Executive Overview
The rapid, relentless scaling of artificial intelligence (AI) and high-performance computing (HPC) has fundamentally rewritten the rules of data center engineering. For decades, traditional air cooling—supplemented by chilled water loops and raised floors—served as the baseline standard for enterprise and hyperscale facilities. However, the emergence of dense, power-hungry server architectures driven by modern graphics processing units (GPUs) has pushed traditional thermal management systems to their absolute limits.
At the recent Data Cloud event in Austin, industry leaders converged to address a looming crisis: how to sustainably cool server racks that routinely exceed 100 kilowatts (kW) of heat dissipation. Among the prominent voices charting the future was Josh Claman, Executive Chairman of Accelsius, who sat down with interviewer Saf Malik for an in-depth conversation.
Claman’s message was both a warning and a blueprint: traditional single-phase water cooling systems are rapidly approaching a technological ceiling. To unlock the full potential of next-generation AI infrastructure, the data center industry must embrace two-phase liquid cooling. According to Claman, this advanced thermal architecture—characterized by superior thermodynamic efficiency, lower flow rates, and remarkable heat transfer capabilities—is poised to transition from a niche alternative to the dominant cooling architecture for AI infrastructure by 2030.
This article explores the core insights shared in Austin, analyzing the mechanical superiority of two-phase cooling, the implications of NVIDIA’s 45°C water-cooling benchmark, regulatory and maintenance considerations, and the structural shift required to future-proof the global digital economy.
Detailed Chronology: The Evolution of Data Center Cooling
To understand why two-phase liquid cooling is commanding the attention of executive boards and engineering teams alike, it is necessary to examine the historical trajectory of data center thermal management.
Era 1: The Air-Cooled Status Quo (1990s–2010s)
For the early history of enterprise computing, central processing units (CPUs) generated manageable levels of heat—typically ranging between 70W and 150W per socket. Facilities relied heavily upon Computer Room Air Conditioner (CRAC) and Computer Room Air Handler (CRAH) units. Cold air was pushed under raised floors, forcing its way up through perforated tiles and across rows of server cabinets. This system operated on sensible heat transfer, where air temperature increased as it absorbed thermal energy from the hardware.
Era 2: The Density Wall and Single-Phase Liquid Emergence (2010s–2020s)
As Moore’s Law slowed and chip designers turned to multi-core architectures, specialized accelerators, and accelerated parallel processing, thermal design power (TDP) per rack began climbing exponentially. Standard air-cooling configurations could no longer deliver enough mass flow of air without requiring prohibitively high fan speeds and massive energy expenditures.
This gave rise to single-phase direct-to-chip liquid cooling. In these systems, a liquid coolant (typically water mixed with glycol) is pumped directly through cold plates mounted on high-heat components like CPUs and GPUs. While single-phase liquid cooling represents a massive leap forward compared to air, it still relies on sensible heat capacity: the liquid warms up as it travels through the loop, necessitating large volumes of fluid, higher pumping pressures, and complex distribution networks.
Era 3: The Generative AI Boom and the Shift to Two-Phase Cooling (Present–2030)
The deployment of hyperscale generative AI clusters—featuring dense arrays of advanced GPUs running continuous training and inference workloads—has pushed individual rack power densities from 10 kW to 40 kW, 100 kW, and beyond. Traditional single-phase plumbing is increasingly strained by these loads, requiring immense flow rates and intricate manifold layouts.
Enter two-phase liquid cooling. As discussed by Josh Claman at Data Cloud Austin, this technology leverages the physics of phase change (latent heat of vaporization) rather than simple sensible heating. By utilizing specialized engineered fluids that boil at low temperatures directly at the chip interface, absorb massive amounts of heat as they transition from liquid to gas, and then condense back into liquid in a closed loop, two-phase systems achieve unprecedented thermal performance. Claman notes that this technological maturation is setting the stage for widespread, mainstream adoption across hyperscale facilities as the industry races toward 2030.
Supporting Context & Metrics: The Mechanics of Two-Phase vs. Single-Phase Cooling
The debate between single-phase and two-phase liquid cooling ultimately boils down to thermodynamics, energy efficiency, and operational risk management.
Thermodynamic Superiority: Latent vs. Sensible Heat
In a single-phase cooling system, heat is removed via sensible heat capacity. The fluid’s temperature rises proportionally to the heat absorbed ($Q = m cdot c cdot Delta T$). To remove massive heat loads, operators must increase either the mass flow rate ($m$) or the temperature differential ($Delta T$), both of which have practical limits. Higher flow rates demand larger pumps, thicker hoses, and more electrical power, eating directly into the facility’s Power Usage Effectiveness (PUE).
Conversely, a two-phase cooling system capitalizes on latent heat. When the engineered dielectric fluid reaches its boiling point at the surface of the thermal evaporator, it undergoes a phase transition into vapor. The amount of energy required to change a liquid into a gas (latent heat of vaporization) is vastly higher than the energy required to raise the temperature of that same liquid by a few degrees. Consequently, two-phase systems can transfer massive thermal loads using a fraction of the fluid volume and significantly lower flow rates.
The NVIDIA 45°C Water-Cooling Benchmark
A critical focal point of Claman’s discussion in Austin centered around hardware manufacturer benchmarks, notably NVIDIA’s 45°C water-cooling benchmark.
As GPUs like NVIDIA’s latest architectures demand higher operating thresholds, thermal engineers are pressured to supply cooler water to maintain optimal junction temperatures and prevent thermal throttling. However, chilling water down to low temperatures requires energy-intensive chiller plants, negating many of the green efficiencies data center operators strive for.
NVIDIA’s benchmarks demonstrate that hardware can maintain high performance and longevity even when coolant entering the system is kept at warmer baseline temperatures (such as 45°C). This is where two-phase technology excels. Because two-phase coolants operate via phase change at tightly controlled, uniform temperatures, they maintain exceptional thermal stability across the entire die surface, eliminating hotspots regardless of incoming water loop temperatures. This synergy allows data centers to operate "free cooling" towers for longer portions of the year, drastically reducing reliance on mechanical chillers and lowering overall energy overhead.
Addressing Industry Concerns: Refrigerants, Maintenance, and Safety
Despite its obvious thermodynamic advantages, two-phase cooling has historically faced skepticism from cautious data center operators regarding environmental impact, maintenance complexity, and fluid safety. During the Data Cloud session, Josh Claman addressed these head-on.
The Regulatory Landscape and Refrigerants
Earlier generations of two-phase cooling relied on fluorochemicals that faced stringent environmental scrutiny due to global warming potential (GWP) and persistence in the environment (such as PFAS regulations).
Claman highlighted how the industry has pivoted toward next-generation, environmentally sustainable engineered fluids. Modern two-phase working fluids are designed to meet rigorous environmental standards, offering zero ozone depletion potential (ODP) and significantly reduced GWP profiles. Furthermore, because two-phase systems operate as closed-loop architectures, the fluid is hermetically sealed within the system, drastically mitigating the risk of fugitive emissions or environmental leakage.
Maintenance and Operational Simplicity
A common misconception is that two-phase systems are overly complex to maintain, requiring specialized on-site chemical engineers. Claman countered this by explaining that modular two-phase architectures—such as those developed by Accelsius—are engineered for plug-and-play deployment.
Because two-phase systems operate at lower internal pressures and flow rates compared to high-pressure single-phase water pumps, the risk of catastrophic leaks is substantially reduced. If a server blade requires servicing, modular cold plates and evaporator units can be disconnected cleanly via drip-less quick-disconnect couplings without exposing technicians to hazardous fluids or draining entire facility loops.
Future Outlook: The Road to 2030
As the digital economy hurtles toward 2030, the intersection of artificial intelligence, edge computing, and power grid constraints will dictate the evolution of data center design.
- Power Density Milestones: Industry analysts project that standard AI server racks will soon exceed 150 kW to 300 kW. At these densities, air cooling is physically non-viable, and single-phase liquid cooling will encounter pumping and piping bottlenecks due to the sheer volume of water required.
- Mainstream Hyperscale Adoption: As outlined by Accelsius leadership in Austin, two-phase liquid cooling is moving out of the early-adopter phase. Hyperscalers, colocation providers, and enterprise data center operators are increasingly standardizing liquid-ready white space.
- Grid Efficiency and Sustainability: With regulatory bodies imposing strict carbon-neutrality mandates and power grids facing unprecedented strain from AI workloads, efficiency is paramount. Two-phase systems offer a pathway to lower PUE metrics, higher server utilization rates, and reduced water consumption compared to evaporative cooling towers.
Conclusion
The conversation between Saf Malik and Josh Claman at Data Cloud Austin underscored a pivotal reality: the thermal challenges posed by artificial intelligence cannot be solved with yesterday’s technology. While single-phase water cooling has provided a necessary bridge, the exponential climb in rack densities demands a more sophisticated thermal architecture.
Through advanced physics, superior thermodynamic efficiency, and alignment with modern hardware benchmarks like NVIDIA’s, two-phase liquid cooling is proving to be the definitive solution for the next era of computing. As the industry marches toward 2030, companies that embrace this transition will secure a decisive competitive advantage in performance, sustainability, and operational resilience.
