Use case
How to compare direct-to-chip, immersion and rear-door options for dense GPU halls, and the site questions that decide between them.
6 tools in this scenario · verified 2026-10-27
A rack of current-generation accelerators pulls more heat from a smaller footprint than the air handling in most existing halls was designed for, so the cooling decision arrives before the hardware order. The options split into direct-to-chip cold plates fed by a coolant distribution unit, immersion in a single-phase or two-phase tank, and rear-door heat exchangers that stay air-side but raise the density an air hall can carry. Each changes a different part of the site: cold plates need facility water and piping to the rack, immersion needs tank space and maintenance access, rear-door units need flow and plant capacity rather than floor space. The comparison is therefore less about the cooling technology than about the building, the service model and the willingness to move heat recovery into the design. Vendors also span the stack: some sell the entire loop, others sell components into a partner's design. The criteria below are the ones that decide which of those arrangements fits an existing hall.
1
Heat rejection path
Establish whether the loop ends in a facility cooling tower, a dry cooler, a chiller or heat reuse. A liquid loop that still dumps heat into the same plant may raise density without improving efficiency.
2
Facility water and piping readiness
Direct-to-chip designs assume supply and return water at defined temperatures and pressures. Check the plant can deliver them, and who is responsible for the piping between the CDU and the rack.
3
Service and maintenance model
Compare how a failed pump, a leak or a coolant top-up is handled, whether service is on-site by the vendor or by the operator's team, and whether the contract covers rapid replacement of the cooling unit.
4
Coverage across the stack
Some vendors supply CDUs plus manifolds plus cold plates, others only one layer. A component supplier may be the better fit when the hall is designed by a partner, but it moves integration risk to that partner.
5
Efficiency evidence beyond marketing
Ask for measured partial PUE or pPUE figures at reference densities, with the boundary stated, rather than a single best-case number. Check whether the figure includes the CDU pumps.
Data Center Sustainability
Liquid cooling hardware for high-density compute: coolant distribution units, piping and coldplates, now part of Ecolab.
Data Center Sustainability
Liquid cooling hardware for AI data centres: coolant distribution units and immersion systems from single racks to gigawatt builds.
Data Center Sustainability
Two-phase direct-to-chip liquid cooling for AI racks, claimed to cut annual OpEx by 35% against single-phase.
Data Center Sustainability
Dutch immersion cooling systems built as one architecture for compute, cooling, power and control at high density.
Data Center Sustainability
Waterless two-phase direct-to-chip cooling for AI accelerators, claimed to cut cooling energy by about 74%.
Data Center Sustainability
DCIM suite for data center assets, capacity, power and environmental monitoring.
Last verified 2026-10-27. Figures on the linked tools are checked against the sources above; confirm pricing and compliance coverage before signing.