AI Data Center Liquid Cooling: From Air Cooling to Chip-Level Liquid Cooling
Adapted for U.S. and Canadian data center operators, infrastructure planners, and enterprise decision-makers.
Artificial intelligence and high-performance computing are changing compute density in data centers and redefining the design boundaries of thermal infrastructure. As GPUs, CPUs, and other accelerators continue to improve performance, power consumption per server keeps rising, and the concentrated deployment of high-power devices in racks further increases localized heat loads.
Historically, data centers relied primarily on airflow to remove heat from servers, then used room-level cooling systems for overall heat rejection. This approach is mature and well-supported, and it remains highly applicable to traditional enterprise IT, general-purpose computing, and low-to-medium-density server environments. However, for sustained high-load scenarios such as AI training, inference, and high-performance computing, relying solely on air to move heat is facing increasingly clear efficiency and density limits.
As a result, data center cooling is shifting from “lowering room ambient temperature” to “directly controlling chip heat.” Chip-level direct liquid cooling (Direct-to-Chip, DTC), represented by cold plates, shortens the heat transfer path and places cooling capacity directly near major heat sources such as CPUs and GPUs, making it an important direction for high-density computing infrastructure.
Why High-Density Computing Changes Cooling Logic
The cooling problem in a data center is fundamentally about how heat—converted from compute power—can be continuously and reliably removed.
Traditional servers have relatively distributed power consumption, and the heat load inside a rack can usually be managed through proper hot/cold aisle design, server fans, and CRAC or CRAH systems. AI servers, however, often integrate large numbers of high-performance GPUs or other accelerators. Compute resources are more concentrated, and the heat generated per rack increases significantly.
This creates three direct changes.
First, heat is more concentrated. Traditional servers can rely on large volumes of air for cooling, while the heat from high-performance GPUs is concentrated in a limited chip area, significantly increasing heat flux density.
Second, rack power continues to rise. As more accelerators are deployed in the same rack, the total heat generated by the servers increases accordingly. Even if the overall room temperature remains within a reasonable range, localized thermal management inside the rack can still become a limiting factor.
Third, the energy consumption of the cooling system itself is receiving more attention. If airflow, fan speeds, and cooling equipment capacity must constantly increase to keep up with IT load growth, the cooling system’s share of total data center energy consumption may also rise.
Therefore, cooling for high-density AI infrastructure is no longer just about “how to lower temperature.” It is about how to reliably deploy more compute capacity per rack and per unit of floor area under limited power, space, and infrastructure conditions.
Air Cooling Still Matters, but Its Boundaries Are Changing
Air cooling has not lost its value because of liquid cooling. On the contrary, in many data center environments, air cooling remains the most mature and easiest-to-maintain thermal solution.
A typical air-cooling heat transfer path is:
Chip → heat sink → server internal air → rack airflow → hot/cold aisle → CRAH/CRAC → refrigeration system
The advantage of this architecture is its high degree of standardization and the mature support ecosystem between servers and facilities. For lower-power-density equipment, air cooling can still provide stable heat removal.
However, as rack power continues to rise, the physical properties of air as a heat transfer medium become an important constraint.
Air has low density and a limited ability to carry heat per unit volume. To remove more heat, greater air volume is required. Increasing air volume means higher fan power, more complex airflow management, and greater room-level cooling capacity.
When high-power devices are concentrated in a limited space, the cooling system may even need to address localized hot spots, rack inlet temperatures, rack airflow short-circuiting, and cooling capacity distribution at the same time.
Therefore, the future data center will not simply move from the “air cooling era” into the “liquid cooling era.” It is more likely to develop a long-term coexistence of air and liquid cooling, with tiered configurations based on thermal load.
Chip-Level Liquid Cooling Changes the Heat Transfer Path
The core change in direct-to-chip liquid cooling is not simply replacing air with liquid. It is a redesign of the heat transfer path from the chip to the facility side.
In a typical DTC liquid cooling architecture, coolant enters a cold plate inside the server through piping. The cold plate forms a highly efficient heat transfer interface with high-power chips such as CPUs and GPUs. After absorbing heat from the chip, the liquid enters the loop and transfers heat through a rack-level or facility-level coolant distribution unit (CDU) to a larger heat rejection system.
The basic path can be summarized as:
Chip → cold plate → coolant → CDU → facility-side heat exchange system → outdoor heat rejection
Compared with air cooling, the greatest structural advantage of liquid cooling is that it moves the heat capture point from the room air layer to the vicinity of the chip.
This means a large amount of heat no longer needs to be released first into the air around the server and then carried away by air. Instead, it enters the liquid loop directly near the heat source.
For GPUs and CPUs with high heat flux density, this architecture is better aligned with the direction of high-density computing.
Why Liquid Is Better Suited to High-Heat-Density Chips
One important reason liquid cooling can support high-density computing comes from the differences in thermal properties between air and water-based coolants.
At near-room temperature, air density is about 1.2 kg/m³, and its specific heat capacity is about 1.0 kJ/kg·K, giving a volumetric heat capacity of about 1.2 kJ/m³·K.
Water has a density of about 1,000 kg/m³ and a specific heat capacity of about 4.18 kJ/kg·K, corresponding to a volumetric heat capacity of about 4.18 MJ/m³·K.
This means that, under the same temperature rise, water can carry far more heat per unit volume than air.
| Parameter | Air | Water |
|---|---|---|
| Density | ~1.2 kg/m³ | ~1,000 kg/m³ |
| Specific heat capacity | ~1.0 kJ/kg·K | ~4.18 kJ/kg·K |
| Volumetric heat capacity | ~1.2 kJ/m³·K | ~4.18 MJ/m³·K |
| Heat-carrying capacity per unit volume | Lower | Significantly higher |
Practical liquid cooling systems do not necessarily use pure water directly. Depending on system design, temperature range, material compatibility, and water quality management requirements, they may use appropriate water-based coolants or other working fluids.
From an engineering perspective, the value of liquid cooling is not just that “liquid carries heat better than air.” More importantly, it can rapidly move heat away from high-heat-flux areas with relatively low flow rates, reducing dependence on large-scale air movement.
How Liquid Cooling Reduces Cooling System Strain
Liquid cooling first reduces dependence on large volumes of airflow inside servers and around racks.
In high-density air-cooled systems, server fans must push large amounts of air through heat sinks. As chip power increases, fan power and airflow requirements also increase.
DTC liquid cooling instead transfers the primary heat directly to the coolant. In servers using a hybrid cooling architecture, high-power components such as GPUs and CPUs use cold-plate liquid cooling, while memory, hard drives, power supply modules, and other lower-heat-load components can still use air cooling.
This approach does not eliminate air entirely. Rather, it assigns components with different heat densities to cooling methods better suited to them.
Therefore, future high-density AI servers are more likely to use hybrid cooling rather than liquid-cooling every component.
Higher Rack Density as a Key Value of Liquid Cooling
For AI and high-performance computing, the ultimate value of liquid cooling comes down to “how much compute capacity can be deployed per unit of space.”
If rack power growth continues to be addressed by increasing airflow, it may require expanding cooling equipment capacity, optimizing hot/cold aisles, increasing air volume, or even retrofitting room infrastructure.
Liquid cooling can separate a large portion of the heat load from the air system, allowing the room air system to handle mainly the remaining equipment and ambient heat load.
This provides more deployment room for high-power racks.
In other words, liquid cooling does not simply improve “heat removal efficiency.” More importantly, it changes the upper limit of data center power density and the way space is utilized.
For AI clusters that need to deploy large numbers of GPU accelerators, there is a close relationship among rack count, room area, and power capacity. If cooling capacity cannot scale simultaneously, it will be difficult to further increase deployment density even if power and compute resources are ready.
As a result, cooling capacity is gradually becoming an infrastructure planning metric as important as power and networking.
Liquid Cooling Can Also Reduce Some Cooling Energy Consumption
Liquid cooling does not automatically mean lower cooling energy consumption. Its actual effect depends on system architecture, coolant temperature, pump power, CDU efficiency, and facility-side heat rejection methods.
From a system structure perspective, however, DTC liquid cooling can reduce the energy required for server fans and large-scale air movement.
At the same time, if the facility uses a higher-temperature liquid loop and combines it with free cooling or dry coolers, some mechanical refrigeration demand can also be reduced.
Under suitable climate and facility conditions, this design may also reduce water consumption from evaporative cooling.
It is important to note that “liquid cooling saves energy” cannot be simplified to mean that all liquid-cooled data centers are necessarily more energy-efficient than air-cooled systems. Liquid cooling systems add pumps, CDUs, piping, heat exchangers, and other equipment, all of which consume energy.
Therefore, evaluating liquid cooling performance should focus on the entire cooling chain, not simply compare cold plates or server fans in isolation.
From Single-Rack Upgrades to Complete Liquid Cooling Infrastructure
Deploying high-density liquid cooling also means that data center infrastructure must change accordingly.
Traditional air-cooled systems focus mainly on hot/cold aisles, supply air volume, return air temperature, and CRAC/CRAH capacity. Liquid cooling systems must further consider:
- Coolant temperature and flow rate
- CDU capacity and redundancy design
- Rack manifolds and quick disconnects
- Piping layout and maintenance space
- Coolant quality and filtration
- Leak detection
- Thermal interface between cold plate and chip
- Facility-side heat rejection capacity
- Pumping power
- Interface standards between IT equipment and liquid cooling infrastructure
This means liquid cooling is no longer just a problem for server vendors. It is a systems engineering challenge involving IT equipment, racks, power distribution, piping, CDUs, cooling equipment, and data center building infrastructure.
Liquid Cooling Deployment Requires Attention to Reliability and Maintenance
Liquid cooling can improve heat removal in high-density computing environments, but it also increases system complexity.
First is liquid management. Coolant must maintain appropriate chemical properties and cleanliness to avoid corrosion, deposition, and material compatibility issues.
Second is piping reliability. Connectors, hoses, manifolds, and quick disconnects must all provide long-term stable operation and minimize leakage risk.
Third, maintenance methods change. Air-cooled server maintenance mainly involves fans, heat sinks, and air filters, while liquid-cooled servers also require attention to cold plates, piping, CDUs, and coolant loops.
Therefore, the operations and maintenance system for a liquid-cooled data center must also be upgraded accordingly, including liquid monitoring, pressure monitoring, flow monitoring, temperature monitoring, and leak detection.
Air Cooling and Liquid Cooling Will Coexist for the Long Term
From the perspective of actual data center construction, the future will not be one technology completely replacing another.
For ordinary enterprise servers, network equipment, storage equipment, and lower-power-density IT devices, air cooling still offers strong economics and maturity.
For high-power GPU servers, high-performance computing nodes, and large-scale AI clusters running continuously, the advantages of liquid cooling are more obvious.
Therefore, a more reasonable direction is to build a tiered cooling system based on the thermal load of different devices.
This hybrid model can avoid fully liquid-cooling an entire data center just for a small number of high-power devices, while also reserving upgrade space for future high-density computing equipment.
Cooling Architecture Will Become a Core Metric in AI Data Center Design
In the past, data center design typically revolved around rack space, power, networking, and storage capacity. The development of AI infrastructure is elevating cooling capacity to an equally important position.
Future data center construction will need to determine the relationship among rack power density, cooling method, power capacity, and network architecture before server procurement.
For new AI data centers, liquid cooling piping, CDU space, rack interfaces, and facility-side heat rejection capacity can be designed directly from the planning stage.
For traditional data centers already in operation, a more suitable approach is incremental liquid cooling retrofits based on the deployment location of high-power racks and actual thermal loads, rather than changing the entire room’s cooling system at once.
This approach can reduce retrofit complexity and better align with the path of gradually adding AI compute capacity to existing data centers.
Conclusion: Cooling Is Shifting from a Supporting System to Compute Infrastructure
The development of artificial intelligence and high-performance computing is fundamentally pushing data centers to evolve simultaneously toward “high compute density,” “high power,” and “high heat density.”
In this process, the cooling system is no longer just a supporting facility that maintains room ambient temperature. It is gradually becoming critical infrastructure that determines compute density, rack design, energy efficiency, and scalability.
Air cooling remains suitable for many traditional computing scenarios. But as the heat flux density of core chips such as GPUs and CPUs continues to rise, simply increasing airflow will face increasingly obvious space and energy constraints.
Chip-level direct liquid cooling shortens the heat transfer path and places cooling capacity directly near major heat sources, providing a new thermal path for high-density AI and high-performance computing. The data center of the future is more likely to form a multi-layered architecture in which air cooling, chip-level liquid cooling, and other liquid cooling technologies coexist.
What truly determines the choice of cooling technology is not whether one technology can completely replace another, but whether compute density, rack power, space, power, climate conditions, and facility architecture can be matched as a whole. As AI infrastructure continues to move toward higher density, cooling capacity will join compute, power, and networking as a foundational element of data center scalability.