres reduce latency-sensitive workload migration to centralized AI facilities.
Companies using
Connected ideas
Sources
Frontier (open questions)
- To be added.