Accelsius says two-phase cooling can cut AI data center energy costs

Accelsius says its two-phase liquid cooling system can let AI data centers run facility water loops as hot as 54 degrees Celsius, or 59 degrees Celsius with double the flow rate, potentially reducing annual energy costs. The company did not disclose customer deployments or pricing, and its claims come from research it commissioned and a white paper shared with The Register.

The pitch is aimed at operators trying to cool high-density GPU racks without adding more chillers, power draw or water use. Accelsius claims each 1 degree Celsius increase in facility water temperature can translate to about 4% in yearly energy savings. In a market where AI clusters are pushing rack power sharply upward and utility capacity is becoming a gating factor, even modest efficiency gains are commercially relevant.

Accelsius CTO Rich Bonner told The Register that low-grade 30 degree Celsius waste heat has limited reuse value, while 50 to 60 degree Celsius heat is closer to shower-water temperatures. Higher loop temperatures also make it easier to use dry coolers, which are industrial radiator-like systems, rather than evaporative cooling systems that consume water. Dry coolers are more useful when the outside air is sufficiently cooler than the facility water, so they are less effective in warmer seasons.

What Accelsius tested

Two-phase cooling is already familiar in data center R&D and some commercial systems. Microsoft showed immersion systems using boiling fluoroketones and hydrofluoroethers in 2021, and ZutaCore uses two-phase fluids in direct-to-chip cooling. Accelsius is not proposing a tank-based immersion design. Its approach resembles a conventional liquid-cooled rack: coolant moves through rack manifolds, passes through finned cold plates attached to servers and other hardware, then returns heat to a coolant distribution unit, or CDU, which exchanges that heat with the building water loop.

The difference is the fluid. Single-phase coolants remain liquid while carrying heat away. Two-phase fluids can change phase as they absorb heat, similar to the refrigerants used in air conditioners and refrigerators.

In Accelsius’ white paper, the company said it modified a Dell PowerEdge XE9680L server with eight Nvidia B200 accelerators to use its NeuCool two-phase coolant and cold plates. Accelsius reported temperature reductions of up to 19 degrees Celsius at the cold plate and 9 to 14 degrees Celsius at the system level versus the prior setup.

Bonner told The Register that operators could use the thermal headroom in different ways, including raising facility water temperatures, reducing or eliminating chiller use in some conditions, shifting saved power toward additional compute, or pursuing more chip performance.

Adoption is the hard part

The technical case still runs into familiar liquid-cooling problems: custom hardware, service responsibility and OEM certification. Accelsius’ B200 test used custom cold plates and a CDU designed for two-phase fluids. Bonner said universal cold plates could let OEMs ship systems compatible with either single-phase or two-phase fluids, but that remains an industry work item.

The challenge grows with rack-scale AI systems. Nvidia’s NVL72 includes 72 GPUs, 36 CPUs, 18 switch ASICs and dozens of NICs, all liquid-cooled. Nvidia rack systems are expected to reach about 250 kW this year and approach 600 kW beginning in 2027. Accelsius’ largest current in-row CDU supports one 250 kW rack, according to The Register. Bonner said the company is developing a unit above 350 kW and has a 2.4 MW CDU in the lab that he expects to become available early next year.

Coolant chemistry is another constraint. Some early two-phase cooling systems used 3M Novec, a PFAS-based fluid. 3M said in late 2022 that it would stop making those chemicals, creating supply and environmental concerns for vendors that depended on them. Bonner said Accelsius’ NeuCool refrigerant blend meets ASHRAE A1 safety requirements, meaning non-flammable and non-toxic.

The commercial test is whether server makers such as Dell, HPE, Lenovo or Supermicro support the design directly. Without OEM approval, enterprise buyers can face warranty and support disputes if a modified liquid-cooled system fails. Bonner said Accelsius wants customers to be able to buy a two-phase system through vendors such as Dell or Supermicro. Until that happens at scale, the company’s cooling gains remain a promising engineering result rather than a default option for AI infrastructure buyers.

This story draws on original reporting from The Register.