Accelsius Expands Austin Lab for Megawatt-Scale AI Cooling Tests

The drive to improve data center cooling continues. 

Austin-headquartered Accelsius has upgraded its Springdale lab in Austin, Texas, with multi-megawatt electrical capacity to test its two-phase liquid cooling system at the loads expected from future AI racks.

The company said it will add 18,000 square feet to its leased space at the site in January 2027.

The added power lets Accelsius test one rack above a megawatt or several racks at lower loads. The company plans to measure its NeuCool system from the plates that cool chips through the coolant distribution unit and the building’s water loop. Those full-system tests should give customers a clearer view of how the equipment performs at higher rack densities and warmer facility water temperatures.

“With multi-megawatt capacity at Springdale, we can run full racks at megawatt-class loads on warm facility water,” Accelsius CEO John Hewitt said in the company’s announcement. Innventure founded Accelsius in 2022 to commercialize cooling technology originally developed at Nokia Bell Labs. 

The company’s expansion comes at a critical time for the U.S. data center buildout. Companies operating or leasing the facilities are seeking ever-growing amounts of energy, but must manage ever-growing amounts of heat produced in high-end computing. 

At the same time, local communities are opposing data centers because of the power needed to run and cool them—as well as well as massive water consumption needed for hyperscale cooling needs. 

AI servers pack more electrical power into each rack, and nearly all of that power becomes heat. Moving heat away from the chips is only the first step. Pumps, distribution equipment and the building’s cooling system must then carry it out of the data hall. The U.S. Department of Energy’s data center design guide describes direct-to-chip cooling as a way to collect heat close to its source, where air systems face a harder task as equipment density rises.

Accelsius uses a fluid that boils as it absorbs heat at the chip and condenses after that heat is transferred away. This two-phase process differs from liquid systems that circulate fluid without a phase change. Neither approach eliminates the need to reject heat outside the building. Energy use and water demand depend on the entire cooling design, including the facility loop and the conditions under which it runs.

The Austin site gives Accelsius a place to test warm-water operation during Central Texas summers. The company pointed to benchmark results it released in July showing lower NVIDIA B200 chip temperatures than a single-phase system under specified warm-water conditions. Those results do not, by themselves, establish annual energy savings at a customer data center; the expanded lab can test performance across racks and supporting equipment.

Accelsius also said the lab will support customers in its HyperStart validation program. It reported that several hyperscale operators are engaged, but did not identify them or disclose resulting orders. Separately, the company plans a 375-kilowatt unit in its NeuCool multi-rack line, above its existing MR250. It said more details will follow.

For data center operators, the useful measure will be whether cooling capacity demonstrated in Austin holds up across a complete system at the water temperatures and power loads their projects require. The lab expansion gives Accelsius the power to run those tests; the company has not yet published their results.