A new class of infrastructure. Because AI is a new kind of load.

AI is not just more compute in the same building. It is a different physical system: higher density per rack, higher continuous load, and a mandatory shift to liquid cooling at the top of the range.

Conventional server rack vs liquid-cooled AI compute rack — Left, a conventional server rack running about 35 to 45 kW under air cooling. Right, a liquid-cooled AI compute rack running 120 to 140 kW.

The three-fold density jump.

NVIDIA's H100 generation ran 35 to 45 kW per rack, cooled by air, in facilities that typically achieved a PUE (Power Usage Effectiveness) of 1.3 to 1.6. NVIDIA's GB200 NVL72 generation runs 120 to 140 kW per rack, cooled by direct-to-chip liquid, in facilities designed for a PUE of 1.03 to 1.10. That is roughly three times the density, and the shift from air to liquid is not optional at that draw.

35-45 kWConventional rackAir cooled

120-140 kWAI compute rackLiquid cooled

3xDensity jumpAir to liquid

1.03-1.10Design PUELiquid-cooled AI

Continuous, not bursty.

AI campuses run continuously at 100 to 750 MW of electrical load. Unlike search or storage workloads, which have peaks and troughs, AI training and inference are close to fully utilised for the useful life of the hardware. That changes power procurement (long-term firm supply, not peak trading), grid design (dedicated 220 kV plus substations), and cooling design (steady heat, not spiky).

The facility efficiency gap.

Liquid cooling reduces facility overhead by up to 15 to 20 per cent at gigawatt scale. On a 1 GW AI campus that translates to hundreds of megawatts of avoided overhead per year, energy that goes to compute instead of to fans and chillers. At full stack economics, this compounds meaningfully.

Interior of a liquid-cooled AI data hall with blue coolant piping and rack rows — Interior of a liquid-cooled AI data hall. Concept image, illustrative.

The race to build AI is a race to build the infrastructure beneath it.

Sources: NVIDIA product specifications for H100 and GB200 NVL72. Uptime Institute facility efficiency benchmarks, 2024 and 2025.

Back to home