Traditional data center racks were designed around 5 to 15 kilowatts. That was the thermal envelope for decades, and the cooling industry built a global fleet of CRAC units, raised floors, and hot-aisle containment around it. AI clusters changed the math permanently. Rack densities are now running 50 to 100 kilowatts, and next-generation GPU configurations are pushing past that. Richard Clifford, VP of Sales and Solutions EMEA at Salute, described the implication plainly: heat management has become a central technical limitation on scaling AI compute, not a facilities afterthought to resolve after the servers are specced.
The physics behind this are not contested. Air moves heat at a rate determined by its specific heat capacity and the volume you can push through a rack. Water carries heat roughly 3,500 times more efficiently per unit volume than air at equivalent conditions. At 15 kilowatts, air works. At 80 kilowatts, you are moving so much air at such velocity that the mechanical energy cost of the fans becomes a material fraction of the server power draw itself. At 130 kilowatts, air has no path. The progression from room cooling to row cooling to rack-level direct-to-chip liquid cooling is not a technology preference. It is the industry following the physics upward along the wattage curve.
Moving from air to liquid cooling is not a configuration change. It requires new piping, heat exchangers, and specialized pumps. Secondary coolant loops inside server chassis need manifolds, quick-disconnect fittings, and leak containment measures that air-cooled facilities never had to plan for. Workforce training is a genuine constraint: the technicians who maintain CRAC units are not automatically qualified to commission and service CDUs and cold-plate manifolds. Salute's Clifford named lifecycle expertise as the differentiator that separates operators who execute liquid cooling migrations smoothly from those who commission at scale and discover thermal validation gaps between vendor specifications after the facility is already running.
Site selection is also shifting. Facilities that support AI workloads need high power availability, large available space, and heat rejection infrastructure rated for loads that older sites were never designed to handle. Environmental compliance adds another variable: water use restrictions in Nevada and Arizona affect the evaporative cooling systems that many facilities still depend on as their primary heat rejection mechanism. Operators who treat cooling as a late-phase infrastructure decision are building themselves into a corner where their power contract and their thermal rejection capacity are mismatched before the first server rack arrives.
The vendor response to this constraint is consistent across the sector. Vertiv acquired Strategic Thermal Labs to add server-side cold-plate engineering. The direct-to-chip cooling market is projected at $17.31 billion by 2032, growing at 26.5% annually. Air cooling and liquid cooling are not competing modalities for AI workloads: air handles everything below 40 kilowatts; liquid handles the rest. The hybrid architecture that operators will actually run through the end of this decade uses both, with RDHx and direct-to-chip loops alongside existing CRAH units for legacy workloads.
The operators who have already committed to liquid cooling infrastructure and built the internal expertise to commission and maintain it are not making a premium bet on an emerging technology. They are building the baseline that AI compute requires. Everyone retrofitting on someone else's timeline will pay more per kilowatt of thermal capacity and wait longer for the components to arrive.