Why data centers are quietly re-plumbing themselves for AI

A standard data center rack in 2020 was designed around roughly 6 kilowatts of power draw, comfortably managed by pushing chilled air through the aisles. New AI racks running current-generation GPUs are landing in the 60-150 kilowatt range, and next-generation designs are projected to reach as high as 370 kilowatts, according to Deloitte's 2026 data center outlook. Air cooling simply doesn't work at that density — there's more heat concentrated in one rack than a fan-driven system can move.

The industry's answer is liquid cooling, and the shift has happened fast. AFCOM's 2026 State of the Data Center Report found 36% of operators have already deployed liquid cooling with another 28% planning to within 24 months, while the specific data center liquid cooling market is projected to grow at better than 16% annually through the next decade. Direct-to-chip cooling — coolant piped straight to a cold plate mounted on the GPU — has become the default approach for the 50-150 kilowatt range, while full immersion cooling, submerging entire servers in dielectric fluid, is now being commercially deployed for the most extreme densities above 100 kilowatts.

This matters beyond the hyperscalers building dedicated AI campuses. One analysis pegged the cost of delaying a liquid cooling transition, once a facility crosses the relevant density threshold, at roughly $3.2 million per megawatt per quarter of delay — a number that shapes colocation pricing and availability for any business renting space rather than building it.

For a business that just leases rack space or cloud compute rather than operating its own facility, the practical relevance is indirect but real: colocation and cloud pricing in capacity-constrained markets increasingly reflects how much a provider has had to spend re-plumbing for AI-density workloads. If a colo contract renewal is coming up, it's a reasonable question to ask directly — not just about uptime, but about cooling capacity and whether the facility can support higher-density racks if your own compute needs grow.