Data Center Knowledge published a piece on May 14 making an argument the cooling industry has watched arrive in slow motion. Its reporting traces Meta's average rack density from roughly 18 kW in 2022 to about 34 kW in 2025, with Nvidia Blackwell NVL72 systems pushing maximum density near 130 kW. Omdia's Vladimir Galabov is quoted saying the industry is now ready for a 600 kW rack. Five years ago a 100 kW rack was treated as an extreme outlier.
The consequence is a physical inversion of the building. Schneider Electric's Steven Carlini describes AI factories as having very limited white space, with the majority of the floor given over to grey space: chillers, generators, transformers, switchgear, and coolant distribution. The servers are becoming the smaller part of the footprint.
When grey space overtakes white space, the cooling and power share of total construction cost rises with it. A facility that is mostly thermal and electrical plant spends a larger fraction of its capital budget on heat rejection, distribution, and redundancy than on the racks doing the compute. That is the structural reason the cooling order book keeps expanding even when individual GPU deployments slip. The 50 kW to 1 MW density transition is not a future curve. It is being priced into projects breaking ground this year.
Carlini also makes the point that converting a legacy data center into an AI factory is extremely challenging. That difficulty is the cooling industry's friend. It pushes hyperscalers toward greenfield builds with thermal plant designed in from day one, and it pushes enterprises that cannot retrofit toward colocation. Both paths route spend through the same vendor base. The brownfield retrofit work that does happen is itself a cooling-led project, not an IT refresh.
The same reporting notes the industry is converging on Open Compute Project specifications and Nvidia reference architectures, and that facilities now run air and liquid loops simultaneously at different temperature bands. Standardization is good for deployment speed and bad for vendor margin. When the rack, the manifold interface, and the temperature setpoints are specified by a reference design, the cooling vendor competes on manufacturing throughput, lead time, and field reliability rather than novel architecture. That is the same dynamic Nvidia's wattage roadmap has been imposing on the sector for the past year.
The planning implication is that thermal capacity, not IT capacity, is now the binding constraint on how much compute a site can hold. Liquid cooling at these densities is non-negotiable, and the building has to be sized around heat rejection first. The useful question for a 2027 site is not how many racks fit on the floor. It is how many megawatts of heat the grey space can move out of the building, because that number sets the ceiling on everything else.