Direct liquid cooling (DLC) has attracted significant investment over the past two years, driven largely by the massive buildout of dense AI training clusters. In the coming years, major AI hardware manufacturers are planning ever-higher computing density, with a corresponding sharp increase in rack power — soon to surpass 200 kW.

Not long ago, the typical question about DLC for data center planning was whether it was needed at all. By 2026, the question for many organizations has shifted to how much DLC capacity they will need over time — AI compute systems have made it clear to many operators they needed to prepare for DLC. Generic servers are also approaching thermal power densities where a change to liquids make sense. As such, expectations of DLC adoption are running high, as highlighted by a recent flurry of acquisitions and investments.

While water cold plates continue to dominate current adoption, the industry has also turned some of its attention to alternative types of cold plates, as well as various immersion systems. Two-phase cold plates, in particular, have seen significant investment activity in recent months as a candidate to supersede water as a coolant.

Placing bets on two phases

IT vendors have thus far favored water cold plates (with 25% propylene-glycol mixes in particular, known as PG25) because of their long, proven track record and industry backing. These systems are closed-loop fluid networks with broad component availability, well-understood material compatibility and established maintenance needs. This preference also aligns with facility operators' inclination toward incremental, non-disruptive changes to data hall layouts as well as the existing technical knowledge of facility staff. Operating procedures for water cold plate systems are familiar to operators versed in facility chilled water systems and water-cooled mainframes.

Investment activity over the past year reflects the current, overwhelming preference for water-based DLC systems (see Investments signal a heated liquid cooling race).

Still, two-phase cold plate systems, particularly those from Accelsius and ZutaCore, have garnered renewed interest from strategic investors such as Johnson Controls and Legrand for the former and Carrier for the latter. The logic behind these partnerships is about strategic positioning, enabling all parties involved to offer complete thermal solutions across both technology cooling and facility loops, from cold plates to heat transport to heat rejection. Furthermore, the water cold plate segment is already saturated, while the two-phase option adds differentiation and relevance.

It also opens up new opportunities in the IT supply chain, primarily with server makers. Server vendors remain crucial to DLC adoption, serving as a channel for cold plates and coolant distribution units (CDUs) manufacturers to integrate and support their DLC equipment in IT systems.

Broader support from IT vendors, along with multiple sources for CDUs, will be critical for commercial success. For system integration, two-phase cold plates are fundamentally similar to those using water in their form factor and placement on the silicon packages (e.g., CPUs and GPUs). However, beyond the cold plate, the technology cooling system starts to diverge from their water-based equivalent — from the size of tubes/hoses and pipes to the manifold and CDU designs.

A key difference is the presence of vapor. In a two-phase system, part (Accelsius, flow boiling) or all (ZutaCore, pool boiling) of the cooling capacity is provided in the form of latent heat, as coolant vaporizes while absorbing heat (hence the cold plate is also called an evaporator). This can be advantageous in cooling silicon hotspots — small areas of extreme heat flux — by providing greater temperature stability and uniformity. Importantly, there is minimal risk of damage from leakage, a major concern shared by many operators.

In the Uptime Institute Cooling Systems Survey 2025, a greater portion of data center operators reported considering two-phase systems for future deployments (see Figure 1 and Operators warming up to dielectric cold plates). Many expect that two-phase cold plates will become more attractive as densification of IT continues and deployments grow in scale.

Figure 1 Operators considering DLC favor water and dielectric cold plates

Reaching boiling point

There are several major considerations that could enable two-phase coolants to challenge the dominance of water-based systems:

• Lower flow rates and pumping energy. A major feature of two-phase cooling is its comparatively lower flow-rate requirement, despite the base liquid having a lower specific heat capacity than water. Nucleation in the liquid at the heat source helps address heat flux hotspots without having to increase liquid replenishment across the entire cold plate, as is typically the case with water coolants. This difference is expected to become more important as silicon thermal design power (TDP) continues to rise alongside any concomitant temperature restrictions (Tcase) as GPUs surpass 2 kW per module and CPUs approach 1 kW per socket. 
  While cold plate systems are expected to keep pace in absolute thermal performance with thermal engineering advances to keep flows under 1.5 liters per kilowatt per minute (at 10°C/18°F temperature rise), total system flow rates will continue to increase as installations grow. Next-generation water-cooled rack systems will each need hundreds of liters of water per minute, requiring larger pipes and hoses, as well as kilowatts of pumping power to overcome pressure in the cold plate fluid network. At the facility scale, two-phase cooling could, therefore, offer significant infrastructure benefits.
• Coolant quality and maintenance. Another key consideration with water cold plate systems is monitoring and maintaining coolant quality. Technology cooling water fluid networks require fine particle filtration, additives to prevent corrosion and biogrowth, and regular contamination checks. Mistakes or material quality issues can be expensive to remedy and have a severe impact on business, particularly as DLC fluid networks grow to multiple megawatts covering dozens of racks. 
  Water leakage at scale is also a concern for all positive-pressure water systems (commercialization of negative-pressure systems, such those from Chilldyne, is still limited), calling for extensive leak detection systems to minimize potentially terminal damage to IT hardware components (often costing several thousands) and to avoid hours of downtime. In contrast, two-phase (and any dielectric-based) systems offer much reduced maintenance needs and overall risk, easing the pressure on facility staff.
• More free cooling opportunities. Nvidia chief executive Jensen Huang inadvertently created some shockwaves at the start of 2026 when he discussed the technical feasibility of using 45°C (113°F) supply water coolant, which in many climates would eliminate the need for compressors (using only dry or evaporative/adiabatic coolers). However, there are numerous trade-offs that explain why most operators avoid this approach: it can create a capacity silo for Nvidia-based IT systems and associated capacity planning issues; it may degrade overall computational speeds and accelerate component wear due to higher temperatures; and it places even more pressure on facility infrastructure to maintain temperature stability during both standard operating conditions and grid power interruptions.
  Two-phase systems can help ease such trade-offs by the nature of physics. The high heat capacity of liquid-to-gas phase change keeps the overall temperature rise between supply and return low. Phase change (nucleate boiling) also provides higher temperature stability and uniformity across the silicon package. This means higher supply temperatures (calibrated just below the two-phase coolant's boiling point, e.g., 50°C/122°F) do not need to come at the expense of hotter silicon areas or increased thermal-mechanical wear. Temperature changes and differences across components in the overall stack between the silicon and the cold plate (e.g., thermal interfaces, package and heat spreader, cold plate base) are the primary source of material stress. 
  On the facility side, adopters of two-phase systems will evaluate — and in some cases will soon deploy — designs without facility water systems, instead pumping the refrigerant directly from the IT chassis to external condensers, similar to direct-expansion (DX) systems. This can reduce infrastructure overhead for dedicated heat rejection via high condensation temperatures and low head pressure in the overall system. Examples of small installations exist already.

Outlook

Longer term (beyond five years), future innovations may lead to tighter integration and co-optimization of two-phase cooling with IT silicon and system design, as dielectric properties make it possible to bring the liquid even closer to microelectronics in the package. This will open up possibilities for higher-performance processor circuit designs with extreme heat fluxes several times greater than supported by current chips, without sacrificing free cooling capabilities.

At the same time, challenges persist on the road ahead to wider adoption of two-phase cold plate systems. First is supply chain diversity: there are only a handful of two-phase developers, and even fewer are technically mature and commercially ready (Boyd Thermal may be the third on the list after ZutaCore and Accelsius), and suppliers are incompatible with each other as they use specialized CDUs.

Second, the use of engineered refrigerants can be contentious for supply, cost and sustainability considerations. Importantly, many are either classified as forever chemicals (Per- and polyfluoroalkyl substances, or PFAS), or their breakdown product stabilizes into such compounds. In some markets, particularly in Europe and Japan, this can be a deal-breaker for some buyers, although ongoing chemical research may resolve these concerns in the future.

Third, water cold plates work well for now, and the industry has largely converged on PG25 coolants. Also, water systems can expand their cooling coverage to virtually all IT components — as demonstrated in supercomputers and next-generation Nvidia rack-scale systems — including memory banks, solid-state drives, network interface chips and power electronics such as voltage converters and regulators. For two-phase closed-loop systems, achieving similar coverage is a much less straightforward step to take, both technically and economically.