Rack power has climbed to 60–100 kW, roughly seven to twenty times the old air cooled baseline, forcing new facilities to be designed around coolant loops, pumps, and non conductive dielectric fluids.
For decades, the standard data-center cooling playbook was simple: push cold air through perforated floor tiles into the front of a rack, then draw the warmed air back out through overhead returns. Most legacy facilities were designed around rack densities of 5–15 kilowatts, with chilled water stepping in once facility loads crossed roughly 500–600 kW (Willem Weber, "Liquid cooling is no longer optional for data centres in the age of AI"). That math held for general-purpose enterprise computing for thirty years.
AI workloads have ended it. Modern AI training and inference clusters now pack 60–100 kilowatts into a single rack, with some deployments pushing past 150 kW, according to Weber, a mechanical engineering consultant at Digital Parks Africa. At those densities, forced-convection air cooling no longer moves heat out fast enough to keep silicon within safe operating temperatures. The enthalpy budget, the amount of heat a given airflow can absorb before it becomes too hot to reuse, collapses.
Liquid changes the equation for a physical reason: water carries roughly 3,500 times more heat per unit volume than air, and pumps move it with less energy than fans move equivalent volumes of air. Coolant distribution units (CDUs) feed secondary manifolds, and dielectric or water-glycol fluids run through cold plates mounted directly on accelerators. Dielectric coolants are non-conductive, so a leak doesn't short the hardware. Operators now plan for fluid handling and leak containment in addition to temperature (Weber).
The shift is more than a hardware swap. Legacy brownfields designed for 5–15 kW racks can't economically retrofit to 60–100 kW loads. The binding constraint is usually the upstream electrical and grid capacity that an AI rack needs, not the cooling loop itself. Greenfield sites, purpose-built AI campuses backed by on-site generation or pre-secured grid connections, hold a structural advantage, which is why the cooling decision and the power decision now have to be made together (Weber).
Digital Realty has launched a direct liquid cooling offering, treating liquid as a product line rather than a custom one-off (Data Center Dynamics). Hyperscaler 2026 outlooks describe liquid and hybrid cooling as the next operating model, not a fringe option (Data Center Knowledge). For data-center REIT investors, the consequences are concrete enough to shift positioning: Equinix's Q1 2026 commentary has already framed AI inference workloads as a driver for data-center real estate fundamentals (AI Consulting Network).
Weber's column is a single-consultant engineering argument, not a market-wide measurement. The 60–100 kW threshold, the loss of air-cooling economic viability at AI scale, and the grid-as-binding-constraint claim all reflect his judgment rather than independently published benchmarking. The direction of the inflection is consistent with industry-side product launches and hyperscaler outlooks, but operators and investors evaluating real exposure will need deployment-specific PUE deltas, named customer sites, and grid-capacity figures that the column does not supply.
AI compute density is structurally rising: larger models, more parameters per server, more accelerators per rack. The cooling decision is no longer separable from the power decision, and the build calendar for liquid-ready facilities, on Weber's framing, already extends 18–36 months out. Operators that haven't started redesigning around pumps and CDUs are effectively choosing which workloads they cannot host.
Hyperscaler 2026 capex disclosures and the first deployment-specific PUE deltas tied to those build-outs will show how much of Weber's thesis is already procurement reality.