Imagine you're trying to keep a barbecue grill cool. With charcoal, you can blow air on it and it works. With a small propane burner, fans still work. With a jet-engine afterburner, no amount of air will save you — you have to physically run cold water through the metal to carry the heat away.
That is the trajectory of AI chips, year by year. An NVIDIA H100 consumed about 700 watts. The Blackwell B200: 1,200 watts. The Rubin generation, due late 2026: 1,800–2,400 watts per chip . A single GB200 NVL72 rack now draws roughly 120 kilowatts — about the same as 100 American homes, concentrated into a vertical refrigerator-sized box. Air cooling tops out around 30 kW per rack. So every Blackwell and Rubin GPU must be liquid-cooled : a cold plate sits directly on the chip, coolant flows through it, and a coolant distribution unit (CDU) carries the heat out of the rack to a chiller on the roof.
This created two new bottlenecks in 18 months: the CDUs that route coolant to and from the chip (Vertiv leads at ~11% market share with a $15B backlog), and the chillers on the roof that ultimately reject the heat (Trane, Carrier, Johnson Controls, Daikin, plus Modine — which pivoted from auto radiators and just signed a $4B hyperscaler chiller LTA). Below those are the specialty refrigerants and dielectric fluids — Honeywell and Chemours stepped in as 3M exited PFAS-based Novec.