Two-phase cooling could cut AI datacenter costs

Liquid cooling is already essential for multi-kilowatt AI accelerators and increasingly dense server racks, but Accelsius argues the industry is still using the wrong fluids. In research commissioned by the startup, the company says switching from single-phase coolants to two-phase fluids could let datacenters run facility water temperatures up to 14°C higher, trimming operating costs by millions.

According to Accelsius CTO Rich Bonner, its approach can support facility water temperatures of up to 54°C, or up to 59°C if an operator is willing to double a facility’s flow rate. The company says every 1°C increase in facility water temperature translates to roughly 4 percent in annual energy savings.

“There’s not that much you can do with a lot of 30°C waste heat, but as you get to 50°-60°C, now you’re talking about shower temperature levels of heat.”

Two-phase cooling is not new. Microsoft was experimenting as far back as 2021 with immersing full server chassis in boiling fluoroketones and hydrofluoroethers, while ZutaCore has been using two-phase fluids in its own direct-to-chip cooling systems.

Accelsius' setup looks closer to a standard liquid-cooling design. Fluids are pumped through rack manifolds into finned cold plates attached to servers, then return heat to a coolant distribution unit (CDU), where it is exchanged with facility water before being expelled to the atmosphere. The difference, the company says, is that its two-phase coolant removes heat far more efficiently.

In a whitepaper shared with The Register, Accelsius said it modified a Dell PowerEdge XE9680L equipped with eight Nvidia B200 accelerators to use its NeuCool coolant and cold plates. The result: temperatures dropped by as much as 19°C at the cold plate and 9°-14°C at the system level.

Recommended reading

UK data centre boom hits a water reality check

Bonner said operators could use that thermal headroom in different ways: raise facility water temperatures and potentially avoid running chillers, use the power savings to add more compute, or push chips harder for more performance. Higher facility water temperatures also reduce the amount of heat that needs to be rejected to the atmosphere, making dry coolers more practical and reducing dependence on water-consuming evaporative coolers.

Barriers to adoption

The catch is hardware compatibility. Accelsius' B200 test system used custom cold plates and a CDU designed specifically for two-phase fluids. That is manageable on a relatively simple system, but much harder on platforms like Nvidia’s NVL72, which combines 72 GPUs, 36 CPUs, 18 switch ASICs, and dozens of NICs under liquid cooling.

Bonner said universal cold plates could eventually let OEMs build platforms that support either single-phase or two-phase coolants. But even if servers become more flexible, two-phase systems still need specialized CDUs sized for modern AI racks.

That matters because rack power is climbing fast. Nvidia systems are reaching 250 kW per rack this year and are expected to push 600 kW beginning in 2027. At the time of writing, Accelsius' largest in-row CDU supports only one 250 kW rack. Bonner said the company is developing a 350-plus kW unit and has a 2.4 megawatt system in the lab that he expects to be available early next year.

There is also the coolant question. Early two-phase systems often depended on fluids such as 3M Novec, based on PFAS chemicals. In late 2022, 3M said it would stop producing those chemicals, forcing suppliers to look for alternatives. Bonner said Accelsius' NeuCool refrigerant blend meets ASHRAE A1 safety requirements, meaning it is non-flammable and non-toxic.

The final hurdle may be vendor support. Retrofitting servers for liquid cooling is not usually the hard part; getting Dell, HPE, Lenovo, Supermicro, or other OEMs to stand behind those systems is. Without that approval, customers risk warranty disputes and expensive downtime. Bonner told The Register that Accelsius' goal is clear: a two-phase system customers can buy directly from Dell or Supermicro.