Written by: Greg Alexander, Senior Thermal Engineer at Motivair by Schneider Electric
Thermal management is a priority as data centers race to add AI capacity. Operators are weighing the pros and cons of single-phase vs. two-phase direct liquid cooling solutions as they seek a quick, effective approach. Single-phase cold plate cooling is the default choice and likely to dominate in the next five to 10 years because it’s easier to deploy, costs less, and aligns with future silicon roadmaps. It has been called the most effective liquid cooling solution and has an estimated 2026 market share of 55%.
Two-phase liquid cooling is more complex, requiring a bigger investment in infrastructure and more time to deploy. It could become more popular in the long term, but for now, operators need to act quickly because demand for AI is exploding. Grand View Research estimates the market will grow at a 30.6% annual rate through 2030.
What is the difference between single and two-phase liquid cooling?
When it comes to direct-to-chip (DTC) options, the choice comes down to single-phase or two-phase DTC cooling. Single-phase solutions use coolant – usually a mix of 75% water and 25% glycol – to remove heat from GPUs and high-density AI chips. Cold plates placed directly over the chip (GPU or CPU) capture heat through convection, transferring it to coolant that circulates in a closed-loop system.
As the name implies, two-phase DTC involves both liquid and vapor. Heat is captured by boiling the coolant – often a dielectric or refrigerant – on the GPU with cold plates similar to those used in single-phase DTC. This allows for heat capture with less temperature rise due to the latent heat of vaporization. This vaporization process ultimately provides cooling to the GPUs in 2-phase DTC.
Is two-phase cooling better than single-phase?
Determining whether single-phase or two-phase is a better approach depends on a data center’s specific requirements. Strictly speaking, two-phase DTC has a higher heat transfer rate at the cold plate, but there is a catch: It involves higher upfront costs and increases the complexity of fluid management.
Currently, two-phase DTC is seldom used, with data center operators opting for the easier, quicker-to-deploy single-phase approach. Single-phase systems are simpler to design, deploy, scale, and integrate with existing infrastructure, giving operators a quicker, more affordable option to cool their fast-expanding AI deployments. Single-phase DTC systems can be rack-mounted to cool only the equipment on the rack or deployed in between rows to handle groups of racks.
When should you use two-phase direct-to-chip cooling?
Two-phase DTC could be preferable in a couple of scenarios:
- The industry flow standard for liquid cooling is 1.5 liters per minute for each kilowatt. Conceivably, chip wattage may eventually become so high that pumping enough coolant per minute for heat extraction won’t be feasible. Such a scenario would favor a two-phase DTC to reduce pumping power.
- Single-phase cold plate cooling allows for temperature variations, with fluid entering and leaving cold plates at different temperatures. With a two-phase process, the fluid boils at a constant temperature, which could be desirable if constant GPU surface temperature is critical to operation.
Neither scenario is occurring as of yet; each would require much denser chips than those in use today. Even the newly announced NVIDIA Vera Rubin can be cooled with single-phase DTC at 45 °C at this point. In the future, processors may reach the densities that necessitate 2-phase DTC cooling.
Balancing speed, cost, and complexity in AI cooling strategies
As AI infrastructure scales at an unprecedented pace, the choice between single-phase and two-phase direct-to-chip cooling ultimately comes down to timing, complexity, and business priorities. While two-phase technologies hold clear long-term promise for ultra-high-density environments, today’s reality demands solutions that can be deployed quickly, cost-effectively, and at scale.
That’s why single-phase DTC continues to lead by offering a practical path to support rapid AI growth without overengineering for future requirements that have yet to fully materialize. For operators, the smartest strategy isn’t choosing one approach over the other indefinitely, but aligning cooling decisions with current workload demands while maintaining flexibility for what’s next. Schneider Electric, with Motivair, delivers a full suite of single-phase DTC solutions that can help data center operators accelerate their AI deployments and achieve a quicker time-to-revenue.
Add a comment