Liquid cooling is no longer a marginal choice for AI data centres. TrendForce projected its penetration to rise from 14% in 2024 to 33% in 2025. The Data Centre Trends Report 2026 from Soben, part of Accenture, reproduces these figures and expects liquid cooling adoption to reach 40% in 2026. The direction of travel is already clear. TrendForce describes liquid cooling as moving from early pilot projects to large-scale deployment as AI chip power and system density increase. In the highest-density environments, traditional air-cooling systems can no longer manage the full thermal load on their own, making it necessary to bring heat removal closer to the processors.
But liquid cooling does not offer a single answer. Direct-to-chip cold plates typically bring water-based coolant into the rack, while two-phase immersion cooling can depend on PFAS-based working fluids. Each addresses the thermal challenge by introducing a different trade-off. Between them, another cooling architecture is beginning to emerge. The question for AI infrastructure cooling is therefore no longer simply whether to use liquid, but which trade-off to accept.
4.0 %
projected liquid cooling adoption in 2026, according to Soben’s Data Centre Trends Report 2026.
Why direct-to-chip liquid cooling moved first
Direct-to-chip cooling removes heat close to its source through cold plates mounted on CPUs, GPUs, and other high-power components. Coolant flows through sealed channels inside the plates and remains confined within the cooling loop. Compared with immersion systems, cold plates can generally be integrated more easily into existing server and rack architectures. For liquid cooling retrofit projects, this relative compatibility reduces the redesign required and helps explain why direct-to-chip liquid cooling gained traction earlier.
The trade-off lies in the fluid used. Conventional cold-plate systems typically rely on water-based coolants, often water-glycol mixtures. Although the fluid remains confined within the loop, it still circulates inside the rack, close to electronic components.
Low facility water use is not no water at the rack
A recent fully liquid-cooled AI infrastructure design illustrates the difference. A closed loop and outdoor dry coolers can sharply reduce, or even eliminate, water consumption for heat rejection in suitable climates. Inside the rack, however, a coolant made of 75% water and 25% propylene glycol still circulates through cold plates mounted on the processors. That is meaningful progress, but it is not “no water at the rack.”
The coolant can be filled once and recirculated within a closed loop, but recirculation does not change the fact that it is water-based. Unlike a dielectric working fluid, a water-glycol mixture does not provide electrical insulation. Safe operation therefore requires the coolant to remain contained throughout the cold plates, hoses, connectors, and distribution points. Cold-plate cooling is not inherently unsafe. Its trade-off is that circuit integrity, leak detection, fluid monitoring, and maintenance remain fundamental when operating a water-based loop inside the rack.
Two-phase immersion cooling has a different catch
Two-phase immersion cooling takes another route. Hardware is submerged in a dielectric fluid that vaporises as it absorbs heat. A condenser then removes the heat and returns the vapour to liquid form. This process transfers energy through the latent heat of vaporisation, rather than solely by increasing the liquid’s temperature. The Soben report describes immersion cooling as more energy efficient than cold-plate cooling, while also noting its higher capital costs and limited suitability for retrofit projects.
Here, one of the trade-offs concerns fluid chemistry. The report identifies PFAS-based working fluids among those used in two-phase immersion cooling and highlights the possibility of regulatory action in Europe and the United States. The regulatory paths differ. The European Union is evaluating a broad restriction on PFAS under REACH. At the US federal level, measures currently address particular PFAS or defined groups of PFAS through environmental controls and reporting requirements, rather than imposing a universal ban on the entire category. Phase change does not, by itself, imply a particular fluid chemistry, and a blanket prohibition covering every PFAS substance is not a settled outcome. Even so, fluid composition, long-term availability, regulatory compliance, and end-of-life management should now form part of the procurement assessment.
The gap: two-phase liquid cooling without immersion
Cold plates can generally be integrated more readily into existing infrastructure, but conventional systems typically bring water-based coolant into the rack. Two-phase immersion cooling uses phase change and a dielectric fluid, but requires the hardware to be submerged and may rely on PFAS-based working fluids. This creates space for another architecture: two-phase direct-to-chip cooling based on a mechanically pumped flow loop.
In this architecture, a pump circulates a dielectric working fluid through a cold plate mounted on the processor, where the plate functions as an evaporator. Through flow boiling, part of the fluid vaporises as it absorbs heat. The resulting liquid-vapour mixture then reaches a condenser, releases the heat, and returns to liquid form. The hardware is not immersed, and the working fluid is designed to remain confined within the closed loop. Phase change occurs inside the cold plate, where the heat load is concentrated.
When used with a PFAS-free dielectric working fluid, this architecture combines phase-change heat transfer with no hardware immersion and “no water at the rack” in the rack-side cooling loop, without relying on a PFAS-based coolant. The dielectric fluid reduces the electrical risk associated with a potential leak, but does not make leakage consequence-free. Pumps, seals, connectors, pressure control, and monitoring remain important. Two-phase direct-to-chip cooling does not eliminate engineering trade-offs. It changes them.
The next question for AI infrastructure cooling
As AI chip power and rack density increase, cooling decisions must account for more than heat-removal capacity. Operators must also consider which fluid circulates inside the rack, how extensively a system modifies existing infrastructure, and how the fluid’s chemistry fits within an evolving regulatory landscape.
The question is therefore not simply how far liquid cooling adoption will grow, but whether the benefits of phase-change heat transfer can be brought to processors through cold plates without water in the rack-side cooling loop or reliance on PFAS-based working fluids.
This is the challenge In Quattro is working to address through its research and engineering activities in advanced thermal management for high-power electronics. Learn more about In Quattro and the expertise behind the company.
Frequently Asked Questions
What is direct-to-chip cooling?
Direct-to-chip cooling uses cold plates mounted on high-power components such as CPUs and GPUs. A coolant or working fluid circulates through channels inside the plates, absorbs heat, and carries it away through the cooling loop.
Does direct-to-chip cooling always use water?
No. Direct-to-chip describes the cooling architecture and where heat is collected, not a specific coolant. Conventional single-phase cold-plate systems typically use water-based coolants, often water-glycol mixtures, while two-phase direct-to-chip systems can use dielectric working fluids.
How does a two-phase direct-to-chip cooling system work?
A working fluid absorbs heat inside the cold plate and partially vaporises. The resulting liquid-vapour mixture reaches a condenser, where heat is released and the vapour returns to its liquid state before the fluid is recirculated through the closed loop.
What is the difference between two-phase direct-to-chip cooling and two-phase immersion cooling?
In two-phase immersion cooling, the hardware is submerged in a dielectric fluid. In two-phase direct-to-chip cooling, the hardware is not immersed: the working fluid circulates within a closed loop, and phase change occurs inside cold plates mounted on selected high-power components.
Are all two-phase cooling fluids PFAS-based?
No. “Two-phase” describes a thermodynamic process, not a particular fluid chemistry. PFAS-free dielectric working fluids can also be used in two-phase cooling systems.
What does “no water at the rack” mean?
It means that water does not circulate within the rack-side cooling loop. It does not imply that no water is used elsewhere in the facility.
Editorial sources:
