Data centres have always produced heat, but artificial intelligence is changing the scale and concentration of the problem. Advanced graphics processors perform enormous numbers of calculations inside increasingly dense server racks. Nearly all the electricity consumed by these components eventually becomes heat that must be removed continuously.
The International Energy Agency reported that global data-centre electricity use increased by approximately 17% during 2025, while demand from AI-focused facilities grew even faster. Its broader analysis projects global data-centre electricity consumption could more than double to around 945 terawatt-hours by 2030. As computing capacity expands, cooling is becoming one of the most important constraints on where data centres can be built and how much equipment they can operate.
Traditional cooling systems remain useful, but the industry can no longer assume that moving cold air through a server room will handle every workload. Engineers are now experimenting with direct liquid cooling, immersion tanks, warm-water systems, dry coolers and waste-heat recovery to keep high-density computing reliable without creating unsustainable energy and water demands.
Why Traditional Air Cooling Is Reaching Its Limits
Conventional data centres generally cool equipment by pushing conditioned air through server racks. Fans inside each server move that air across processors, memory and power components before the heated air returns to the cooling system.
The method works effectively when heat loads remain within manageable limits. Hot-aisle and cold-aisle layouts, containment systems, variable-speed fans and better airflow management can all increase efficiency. The U.S. Department of Energy’s data-centre design guide continues to identify optimized airflow and higher equipment operating temperatures as important efficiency measures.
However, air has a limited ability to carry heat compared with liquid. As rack power rises, operators must move larger volumes of air at higher speeds. That requires bigger fans, more space and additional electricity. Uneven airflow can also create hotspots even when the room’s average temperature appears acceptable.
The challenge is especially severe in AI clusters, where large numbers of powerful accelerators operate close together. Cooling systems must remove heat from individual chips before it spreads through the server and into the surrounding room.
Direct-to-Chip Cooling Brings Liquid Closer to the Heat
Direct-to-chip cooling is emerging as one of the most practical answers. Instead of depending entirely on air, specially designed cold plates are attached to processors and other high-temperature components. A liquid flowing through channels inside the plates absorbs heat close to where it is produced.
The heated liquid then travels through rack manifolds and a coolant distribution unit, where the heat is transferred to a facility-level water loop or another rejection system. Fans may still cool components that are not connected to cold plates, but the largest heat sources are managed directly.
This approach can support much higher rack densities while reducing the amount of energy devoted to moving air. It can also make cooling more predictable because heat is captured before it enters the server room.
A life-cycle assessment described by Microsoft found that cold-plate and immersion technologies could reduce total cooling-related energy demand by approximately 15% to 20%, greenhouse-gas emissions by 15% to 21% and water consumption by 31% to 52% compared with air cooling across the assessed data-centre life cycle. The precise result depends on climate, electricity source, equipment design and operating conditions.
Existing Data Centres Need Retrofit-Friendly Solutions
Building a new liquid-cooled facility is easier than converting an established air-cooled site. Existing buildings may not have the pipework, floor loading, leak-detection systems or heat-rejection equipment needed for full liquid cooling.
Retrofit technologies are therefore becoming a major area of innovation. Google’s Brazos cooling system, introduced for wider industry use in 2026, is a rack-mounted closed-loop liquid-to-air system designed to operate liquid-cooled hardware inside an existing air-cooled environment. It captures heat from components through liquid and releases it into the data centre’s hot aisle through a liquid-to-air heat exchanger.
This type of hybrid design could allow operators to modernize selected racks without reconstructing an entire facility. High-density AI equipment could use liquid cooling while conventional servers remain air-cooled.
Rear-door heat exchangers provide another transitional option. A liquid-filled heat exchanger is installed behind a server rack, absorbing heat from the outgoing air before it enters the room. The technology does not capture heat as directly as a cold plate, but it can reduce the burden on room-level cooling systems.
Immersion Cooling Removes Air From the Equation
Immersion cooling takes a more radical approach by placing electronic equipment inside a non-conductive fluid. Because the fluid does not conduct electricity, it can contact servers without causing the short circuits that ordinary water would create.
In single-phase systems, the fluid remains liquid as it circulates through an external heat exchanger. In two-phase systems, it boils at a carefully selected temperature, carries heat upward as vapor and then condenses back into liquid.
Immersion can eliminate many server fans, bulky air channels and conventional heat sinks. Microsoft researchers have also explored three-dimensional server arrangements that take advantage of immersion cooling’s ability to remove heat without traditional airflow paths. This could allow future computing equipment to be packed more densely than standard rack-mounted hardware.
However, immersion cooling introduces its own complications. Servers must be designed or adapted for submerged operation. Maintenance practices change, fluids must remain stable and compatible with components, and operators need systems for filtering, monitoring and containing those fluids.
Two-phase systems also require careful fluid selection because environmental impact, cost and long-term availability can affect whether the technology is viable at large scale.
Warm-Water Cooling Could Reduce Chiller Dependence
Liquid cooling does not necessarily require extremely cold water. Because water transfers heat more effectively than air, some systems can cool processors using water that would feel warm by conventional cooling standards.
Higher coolant temperatures make it easier to release heat through outdoor dry coolers without operating energy-intensive mechanical chillers. They also produce warmer return water, which is more valuable for heat recovery.
The National Renewable Energy Laboratory’s high-performance computing facility uses component-level warm-water cooling in a chiller-less design. Its system captures waste heat that can be reused within the surrounding building rather than immediately rejecting all of it to the atmosphere.
NVIDIA has also described future systems using approximately 45°C liquid loops to capture heat directly from chips. According to the company, the higher operating temperature can expand the number of hours in which outdoor dry coolers can reject heat without mechanical refrigeration.
Water Use Is Changing the Cooling Debate
Evaporative cooling can consume less electricity than compressor-based refrigeration, but it loses water as part of the cooling process. That trade-off has become controversial in water-stressed communities where data centres compete with households, agriculture and industry for limited supplies.
The next generation of designs is attempting to reduce or eliminate routine water evaporation. Microsoft’s zero-water cooling design uses closed-loop chip-level cooling and is expected by the company to avoid more than 125 million litres of cooling water per year at each facility built with the design. The figure concerns cooling consumption and does not mean the entire site operates without any water.
Dry coolers can reject heat directly to outside air without evaporating water, although their performance depends on climate and coolant temperature. Hybrid systems can switch between dry and evaporative operation, using water only when outdoor conditions make it necessary.
Better cooling decisions will therefore require operators to evaluate electricity and water together. A system that saves water but requires significantly more power may shift rather than solve its environmental burden.
Waste Heat Could Become a Useful Product
Data centres do not destroy heat after removing it from processors. They simply transfer it somewhere else. Capturing that heat at a higher temperature creates opportunities to reuse it for offices, greenhouses, industrial processes or district-heating networks.
Google’s Hamina data centre in Finland uses seawater for cooling and is connected with an offsite heat-recovery project. NREL has demonstrated a more localized approach by reusing warm-water computing heat within its Energy Systems Integration Facility.
Heat reuse is not practical everywhere. A suitable customer must be located nearby, and demand may be seasonal. Long pipe networks are expensive, while low-temperature waste heat may require heat pumps before it becomes useful.
Even so, heat recovery changes the objective of cooling. Instead of treating server heat purely as waste, facility designers can view it as an energy stream that should be captured at the highest practical temperature.
The Future Will Use Several Cooling Technologies
There is unlikely to be one universal cooling system for every data centre. Conventional air cooling will remain suitable for lower-density equipment. Rear-door exchangers and rack-level systems will help upgrade existing facilities. Direct-to-chip cooling is likely to become common in high-performance AI infrastructure, while immersion may support specialized equipment and unusually dense designs.
The strongest future systems will combine hardware, software and facility engineering. Sensors will identify hotspots, control systems will adjust pumps and fans in real time, and computing workloads may be shifted according to cooling capacity, electricity availability and outdoor conditions.
The hunt for better cooling is therefore becoming inseparable from the future of computing itself. As AI systems become larger and more powerful, successful data centres will not simply generate more processing capacity. They will remove, reuse and manage heat with far greater precision.
