Where air cooling runs out
Traditional data centre cooling was designed around racks drawing 10 to 15kW. Fans, cold aisles and room air handlers handled that comfortably.
Air reaches its limit well below AI rack densities, because of what air can physically carry.
Why the old model broke
AI racks sit far above the point where air begins to fail. Air cannot move the heat fast enough or precisely enough to hold the temperatures GPU performance depends on.
The failure is not a hot aisle but silicon throttling, with the hardware losing the performance it was bought for.
Why liquid solves it
Water carries far more heat per unit volume than air, and it can be delivered directly to the chip and controlled at rack level.
Liquid cooling itself is long established. What has changed is the scale it now runs at.
What that changes about testing
Proving one of these buildings is a different job now. You are validating a pressurised fluid circuit and a secondary cooling loop, with flow rates and delta T across every rack, water quality and thermal response under a changing GPU load.
That calls for load banks that connect to the cooling loop and produce data on how the infrastructure will actually perform: the base build range to prove the plant, and the end-of-line GPU emulators to replicate the racks.