Cold Plates and Coolant Distribution Units

In 2025, Google completed the contribution of a coolant distribution unit specification, which it calls Project Deschutes, to the Open Compute Project, or OCP. The Open Compute Project described it as the first coolant distribution unit specification contributed to the community, and the design is rated for about two megawatts of cooling. Equipment makers showed their own builds of the design at the OCP Global Summit that October. A shared specification for a piece of plumbing may seem an odd thing to celebrate. It makes sense once you see where that unit sits in the chain.

Direct-to-chip liquid cooling, sometimes called direct liquid cooling, replaces the air-cooled heat sink on the hottest components with a cold plate. A cold plate is a block of metal, usually copper, with fine internal channels. It is clamped to the processor with a thermal interface material between them, and a liquid coolant flows through the channels, picking up heat directly from the package. Tubes carry the coolant from each cold plate to a manifold at the back of the rack, and quick-disconnect fittings let technicians remove a server without draining the loop. The coolant is typically water treated with additives, often a mix with propylene glycol, chosen to resist corrosion and biological growth.

The loop that runs through the servers is deliberately kept small and clean. ASHRAE, the American Society of Heating, Refrigerating and Air-Conditioning Engineers, calls it the technology cooling system. It does not connect directly to the building's larger water system, which ASHRAE calls the facility water system. The two loops meet in a heat exchanger inside the coolant distribution unit, or CDU. The CDU also contains pumps that drive the technology loop, filters that keep particles out of the fine channels in the cold plates, and controls that hold temperature and pressure steady. Google's Deschutes specification, for example, includes fine side-stream filtration and redundant power feeds to the pumps, which tells you what the designers were worried about.

Separating the loops is the most important design choice in the whole arrangement. Facility water may come from a cooling tower, a dry cooler or a chiller plant, and it can carry dirt, scale and microbes that would clog a cold plate. Keeping the server loop isolated lets operators control its chemistry tightly and limits the volume of liquid that could leak near electronics. The CDU is the firewall between building plumbing and computer plumbing.

Now to temperature, which is where the efficiency case lives. The ASHRAE Thermal Guidelines define liquid cooling classes for the facility water supply, labeled W17, W27, W32, W40 and W45, where the number is the maximum supply temperature in degrees Celsius, plus an open-ended class called W+. Because a cold plate sits in direct contact with the chip, it can do its job with surprisingly warm coolant. If the facility water can be supplied at 30 degrees Celsius or more, the building can reject heat to outdoor air with dry coolers for most or all of the year in many climates, with little or no refrigeration. Warm-water cooling is the quiet reason liquid can improve efficiency and not merely capacity.

There is a catch. Chip makers want cooler coolant as their parts get hotter, because it gives them more thermal margin, while facility engineers want warmer coolant because it saves energy and water. That negotiation is live, and specifications published by hardware vendors, standards bodies and the Open Compute Project are where it gets settled. A 2024 white paper published through the Open Compute Project, with contributors including AMD, Intel, Meta, NVIDIA and Vertiv, argued for a coolant temperature of about 30 degrees Celsius as a durable planning point. Faraday's read is that anyone designing a building today should find out which supply temperatures the hardware on their roadmap actually requires, because the answer decides whether they need chillers.

A final practical point concerns the heat that cold plates do not capture. A cold plate covers the processors and often the memory stacks beside them, but power supplies, network cards and other components still shed heat into air. Vendors describe the share captured by liquid differently from product to product, and quoted figures vary, so treat any single percentage with caution. The design consequence is that a liquid-cooled hall still needs some air handling, sized for the residual load, which is why the hybrid hall described in the previous module is the norm.

Put together, direct-to-chip cooling is less a new machine than a rearrangement of old ones. Pumps, heat exchangers, filters and pipes are familiar to any mechanical engineer. What changed is that they now reach the rack, and the building has to be drawn around them.