The chips that power modern AI models now draw more than 1,000 watts apiece, more than a typical household appliance and roughly triple what a server chip consumed five years ago. Cooling them has become a constraint as real as the silicon itself, and two of the industry’s biggest names are pushing the answer: liquids instead of air.
Nvidia laid out a full liquid-cooling plan for its Rubin platform in a company blog post this week, describing a design in which coolant circulates through the racks that hold its next-generation accelerators. Google, for its part, has introduced a technical route it calls Brazos that significantly boosts the performance of traditional data centers, an approach that also leans on liquid cooling for the hottest components.
The two announcements have revived a debate that data-center engineers have been having for years: whether the industry is ready to abandon the air-cooled racks that have served it for decades. Industry insiders say three technical routes now coexist—air cooling for older and lighter workloads, cold-plate liquid cooling that moves heat through metal plates attached to chips, and immersion cooling that submerges entire servers in dielectric fluid.
Because infrastructure, geography and climate differ around the world, most engineers expect a hybrid future: liquid cooling at the core, with other routes running in parallel. In desert climates, water scarcity pushes operators toward air or immersion systems that recycle coolant; in temperate regions with cheap power, cold-plate systems dominate. The choice is rarely technological purity; it is a calculation of cost, water, power and maintenance.
The physics is unforgiving. A standard air-cooled rack can dissipate roughly 30 to 40 kilowatts of heat before airflow becomes impractical; liquid-cooled racks handle well over 100 kilowatts. As chips draw more power, the crossover point has arrived faster than most operators predicted. Analysts estimate that liquid cooling will account for the majority of new AI data-center capacity within the next few years, and equipment makers including Vertiv and nVent have reported record backlogs for cooling products.
The pressure is not only thermal but electrical. Data centers around the world are hitting power-supply limits, and utilities are telling operators they must wait years for new connections. Cooling systems that squeeze more compute out of each megawatt are therefore valuable twice over: they keep chips within safe temperatures and they stretch scarce electricity further. Regulators in several countries have begun tying energy-efficiency standards to how facilities manage heat.
That shift is a gift to a supply chain that has been building for it. In China, where the push has been especially visible, a swath of listed companies has laid out AI liquid-cooling businesses covering the full chain: cold plates that sit on the chips, coolant distribution units that pump fluid through the racks, the cooling fluids themselves, and complete cooling systems that integrate everything.
Companies such as Envicool, Goaland and Tongfei, among others, have expanded capacity and signed contracts with data-center operators, according to company filings and industry reports. The Chinese market has moved fast partly because the country’s data-center industry is young enough to build for liquid cooling from the ground up, rather than retrofitting decades-old facilities. Analysts in Shanghai track the segment as a proxy for domestic AI spending.
The retrofit problem is the industry’s quieter challenge. Most of the world’s data centers were built for air cooling, and converting them is expensive: pipes, pumps and heat exchangers must be threaded through buildings designed around fans and vents. Hyperscalers building new campuses can design for liquid from day one; enterprises with older facilities face a harder choice between spending on conversion or accepting lower-density deployments.
Nvidia’s role in the shift is central. Because its accelerators define the thermal demands of the modern data center, its decisions effectively set the industry’s cooling agenda. When the company says Rubin will be liquid-cooled, every operator planning a Rubin deployment must plan the plumbing to match. That, more than any conference presentation, is what drives the cooling market.
The economics are not trivial. Liquid cooling adds cost per rack, and maintenance requires skills that most data-center staffs do not yet have. But the alternative—running chips at reduced power to keep them cool—costs more in lost compute than the cooling systems save. The math favors the liquids, particularly as electricity prices climb and regulators tighten energy-efficiency rules for large facilities.
For suppliers, the opportunity is measured in the hundreds of billions of dollars of data-center construction expected over the next several years. For operators, the question is no longer whether to adopt liquid cooling, but which flavor: cold plates for the mainstream, immersion for the densest deployments, and air for everything that runs cool enough to keep it.
The industry appears to have settled the debate in practice even as it continues in conference halls. Water is moving through the racks. The only open question is how fast, and who supplies it.
