The Cooling Wars: Air, Liquid, and the 600-Kilowatt Rack
The average server rack still draws under 9 kW. AI racks are heading past 600. Inside the industry's forced march from air conditioning to liquid cooling, and the water politics that come with it.
Here is the strangest fact in data center engineering right now: the industry’s average rack still draws less than 9 kilowatts, according to the Uptime Institute’s 2025 global survey, while the racks being installed for AI training draw fifteen times that, and the ones on NVIDIA’s roadmap draw seventy times that. Two different industries are sharing the same buildings, and only one of them can still be cooled with air.
Where air runs out
Air cooling has carried computing for sixty years, and for most workloads it still works fine. Contain the hot aisles, push chilled air through the cold ones, and a well-designed room handles 20 to 25 kW per rack; stretch the aisles and it might reach 30. Uptime’s 2025 cooling survey found that 75 percent of operators still rely on perimeter air cooling, and most say liquid only becomes necessary above roughly 20 kW per rack.
AI blew through that ceiling in one product generation. NVIDIA’s GB200 NVL72, the rack-scale system behind most frontier-model training, draws about 120 to 140 kW per cabinet. Its successor is mainstream in 2026 deployments at similar density. And the Rubin Ultra “Kyber” racks arriving in late 2027 are specified at around 600 kW, with NVIDIA executives openly discussing megawatt racks for the generation after. No amount of fan engineering moves that much heat with air; the physics simply gives out.
The industry’s average PUE (total facility power divided by IT power) has been stuck at 1.54 for six consecutive years. Google’s fleet runs at 1.09.
Liquid goes mainstream
The answer is direct-to-chip liquid cooling: cold plates clamped to the processors, with coolant loops carrying heat out of the rack far more efficiently than air ever could. In Uptime’s 2025 survey, 22 percent of operators had deployed it and another 61 percent said they’d consider it. Among AI hardware specifically, the shift is faster: TrendForce estimates liquid cooling covered about a third of AI chips in 2025 and will pass half in 2026. Google, the most aggressive adopter among hyperscalers, already liquid-cools more than 80 percent of its AI servers.
The vendor landscape is shifting with it. When AWS unveiled its in-house In-Row Heat Exchanger system for Blackwell racks in July 2025, co-engineered with NVIDIA in eleven months, shares of cooling incumbent Vertiv fell more than 6 percent in a day. Cooling, long a sleepy corner of mechanical engineering, has become a competitive weapon.
Immersion’s setback
The most radical approach, submerging entire servers in dielectric fluid, has split into two very different stories. Single-phase immersion, using mineral-oil-like fluids, works today and handles 100–120 kW racks. Two-phase immersion, which uses engineered fluids that boil off the chips and recondense, promised even more, until its chemistry collided with reality. The key fluids are PFAS, the “forever chemicals” now facing an EU-wide restriction proposal and multibillion-dollar litigation. 3M, the dominant producer, exited PFAS manufacturing entirely at the end of 2025, and Microsoft and Meta shelved their two-phase research. A Microsoft lifecycle study published in Nature in 2025 offered the industry a tidy exit: optimized cold plates and single-phase immersion match two-phase efficiency without the toxic fluids.
The water question
Cooling is also where data centers touch their most politically sensitive resource. Evaporative cooling, the cheapest way to reject heat, consumes real water at scale: Google’s data centers used 6.1 billion gallons in 2024, up more than 40 percent since 2021, and an AP investigation found most operators won’t disclose site-level figures at all.
The engineering response is closed-loop designs that give up some efficiency to consume almost nothing. Microsoft’s newest generation, including its Fairwater AI campuses, fills its loops once (about what 20 homes use in a year) and replaces the water every six-plus years; the company’s fleet-wide water intensity has already improved 39 percent since 2021. Expect “zero-water” to become a standard bullet point in every new campus announcement, because water access is now a permitting issue, not just a sustainability one.
The bottom line
For thirty years, cooling was the part of the data center nobody outside the industry thought about. The AI buildout has made it the binding constraint: the difference between a 20 kW rack and a 600 kW rack is not an upgrade, it’s a different kind of building: different plumbing, different floor loading, different economics, different politics. The companies that solve heat at scale will decide how fast the AI era actually arrives.