KRAMBU & Elkhorn Products have just re‑written the playbook for high‑density AI cooling. During a live test at the company’s Newport facility, the new Elkhorn 970 v21 hydrovaporization cooler displaced a conventional chilled‑water plant, removed every kilogram of synthetic refrigerant from the mechanical room, and cut cooling power by roughly half.
In legacy water‑cooled plants, bringing 7 °C supply water to a megawatt of liquid‑cooled GPUs typically costs about 170 kW of electrical power and drives facility‑level power‑usage effectiveness (PUE) to roughly 1.18. Air and adiabatic chillers fare even worse, demanding 250–270 kW and pushing PUE above 1.22. By contrast, the Elkhorn 970 v21 performs the same duty with approximately 85 kW, thanks to an effective coefficient of performance near 11.8. That single change drops the Newport site’s PUE to about 1.09 and delivers fifty‑percent energy savings versus the best chilled‑water plants, or sixty‑five percent versus typical air‑cooled infrastructure.
The step‑change in efficiency comes from four intertwined design choices. First, the Elkhorn cycle is entirely compressor‑free, relying on a low‑horsepower water pump instead of energy‑hungry vapor‑compression machinery and avoiding the twenty‑five‑percent heat‑of‑compression penalty. Second, it exploits water’s exceptionally high latent heat, which lets the system move the same thermal load with roughly half the mass‑flow and motor power required by hydrofluorocarbon refrigerants. Third, the process operates under vacuum, eliminating thick‑wall pressure vessels, oil‑separation hardware, and other parasitic components that plague conventional chillers. Finally, because the working fluid is simply R‑718—plain water—there are no high‑GWP chemicals onsite, no leak‑checks, and no direct Scope 1 emissions from cooling operations.
April’s technology demonstration underscored these gains. A two‑horsepower Elkhorn 970 prototype removed a three‑ton GPU load—about 10.5 kW—while drawing just 1.93 kW of electrical power. Scaled to a full megawatt of compute, that performance would save roughly 745 megawatt‑hours per year and sidestep about 300 metric tons of carbon dioxide equivalent, assuming a grid intensity of 0.4 kg CO₂e per kilowatt‑hour.
Travis Jank, CEO of KRAMBU, framed the result in blunt data‑center economics: “For AI workloads, megawatts are the new currency. By halving cooling energy and ditching Freon entirely, the Elkhorn 970 lets us invest more power budget where it matters—into GPUs that accelerate our customers’ models.” Heather Jones, CEO of Elkhorn Products, added that KRAMBU’s deployment “proves that natural‑refrigerant systems can outperform legacy chillers at scale” and marks the transition from laboratory validation to real‑world production.
The Newport pilot now serves as a template for hyperscale roll‑outs wherever liquid‑cooled AI clusters are straining electrical capacity and sustainability budgets. Organizations looking to double compute density without expanding utility infrastructure—or aiming to retire synthetic refrigerants altogether—have a new option ready for production. For more technical depth, Elkhorn’s white paper details the documented fifty‑to‑ninety percent cooling‑energy reduction across multiple load profiles, and both Elkhorn Products and KRAMBU are fielding requests for site demos.
