Data centers consumed 485 TWh of electricity in 2025. Roughly 30 percent of that, about 145.5 TWh, went to cooling. That number exceeds Sweden's total annual power consumption. It is not a rounding error. A single NVIDIA GB200 chip draws 1,200 watts and dissipates every watt of that as heat. xAI's Colossus 1 runs 220,000 GPUs at 300 megawatts combined. Without active cooling, that hardware reaches 1,200°C within an hour.
Researchers at the University of Illinois Urbana-Champaign have published results that change what a cold plate fin structure can look like and how much work it can do. Behnood Bazmi and Nenad Miljkovic combined two methods that do not usually appear together in the same fabrication pipeline: topology optimization, which uses algorithms to generate complex internal geometries no human designer would specify, and electrochemical additive manufacturing, which 3D-prints pure copper at feature resolution as fine as 30 to 50 micrometers. The results are published in Cell Reports Physical Science.
Conventional cold plate fins are rectangular or cylindrical. Those shapes are constraints of subtractive manufacturing, not constraints of physics. Topology optimization removes the manufacturing constraint and asks the algorithm to find whatever internal geometry maximizes heat transfer while minimizing fluid resistance. The output is irregular and jagged, the kind of structure that looks wrong to someone trained on machined parts but is thermally correct.
The electrochemical additive manufacturing process, ECAM, deposits pure copper at those algorithmic geometries with feature resolution down to 30 micrometers. That is roughly one-third the width of a human hair. Conventional copper machining cannot produce internal channels at that scale. The combination of the two methods unlocks a design space that was previously fabrication-limited.
Against a conventional cold plate baseline, the UIUC results show 32 percent better cooling performance and 68 percent reduction in pressure drop. Both numbers matter. Better cooling reduces the facility-side thermal load. Lower pressure drop reduces pump energy and allows higher flow rates without proportional energy cost increases.
The researchers applied their optimized cooling model to a conventional 1-gigawatt air-cooled facility as a comparison baseline. That facility would require roughly 550 megawatts of cooling energy, more than half the total facility power draw going to move heat rather than perform compute. The optimized liquid cooling system in the UIUC model reduces facility cooling energy to approximately 11 megawatts. Cooling drops from 30 to 35 percent of facility consumption to 1.1 percent. The projected PUE under that scenario is 1.011, near the theoretical limit for any real infrastructure.
The industry talks about PUE improvement in tenths of a point. Moving from a PUE of 1.4 to 1.3 is presented as a serious operational achievement. The UIUC number puts a ceiling on what optimized liquid cooling can produce at scale: essentially zero overhead per watt of compute delivered.
ECAM is not a production process at data center scale. The fabrication technique is validated at the component level. Moving from laboratory-produced copper cold plates to high-volume manufacturing at the throughput a hyperscale build requires is a different problem with its own lead time, tooling cost, and supply chain structure.
The research establishes a performance target and a fabrication path. What it does not establish is a delivery timeline. Companies like ACT, which just announced a 500,000-unit annual cold plate production ramp in Lancaster, Pennsylvania, are scaling proven single-phase and two-phase cold plate architectures using conventional methods. The UIUC result shows where the physics ceiling is. The production question is how quickly the manufacturing process catches the geometry.
Cooling represents the single largest efficiency opportunity in AI infrastructure. The GPU wattage roadmap is public. Nvidia's Vera Rubin Ultra targets one megawatt per rack. At that density, the difference between a PUE of 1.4 and 1.011 is not a footnote in an annual sustainability report. It is hundreds of megawatts of avoided generation capacity across the buildout. The UIUC paper points at what liquid cooling can become when the geometry is no longer constrained by what a mill can cut.