Cooling the AI Boom: The High-Stakes Shift to Two-Phase Liquid

AUSTIN, Texas – November 03, 2026 – As artificial intelligence workloads drive rack power densities far beyond the limits of traditional air cooling, the physical infrastructure of the internet is reaching a boiling point. The rapid deployment of high-performance computing has forced a reckoning in data center design, shifting the industry's focus from mere processing power to the fundamental thermodynamics of heat removal.

At the center of this transition is Accelsius, an Austin-based thermal management company founded by Innventure, Inc. Today, Accelsius announced its participation in the National Science Foundation (NSF) Center for Energy-Smart Electronic Systems (ES2). As part of this federal data center energy efficiency research program, the company is providing its NeuCool IR150—an integrated IT rack and coolant distribution unit—to Villanova University's Laboratory for Advanced Thermal and Fluid Systems.

The installation, scheduled for early October 2026, represents much more than a standard equipment donation. It is a critical stress test in a broader, high-stakes race to commercialize two-phase direct-to-chip liquid cooling. With the backing of hyperscale giants and federal research dollars, the initiative aims to prove that boiling dielectric fluids can sustainably cool the next generation of artificial intelligence without triggering an environmental backlash.

The Consortium Racing to Avert an Energy Crisis

The explosive power requirements of AI accelerators have forced technology giants like Microsoft, Meta, and Google into joint academic testbeds with emerging cooling startups. The NSF ES2 operates as an Industry-University Cooperative Research Center, and its stated objective is nothing short of transformative: to improve the energy efficiency of United States data centers by 20 to 35 percent. By the consortium's own estimates, achieving this efficiency gain would be the equivalent of removing one to two power plants from the grid every three years.

“The largest challenges in data center energy efficiency are not going to be solved by any one group— they require sustained collaboration between research communities and the manufacturers building AI infrastructure at scale,” said Dr. Richard Bonner, Chief Technology Officer of Accelsius. “Our participation in ES2 continues a tradition at Accelsius of prioritizing and publishing research that brings facility energy consumption down.”

Villanova is one of three university sites participating in ES2, alongside The University of Texas at Arlington, where Accelsius already supports the COOLERCHIPS project led by Dr. Dereje Agonafer under ARPA-E. These academic laboratories serve as neutral proving grounds. By subjecting proprietary hardware to rigorous, independent system-level testing, researchers can validate vendor claims regarding thermal resistance, Power Usage Effectiveness (PUE), and Water Usage Effectiveness (WUE). For hyperscalers looking to deploy new cooling architectures across millions of square feet of server space, this independent validation is the ultimate prerequisite.

Breakthrough Efficiency or an Operational Gamble?

The technical debate currently dividing data center architects centers on the choice between single-phase and two-phase liquid cooling. As of 2026, single-phase direct-to-chip cooling—which pumps a liquid like a water-glycol mixture through cold plates—dominates the market with an estimated 55 percent share. It is a mature, well-understood technology. However, it is fundamentally limited by the sensible heat capacity of water, requiring high flow rates, massive pumps, and the inherent risk of placing water adjacent to multi-million-dollar computing clusters.

Two-phase direct-to-chip cooling, the architecture championed by Accelsius, takes a radically different approach. It utilizes a closed-loop system containing a dielectric refrigerant. As the fluid passes over the hot processor, it boils, absorbing massive amounts of thermal energy through the phase change from liquid to vapor (latent heat transfer). The vapor then travels to a condenser, returns to a liquid state, and repeats the cycle.

This method is remarkably efficient. Accelsius claims its NeuCool system can cool upwards of 4500 watts per socket with a thermal resistance of just 0.020°C/W, targeting a facility PUE of 1.08. Furthermore, because the refrigerant is dielectric, a leak will not short-circuit the IT equipment.

“We’re excited that Accelsius is joining the ES2 Center, which has become a leading research site for efficient cooling, WUE optimization, and waste heat reuse,” said Dr. Alfonso Ortega, James R. Birle Professor of Energy Technology and ES2 Site Director at Villanova University. “This donation of Accelsius technology will provide our researchers with a rack-level understanding of two-phase direct-to-chip as a scalable, sustainable option to address rising thermal demands.”

Yet, despite the thermodynamic advantages, two-phase cooling introduces profound operational complexities. The systems require precise pressure management to facilitate the boiling process, and the industry has historically struggled with fluid containment and maintenance protocols at hyperscale.

The Chemical Elephant in the Server Room

The most significant hurdle facing two-phase liquid cooling is not thermodynamic, but chemical. The dielectric refrigerants historically used in these systems belong to a class of chemicals known as PFAS (per- and polyfluoroalkyl substances). Dubbed "forever chemicals" due to their environmental persistence, PFAS are facing intense global regulatory scrutiny.

The Environmental Protection Agency (EPA) in the United States and the European Chemicals Agency (ECHA) are both aggressively moving to restrict or outright ban thousands of PFAS compounds. The market leader in engineered fluids, 3M, announced it would phase out PFAS manufacturing by the end of 2025, sending shockwaves through the thermal management sector.

Accelsius currently utilizes a fourth-generation PFAS coolant known as R1233zd. The company maintains that its fluid has a global warming potential of less than one and converts into trifluoroacetic acid (TFA), which it characterizes as a naturally occurring compound. However, environmental regulators increasingly view TFA as a hazardous, highly mobile short-chain PFAS. As one industry sustainability analyst noted, deploying closed-loop systems mitigates daily exposure, but fugitive emissions during maintenance and end-of-life fluid disposal remain critical liabilities for data center operators seeking to meet strict ESG mandates.

Recognizing this existential regulatory threat, the industry is pivoting. Accelsius recently shared promising results at the Open Compute Project (OCP) EMEA conference regarding a new, non-PFAS coolant fluid, R-4101A, developed in collaboration with SK Innovation. The ability of the NSF ES2 consortium to test and validate these next-generation, environmentally benign fluids will likely determine whether two-phase cooling can survive the impending regulatory cliff.

From Federal Grants to Hyperscale Deployments

Programs like NSF ES2 and ARPA-E's COOLERCHIPS serve as the vital bridge between academic theory and commercial reality. For a hardware startup, federal research sponsorship validates early-stage technologies, de-risking them for high-stakes enterprise deployments.

The commercial trajectory of Accelsius suggests this strategy is yielding dividends. Bolstered by its participation in the NVIDIA Inception program, the company started 2026 by securing a $65 million Series B funding round led by Johnson Controls, with participation from Legrand. This influx of capital has accelerated their manufacturing capabilities at their Austin, Texas facility and fueled product expansion, including the release of the NeuCool MR250, a row-based coolant distribution unit designed for hyperscale environments.

The ultimate proof of concept arrived late last year when Accelsius secured an agreement to deploy its NeuCool technology at a massive 300-megawatt data center campus being built by DarkNX in Ontario, Canada. Deployments of this magnitude indicate that enterprise IT, colocation providers, and AI neoclouds are increasingly willing to embrace the complexities of two-phase cooling to unlock unprecedented compute densities.

As the NeuCool IR150 comes online at Villanova University this October, the data generated will do more than just inform academic papers. It will provide the concrete operational benchmarks that the data center industry desperately needs. In the race to support the artificial intelligence revolution, the technology sector is learning that the future of computing depends entirely on how effectively we can manage the heat it leaves behind.