MarketsandMarkets sizes the data center pump market at $2.74 billion in 2026, reaching $7.35 billion by 2032 at a 17.9 percent CAGR, in its report summary. Cooling and thermal management is the largest application, which covers chilled water, condenser water, and the secondary loops feeding CDUs. Centrifugal is the largest pump type, split further into inline, end suction, and split case configurations, with positive displacement as the other category. The report calls Asia Pacific the fastest-growing region and attributes it to demand: cloud buildout and AI infrastructure investment across China, India, Indonesia, and Malaysia. Named vendors are Grundfos, Wilo, Xylem, Armstrong Fluid Technology, KSB, Sulzer, Ebara, Flowserve, ITT, and Pentair.
Not one of those ten is a data center cooling specialist. They are industrial pump builders with century-old catalogs, selling through mechanical contractors and consulting engineers, competing on a curve standardised long before anyone put a GPU in a rack. That routing decides two things. Pump selection happens in the mechanical package, which means it is made by a consulting engineer optimising for first cost and code compliance rather than by the team that owns GPU uptime. And a commodity sold through contractors gets value-engineered, so the pump that arrives is often not the pump that was specified.
Comparing this forecast to the cold plate market takes care, because the scopes overlap rather than nest. The $2.74 billion figure spans facility pumps across the whole estate, most of which never touch a cold plate, while the direct-to-chip forecast of $3.33 billion in 2026 rising to $17.31 billion in 2032 already contains the pumps inside every CDU it counts. Same for the $1.80 billion valve forecast. What the numbers do support is a directional read: the components with software, controls, and integration attached are forecast to grow faster than the rotating equipment underneath them.
A direct-to-chip loop holds die temperature by holding flow, and flow is a pump curve intersected with system resistance. Those two inputs move in opposite directions depending on what changes. Foul a strainer or throttle a manifold and resistance climbs, the operating point walks up and left, and flow drops, which is the case that starves a cold plate. Add racks in parallel and total flow rises while per-rack flow falls, which is the case that starves the racks that were already there. Neither failure is visible from the CDU's own gauges. Redundancy inside a CDU is normally N+1 on the pump, so one failure is a maintenance ticket and two inside the same window is a thermal event across the row that unit serves.
Water's vapor pressure roughly doubles between 30C and 45C, from about 4.2 kPa to 9.6 kPa, which costs roughly half a metre of available net positive suction head. In a pressurised closed secondary loop that is absorbed easily. In an atmospheric tank or any sub-atmospheric design it is a real bite out of a small budget, and it lands at the same time NVIDIA is pushing facility supply water toward 45C to delete the chiller plant. Ask for the NPSH available against NPSH required at the top of the operating range and at the warmest facility water the site will ever run, not at catalog conditions.
The procurement question that costs money later is spares, and headquarters is a bad proxy for it. Grundfos, Wilo, KSB and Sulzer all manufacture in the United States; Xylem, Flowserve and ITT are US-headquartered; Armstrong is Canadian; Ebara is Japanese. Ask where the specific model number is built, where the seal kit and mechanical seal ship from, and what the stocked quantity is at the US distributor. Lead time on a new pump is measured in weeks. The seal kit for a pump that fails in year three, built on a European line, held by a distributor who does not stock it, is measured in whatever downtime a GPU hall can absorb, and nobody negotiates that clause at the top of a build.