Liquid cooling, beyond the rack

The Open Compute Project's cold-plate work covers the technology cooling system from the cold plate to the coolant distribution unit, including tubing, manifolds and related interfaces. That scope is a reminder that direct-to-chip cooling involves more than a metal plate on a processor.¹

The U.S. Department of Energy illustrates how heat can move from an IT-side liquid loop through a coolant distribution unit into a separate facility loop and ultimately a cooling tower. The exact configuration varies. Separating the loops can help manage their different operating requirements; it does not make the heat disappear.²

For a review team, the first useful drawing is therefore a boundary diagram. Mark which party owns the cold plates, the distribution equipment, the building loop and the final heat-rejection equipment. Then put a meter, a maintenance obligation and a failure response at each boundary. This is our analytical framework, rather than a claim that all facilities use the same design.

Power usage effectiveness, or PUE, compares total facility energy with IT equipment energy over the same measurement boundary and period. A lower ratio indicates less facility overhead per unit of IT energy. It does not establish how much useful computation the IT equipment produced.

DOE's data-center design guide treats IT systems, environmental conditions, airflow, cooling, electrical systems and heat recovery as linked areas for efficiency improvement. It also cautions that no single guide supplies the most efficient design for every situation. The location, equipment and operating profile still matter.³

Our implication is that a procurement scorecard should pair energy measures with service measures. Include useful workload delivered, availability, maintenance time and the cost of replacing equipment. A facility could improve PUE while running underused servers, or support much more valuable work at a slightly higher overhead ratio. A single ratio cannot choose between those cases.

Consider a fictional installation carrying a constant 10 MW IT load for a full year. At a PUE of 1.40, total facility energy is 122,640 MWh. At 1.20, it is 105,120 MWh. The difference is 17,520 MWh per year. At an assumed electricity price of $80/MWh, the gross energy-cost saving is $1,401,600 per year.

This is arithmetic, not a measured liquid-cooling result. The PUE values, load and electricity price are hypothetical. We assume the same IT work and operating hours in both cases, so the comparison isolates facility overhead. In practice, load profiles, weather, pump power and equipment controls can change the result.

A retrofit costing an assumed $5 million would have a simple gross energy-only payback of about 3.6 years under those conditions. That number excludes financing, maintenance changes, downtime, taxes and the value of added capacity. It is not a complete investment return. If the retrofit cannot be installed without interrupting a valuable service, the implementation schedule may matter as much as the annual energy saving.

A closed technology loop can circulate the same fluid while the external heat-rejection system consumes water. Conversely, a design can use a different heat-rejection approach. The presence of liquid at the processor is therefore not enough to infer a site's water consumption.

DOE's cooling-water guidance connects water use to the heat load and the efficiency of successive heat-removal steps. Its discussion includes cooling-tower operations and the value of reducing the amount of heat that must be rejected. This supports evaluating the whole facility, rather than describing any liquid-cooled installation as automatically water efficient.²

Reliability also requires a separate argument. Ask what a leak detector triggers, which valves isolate a fault, how a failed distribution unit is serviced and whether qualified staff and spare parts are available. The relevant outcome is not simply that the loop operated in a demonstration. It is that the facility can continue meeting its service obligations through predictable maintenance and credible fault conditions.

Start with a baseline using the same workload, ambient conditions and measurement boundaries. Request time-series data rather than a single favorable operating point. Document which pumps, fans and supporting equipment are inside each energy calculation. A claimed saving without a boundary is difficult to compare.

Next examine the installation program. A product that performs well in a new building may require extensive plumbing, structural changes or commissioning work in an existing one. A repeatable bill of materials and a documented service procedure can make expansion easier to plan, but neither should be inferred from the technical prototype alone.

Finally, connect the engineering evidence to a contract. Identify who warrants the operating envelope and who pays when an equipment change requires a new fluid, connector or control strategy. The strongest thesis combines measured performance with an economical path to installation and long-term support. A technology trend by itself is not a margin forecast.

This sample combines primary technical guidance with original evaluation questions. It does not certify a cooling design, estimate an industry market size or reproduce proprietary performance testing. The cover's simplified pipes and cooling tower are illustrative, not a statement that every facility uses those components.

The numerical example assumes constant load, 8,760 annual hours and a uniform electricity price; it omits demand charges and time-of-use effects. All figures labeled illustrative are author-selected inputs or their calculated outputs. Sources were checked on 22 October 2026. A real project requires site-specific thermal, mechanical and electrical engineering.