Liquid cooling cuts data center energy

Data-center cooling just got a lot more interesting. The news is not that liquid cooling exists — that part is already old. The news is that a University of Illinois Urbana-Champaign team says a new kind of pure-copper cold plate, built with topology-optimized shapes and high-resolution metal printing, could push cooling energy down to about 1.1% of a data center’s total load. That is a huge jump from the rough “cooling eats 30% or more” world people still use as the mental model for air-cooled facilities. ### What actually changed? On May 7, the Illinois team published results in *Cell Reports Physical Science* showing a direct-to-chip liquid-cooling design that beat more conventional cold plates on both heat removal and pumping power. The key trick was not just switching from air to liquid. It was redesigning the tiny fins inside the cold plate so coolant moves through a shape that sheds heat better without demanding as much energy to circulate. (mechse.illinois.edu) ### What is a cold plate? A cold plate is the metal block that sits against a hot chip and carries heat into flowing liquid. In a direct-to-chip setup, you are not trying to chill the whole room first. You are grabbing heat right where the GPU or CPU makes it. That matters because modern AI chips are so power-dense that moving heat with room air alone starts to look wasteful and, at the high end, physically inadequate. (cell.com) ### Why is this better than normal liquid cooling? Most cold plates still use pretty simple internal geometries — rectangles, cylinders, basic fins. The Illinois group used topology optimization, basically software that keeps iterating toward a better answer, and then fabricated the resulting jagged pure-copper structures with electrochemical 3D printing. The result was a design that improved cooling performance over standard finned cold plates while also lowering the pumping energy penalty that usually comes with tighter, more aggressive liquid paths. (coolitsystems.com) ### Where does the 1.1% figure come from? It comes from the team’s data-center energy analysis, not from a giant production campus already running this exact hardware. Their estimate says that, under the study assumptions, cooling would account for only around 1.1% of total data-center energy use. One writeup translated that into a drop from about 550 MW of cooling power to 11 MW per gigawatt of compute load. So this is promising — but it is still a modeled system result tied to this specific design. (cell.com) ### Why does AI make this urgent? Because chip heat is rising faster than air-cooling economics can comfortably follow. CoolIT’s 2025 white paper lays out the direction of travel pretty clearly — GPUs already exceed 600 W, server CPUs are heading toward 500 W to 600 W, and rack densities above 70 kW are becoming normal enough that old airflow assumptions break down. Liquid cooling is moving from “nice for HPC” to “needed for dense AI.” (cell.com) ### Does this line up with what operators already see? Yes — in a broader sense. Earlier work from NVIDIA and Vertiv modeled hybrid air-liquid facilities and showed that as more of the IT load moves to liquid cooling, operators can raise water and air temperatures and cut overhead energy. And hyperscale operators are already proving that whole-facility efficiency can get very close to the theoretical floor — Google says its 2024 fleet-wide annual PUE was 1.09. (coolitsystems.com) The Illinois result fits that direction, even if its exact number is from a research model. ### What is the catch? Manufacturing and deployment. Fancy microstructured copper parts have to be made reliably, integrated into server designs, and supported by pumps, coolant distribution units, controls, and maintenance practices that work at scale. A better cold plate is not the whole stack. But turns out it is a very important piece of the stack, because if you can remove more heat at the chip with less pumping power, the rest of the facility gets easier too. (vertiv.com) ### Bottom line? Liquid cooling was already the direction of travel. This week’s Illinois result makes the case sharper: the next gains may come not from cooling the room better, but from cooling the chip much more intelligently. (mechse.illinois.edu) (cen.acs.org)