Direct-to-Chip and Immersion Liquid Cooling: Dramatically Cutting Water Use in AI Data Center

Neil L. Rideout

9/25/20265 min read

Direct-to-Chip and Immersion Liquid Cooling: Dramatically Cutting Water Use in AI Data Centers

Artificial intelligence is transforming every industry, but the physical infrastructure powering large language models, generative AI, and high-performance computing is straining under intense heat loads. Traditional air-cooled data centers rely heavily on evaporative cooling towers that consume vast quantities of water. As AI clusters scale to tens or hundreds of megawatts, water scarcity and sustainability pressures are forcing a rapid shift toward advanced liquid cooling. Two technologies stand out: direct-to-chip (D2C) liquid cooling and immersion cooling. Both deliver dramatic reductions in water use while enabling higher rack densities and improved energy efficiency.

The Water Problem in AI Data Centers

Modern AI training and inference servers pack GPUs and accelerators that dissipate 700 W to more than 1,500 W each. A single rack can exceed 100 kW; some hyperscale deployments push beyond 150 kW per rack. Air cooling struggles at these densities. Cooling towers evaporate water to reject heat, and a large data center can use millions of gallons per day. In water-stressed regions such as the American Southwest, parts of Europe, and many Asian markets, this consumption has become a major constraint on new capacity.

Water usage effectiveness (WUE) measures liters of water consumed per kilowatt-hour of IT energy. Conventional facilities often operate at WUE values of 1.5–2.0 L/kWh or higher when evaporative cooling is primary. Liquid cooling systems can drive WUE close to zero by eliminating or sharply reducing evaporative losses. The result is not only environmental benefit but also lower operating risk and greater siting flexibility.

Direct-to-Chip Liquid Cooling Explained

Direct-to-chip cooling, sometimes called cold-plate or rear-door heat exchanger cooling when combined with facility loops, places a metal cold plate in direct thermal contact with high-power components—primarily CPUs, GPUs, and memory modules. A dielectric or water-glycol coolant circulates through microchannels in the cold plate, absorbing heat at the source. The warmed fluid travels to a coolant distribution unit (CDU) that transfers the heat to a facility water loop or, in some designs, to outdoor dry coolers.

Because heat is captured before it enters the room air, the need for high-volume air movement and evaporative cooling drops dramatically. Many D2C deployments still use a small amount of facility water for the secondary loop, but the primary coolant loop is closed. When paired with dry or adiabatic coolers, water consumption can fall by 80–95 % compared with traditional evaporative systems.

Key advantages of direct-to-chip cooling include:

Compatibility with existing server form factors. Vendors offer cold-plate kits that retrofit standard 1U and 2U servers or integrate into new designs.
High heat-removal capacity. Modern cold plates handle 1,000 W+ per GPU while maintaining safe junction temperatures.
Incremental deployment. Operators can cool only the hottest components and continue using air for lower-power devices, easing the transition.
Proven reliability. Data-center operators have run D2C systems for years in high-performance computing environments; the technology is mature and supported by major OEMs.

Challenges remain. Plumbing complexity increases with manifold design, leak detection, and maintenance procedures. Coolant chemistry must prevent corrosion and biological growth. Facility teams need training for liquid handling. Despite these hurdles, hyperscalers and colocation providers are rapidly adopting D2C for AI clusters because it enables denser racks without the full infrastructure overhaul required by immersion.

Immersion Cooling: Submerging the Server

Immersion cooling takes liquid cooling further by submerging entire servers or server boards in a dielectric fluid. Two main approaches exist:

Single-phase immersion keeps the fluid below its boiling point. Heat transfers by convection to a heat exchanger that rejects it to a secondary loop.
Two-phase immersion allows the fluid to boil at the component surface. Vapor rises, condenses on a cooler surface, and returns as liquid, providing highly efficient heat transfer with minimal pumping power.

Dielectric fluids—engineered hydrocarbons, fluorocarbons, or synthetic esters—are non-conductive and safe for electronics. Servers are modified to remove fans and sometimes air baffles, then lowered into sealed tanks. Heat is extracted through the fluid to external coolers that can operate as dry systems in many climates.

Immersion delivers the highest cooling densities currently available—well over 200 kW per rack in some designs—and can virtually eliminate water use when paired with dry coolers or free-cooling heat rejection. Because the fluid surrounds every component, hot spots are minimized and component temperatures remain more uniform, potentially improving reliability and allowing higher sustained performance.

Additional benefits include:

Dramatic reduction in fan energy. Server fans and computer-room air handlers consume significant power; immersion removes most of that load.
Noise reduction and simpler airflow management.
Potential for higher ambient temperatures and free cooling in a wider range of climates.
Space efficiency: tanks can be densely packed and stacked in some configurations.

Trade-offs exist. Immersion requires purpose-built or heavily modified servers, specialized tanks, and fluid management systems. Fluid cost and long-term availability matter. Maintenance involves lifting servers from the bath, which changes operational procedures. Two-phase systems add complexity around vapor management and fluid purity. Nevertheless, for greenfield AI facilities or dedicated high-density pods, immersion is increasingly competitive.

Quantifying the Water Savings

Independent studies and operator reports consistently show large water reductions. A traditional air-cooled AI cluster relying on cooling towers may consume 1.8–2.5 liters of water per kWh of IT load under design conditions. Direct-to-chip systems with dry or hybrid coolers routinely achieve WUE below 0.2 L/kWh, and pure dry systems approach zero. Immersion installations paired with dry coolers report near-zero water use for cooling.

At the scale of a 100 MW AI campus, the difference is enormous. Reducing average WUE from 1.8 to 0.1 L/kWh saves roughly 1.5 billion liters of water per year—enough to supply tens of thousands of households. In drought-prone regions this translates directly into regulatory approval and community acceptance.

Energy efficiency also improves. Liquid cooling raises the temperature of the heat rejection stream, enabling higher free-cooling hours and more efficient chillers when mechanical cooling is required. Power usage effectiveness (PUE) values of 1.1 or lower are realistic for well-designed liquid-cooled facilities, compared with 1.3–1.5 for many air-cooled sites. Lower PUE reduces both energy cost and the carbon associated with electricity generation.

Implementation Considerations and Best Practices

Successful liquid-cooling deployments share several characteristics:

Early design integration – Cooling architecture must be decided before server procurement and facility layout. Retrofitting large existing halls is possible but more expensive than designing for liquid from the start.
Coolant selection and chemistry management – Whether water-glycol for D2C secondary loops or dielectric fluids for immersion, fluid properties, corrosion inhibitors, and filtration require ongoing attention.
Leak detection and containment – Sensors, drip trays, and automatic isolation valves are standard. Immersion tanks are sealed by design, reducing risk.
Training and operational change – Facilities teams need new skills. Many operators partner with specialized liquid-cooling vendors for initial deployment and knowledge transfer.
Hybrid approaches – Not every watt needs immersion. Many sites use D2C for GPUs and CPUs while retaining air cooling for storage and networking, optimizing cost and complexity.
Heat reuse opportunities – Liquid systems produce higher-temperature heat streams that are more useful for district heating, industrial processes, or absorption chilling, further improving sustainability metrics.

Standards bodies and industry consortia are accelerating adoption. Open Compute Project, ASHRAE, and various liquid-cooling alliances publish guidelines on cold-plate interfaces, fluid specifications, and safety practices. Major server OEMs now offer liquid-ready platforms, reducing integration risk.

The Road Ahead

AI demand continues to grow faster than efficiency gains in silicon. Without a fundamental change in cooling, water and energy constraints will limit the pace of deployment. Direct-to-chip and immersion cooling provide a practical, near-term solution that cuts water use by an order of magnitude while unlocking the denser, more powerful clusters the industry needs.

Over the next five years we can expect:

Wider standardization of cold-plate and immersion interfaces.
Lower fluid costs and improved long-term fluid stability.
Greater integration of liquid cooling into modular data-center designs.
Increased focus on heat recovery as a revenue or sustainability offset.
Regulatory incentives or requirements favoring low-WUE designs in water-stressed regions.

Operators who adopt these technologies early gain both environmental credibility and competitive advantage in power- and water-constrained markets. The transition from air to liquid is no longer experimental; it is becoming the baseline for next-generation AI infrastructure.

Direct-to-chip and immersion liquid cooling are not merely incremental improvements. They represent a structural shift that allows AI data centers to scale responsibly, dramatically reducing water consumption while supporting the computational intensity that modern artificial intelligence demands. As the industry continues its rapid expansion, these cooling methods will be essential to balancing technological progress with environmental stewardship.

Contact

Head Office

Green Life Enterprises LLC
8 Biltmore Est
Suite 310,
Phoenix, AZ 85016

greenlifedatacenters@gmail.com

+1-813-220-0001

Canadian Office

Green Life Enterprises LLC
3142 Nicholson Ave
Suite 10
New Waterford, Nova Scotia B1H 1N8