Before AI came along, data centers cooled themselves in a very simple way: blow cold air into the racks, suck hot air out, and an air conditioner plus a bunch of fans was enough.
That method worked for thirty years, until GPU power curves completely broke it. By the time we reached NVIDIA’s GB200 generation, power per rack had already surged past 120 kilowatts, and air cooling simply no longer made physical sense. The entire industry is not “choosing” liquid cooling; it is being forced by the laws of physics to switch to liquid cooling.
First, let’s make clear why air cooling is not enough.
Air has a very low specific heat capacity and thermal conductivity. For the same volume, it can remove one to two orders of magnitude less heat than water. GPU power has jumped from the A100’s 400 watts to the B200’s 1,000 watts, and then to Rubin, the next generation, at more than 1,400 watts. The heat flux on the chip surface is already approaching the level of an electric iron, and no amount of airflow can blow it away.
More importantly, GPUs are no longer scattered one by one; now 72 of them are packed tightly together in the same rack and interconnected. Once heat sources are concentrated to this density, the only way air cooling could work would be to raise the data center ceiling height to absurd levels, which is not economical.
There are two main liquid cooling paths.
The first is called Direct-to-Chip: direct-to-chip cooling. You attach a metal cold plate to each GPU, with coolant flowing inside it, carrying heat away directly from the back of the chip.
The second is called Immersion: immersion cooling. The entire server is submerged in an insulating liquid, and the liquid itself circulates to carry away heat.
Each path has its supporters, but in the AI rack scenario, the industry has basically converged on Direct-to-Chip, and NVIDIA’s official GB200 NVL72 design uses direct-to-chip cooling.
The reason is simple: direct-to-chip requires the smallest changes to the server structure and is compatible with the existing data center ecosystem; immersion cooling may be more efficient, but it requires redoing the entire system, maintenance, and supply chain, and the ramp-up is too slow.
What liquid cooling brings is not as simple as “swapping out a heatsink”; it means redoing the entire data center’s electrical, mechanical, and water systems.
You need a coolant distribution unit (CDU) to distribute coolant to each row of racks; you need a secondary-side heat exchanger to transfer heat to the building’s chilled-water loop; you need leak detection, quick shutoff valves, and backup pumps; and you also need to redesign the internal water connections and quick couplers inside the racks.
Supermicro’s CDU solution for GB200 NVL72 reaches 250 kilowatts per unit, and a whole data center needs dozens or even hundreds of such units. So old data centers are almost impossible to retrofit; you can only build new ones. And building new ones means an entire round of engineering, permits, and equipment procurement.
That has pushed a group of companies that used to be very “traditional” to the front line of AI infrastructure.
Vertiv is the most typical example. It makes CDUs, precision air conditioners, rack power distribution, and UPS systems as a full stack. The market re-rating has been very direct: over the past year, its stock price has risen more than 250%, and quarterly profit has nearly doubled year over year.
In 2025, Eaton spent $9.5 billion to acquire Boyd Thermal, stepping into the direct-to-chip cooling market in one move and formally competing head-on with Vertiv.
Beyond these two, established HVAC and electrical companies such as Schneider and Trane are also increasing their bets on this track. These are companies that, over the past decade, never appeared on the list of “AI concept stocks.”
When looking at this chain, there is one thing to remember: liquid cooling is not an optional extra; it is a necessity.
Every new AI rack shipment will bring with it an order for a CDU, a power distribution cabinet, and a rack-level cooling loop.
The GPU shipment cycle determines the order cycle for liquid cooling equipment, and that order cycle is pulling a group of traditional electrical engineering companies out of a decade-long low-growth rut.
In the air-cooling era, data centers were an IT engineering matter; in the liquid-cooling era, data centers are an electrical and thermal management engineering matter. This is a shift in the center of gravity that happened quietly, but its impact runs deep.