AI racks are moving to direct-to-chip liquid cooling across Nvidia, Google and AMD platforms. Capacity is being added fast enough that commodity parts will lose their premium, while coolant distribution units, controls and whole-system integration keep theirs.
Liquid cooling has become a standard part of high-density AI infrastructure. CITIC Securities argued on September 24 that AI systems are entering an “all-liquid cooling” phase, citing Nvidia’s Rubin, Google’s TPU generations, AMD’s Helios racks and large cloud deployments. Server makers are already building for it. Super Micro’s current platform disclosures include liquid-cooled configurations for Nvidia Vera Rubin systems and AMD Helios.
The physics support the shift. Accelerator density has raised the amount of heat each rack must shed. At the same time, operators want to reclaim the power used by fans and conventional air conditioning. Direct-to-chip loops, coolant distribution units, pumps and manifolds are now critical as rack power keeps climbing.
For investors, the key question is which parts of the cooling chain will keep their margins once supply catches up. Our answer is that supply will scale faster than pricing power, and the value will settle in only a few places.
Why supply will catch up
Most of a direct-to-chip cooling system is made from well-understood components. Cold plates are machined copper blocks with internal channels. Manifolds are plumbing. Pumps and heat exchangers come from mature industrial supply chains. None of these parts needs the years of process development that limit leading-edge chips or advanced packaging.
Manufacturers are adding capacity quickly. Taiwanese thermal specialists that grew up making PC and server heat sinks, such as Asia Vital Components and Auras, have expanded cold-plate production for AI servers. Industrial groups are buying their way in: Schneider Electric agreed in 2024 to buy 75% of Motivair for $850 million. Eaton paid $9.5 billion for Boyd Thermal, about 22.5 times estimated 2026 adjusted EBITDA, in a deal that closed in March. Ecolab completed its roughly $4.75 billion purchase of CoolIT Systems in July. The Eaton price looks like a challenge to this piece’s argument, since Boyd makes cold plates, the part we expect to commoditise. Eaton’s own rationale answers it. Boyd expects about $1.5 billion of its $1.7 billion of 2026 sales to come from liquid cooling, spanning CDUs, cooling loops and cold plates. Eaton bought it to sell a combined power and cooling offering from “the chip to the grid”. It paid 22.5 times earnings for a system supplier, not for a cold-plate maker. Several of today’s shortages will turn into surpluses within one or two accelerator generations.
Standardisation speeds that up. Nvidia publishes reference designs for its rack systems, and the Open Compute Project has published a universal quick-disconnect specification since 2020, with a second version in progress. When a hyperscaler can source an interchangeable part from several qualified suppliers, the part’s price moves towards its manufacturing cost.
Where pricing power survives
The parts that resist commoditisation share three features. Failure is costly, integration with the facility is hard, and the supplier carries responsibility for how the system behaves over time.
Coolant distribution units (CDUs) are the clearest case. A CDU manages the flow, pressure, temperature and quality of coolant between the facility water loop and the racks. If it fails, the result is a hall of throttled or shut-down accelerators worth hundreds of millions of dollars. Buyers choose CDUs on reliability records, controls software and service coverage rather than unit price. Vertiv, Schneider Electric through Motivair, and CoolIT Systems, now owned by Ecolab, are among the suppliers hyperscalers qualify, and qualification takes long enough to protect margins.
Facility-level heat rejection and controls are the second case. Liquid-cooled halls need chillers, dry coolers, heat exchangers and building controls sized for the new loads, and the choices interact with power and water constraints on each site. Suppliers that can design, deliver and maintain the whole thermal chain from chip to roof, a market Vertiv and Schneider are both built to serve, capture integration value that component makers cannot.
Quick-disconnects fall in between. A leak at a coupling inside a rack is expensive, so buyers favour proven suppliers such as Stäubli, Parker and Danfoss. But as common specifications spread, the premium will narrow, and the part will be priced like a precision industrial fitting rather than a scarce AI component.
Cold plates will commoditise first. They are designed around each accelerator’s reference layout, made by many suppliers and rebid with each new chip generation. Revenue will keep growing with accelerator volumes, but margins will fall as capacity catches up with demand.
The hyperscaler’s bargaining power
Customer concentration pushes in the same direction. A small number of hyperscalers and server integrators buy most AI cooling hardware. They qualify several suppliers on purpose, change designs between accelerator generations and can bring component design in-house. Suppliers that sell a component lose bargaining power as soon as a second source is qualified. Suppliers that sell a maintained system, with software, service and uptime commitments, keep it.
Service contracts add to that advantage. A liquid-cooled hall needs coolant chemistry monitoring, filter changes, leak detection and fast repair, all under uptime guarantees. That recurring revenue is harder to rebid than a hardware order, and it goes to the supplier whose technicians are already on site. Vertiv and Schneider have large field-service organisations built for data-centre power equipment, which they can extend to cooling. Component makers mostly do not.
This is also why the growth of the market is a poor guide to returns. Cooling spend per rack will keep rising with power density, but the extra spending will not be shared evenly. Investors should expect cold-plate makers’ revenue to outgrow their margins, while margins at CDU and thermal-system suppliers hold up.
Is the premium already priced?
Largely, yes. On September 23 Vertiv traded at about 38 times forward EV/EBITDA on TIKR’s estimates, roughly twice Schneider Electric’s 18.5 times and well above industrial peers. The market has already identified the system suppliers as the winners. What it has not fully priced is the other side of the split: cold-plate makers whose revenue will keep growing as their margins come under pressure through 2027.
The risk to the whole thesis is standardisation reaching the CDU through the Open Compute Project, which would expose the last protected part of the chain to rebidding. Until that happens, our view is that the better trade is to avoid the commoditising components rather than chase Vertiv at twice Schneider’s multiple. Schneider owns the same CDU franchise through Motivair at half the valuation.