Definition: Liquid cooling removes heat from computer chips with a flowing liquid instead of moving air. In AI data centers the most common form is direct-to-chip cooling, where a metal cold plate sits on each processor and coolant circulates through it and out to a heat exchanger.
TL;DR: Air stops working as a coolant when a rack draws far more power than an office server room ever did. A flagship AI rack draws about 120 kW, so operators pipe liquid to the chip. Liquid cooling can cut energy and water use compared with air, according to a 2025 Nature study, but adoption is still partial: in Uptime Institute's 2025 cooling survey, 22% of respondents used direct liquid cooling. Build an app free →
Why AI Needs It
Every watt a chip draws leaves as heat. Better than 99% of the electricity a data center uses ends up as heat, so power density and heat density are the same number.
A general-purpose rack draws roughly 10 kW. A rack-scale AI system draws roughly 120 kW. Air can carry only so much heat per volume, so at this density the fans and airflow needed become impractical. Water carries roughly 3,500 times more heat than the same volume of air, so a modest pipe replaces a wind tunnel.
The chips themselves also sit close together. Stacked HBM sits beside the processor, and rack-scale systems link 72 accelerators through one NVLink fabric. That packing concentrates heat, and liquid cooling removes it.
How It Works
In direct-to-chip cooling, the coolant never touches the chip. It touches a plate that touches the chip.
The Four Main Approaches
| Approach | How it works | Trade-off |
|---|---|---|
| Air cooling | Fans push air across heat sinks | Simple and familiar, limited by density |
| Direct-to-chip (cold plate) | Liquid flows through plates on the chips | Works in existing halls; some air cooling still needed for other parts |
| Single-phase immersion | Servers sit in a bath of dielectric fluid | Simpler tanks and cheaper fluids; less chip-level performance |
| Two-phase immersion | Fluid boils at the chip and condenses | Very high density; complex tanks and fluorinated fluids under regulatory scrutiny |
In the cold plate designs the coolant is often a water and glycol mix. In immersion the fluid is a dielectric, meaning it does not conduct electricity.
What the Research Says
In May 2025, Microsoft researchers published a cradle-to-grave life-cycle assessment in Nature comparing air cooling, cold plates, and single-phase and two-phase immersion. Against air cooling, the liquid options cut greenhouse gas emissions by 15% to 21%, energy demand by 15% to 20%, and blue water consumption by 31% to 52%. Cold plates performed about as well as the two immersion methods.
Read that with its limits. The study focused on general-purpose server chips, not AI accelerators, and it compared cooling methods, not whole facilities.
Liquid Cooling and Water Use
The common worry is that AI uses too much water. The answer depends on how the heat leaves the building:
- Evaporative cooling towers reject heat by evaporating water. They are efficient in energy but use water.
- Closed-loop liquid cooling with dry coolers reject heat to outdoor air without evaporation. Berkeley Lab's data center center of expertise notes that a system running on high-temperature coolant can use dry coolers, which removes evaporative water use almost anywhere in the world.
Microsoft describes a datacenter design, used from August 2024, that recirculates water in a closed loop through direct-to-chip cooling. Microsoft says the design avoids more than 125 million liters of water per year per datacenter, and that water is still used for restrooms and kitchens. The trade-off is that mechanical cooling raises power use, which Microsoft calls "nominal."
NVIDIA makes a similar argument for its own systems and claims 300 times better water efficiency than air-cooled designs. That figure is NVIDIA's own marketing claim, so treat it as a claim rather than a measurement.
Worked Example: A 100 MW Site
The data center page shows what PUE does on a 100 MW connection. At a PUE of 1.6, about 62.5 MW reaches compute. At 1.1, about 91 MW reaches compute. Liquid cooling is one of the main ways operators move from the first number toward the second, because it cuts the power spent on fans and chillers. That is why the choice is financial before it is thermal.
Adoption
Uptime Institute's 2025 cooling survey lists perimeter air cooling at 75% of respondents and direct liquid cooling at 22%. Uptime's 2025 global survey adds that average rack densities are still rising slowly, with few facilities above 30 kW. The number of liquid-cooled AI sites is growing, but the installed base of the world's data centers is still mostly air-cooled.
Common Mistakes
- Treating liquid cooling as one technology. Cold plate and immersion differ in cost, risk, and maintenance.
- Assuming liquid cooling means zero water use. Closed loops avoid evaporation, but the site still needs a way to reject heat.
- Quoting vendor multiples as measurements. Water and energy claims depend on the baseline, and most baselines are older air-cooled designs.
- Forgetting the supply chain. Pipes, manifolds, heat exchangers, and trained technicians are constraints too.
Connection to Taskade
Cooling sits in the provider's building, not in your workspace. When you use AI agents or automations on Taskade, model calls run on secure infrastructure that you never size or cool. The link to you is cost: cooling is a large share of what it takes to serve a token.
What You Would Build in Taskade
A "sustainability log" project: an automation records each week's AI-heavy workflows, and an agent summarizes which ones could run less often. Describe yours and build it free →
Related Concepts
- Data Center - the building that hosts the racks
- GPU - the chip that creates the heat
- HBM - the memory that sits beside it
- NVLink and Interconnects - why racks are dense
- TPU - Google says its eighth generation uses fourth-generation liquid cooling
- Inference Cost - where energy shows up on a bill
- Tokens Per Second
Frequently Asked Questions About Liquid Cooling
What is liquid cooling in a data center?
It is a method that removes heat from chips with a liquid instead of air. The most common form for AI is direct-to-chip cooling, where a cold plate on each processor carries heat into a coolant loop.
Why do AI data centers need liquid cooling?
Because AI racks draw far more power than traditional racks, roughly 120 kW against 10 kW, and air cannot remove that much heat practically. Liquid carries far more heat per volume.
What is the difference between direct-to-chip and immersion cooling?
Direct-to-chip runs liquid through plates on the chips while the rest of the server stays in air. Immersion submerges whole servers in a dielectric fluid. Immersion can reach higher densities but needs different tanks and maintenance.
Does liquid cooling use less water than air cooling?
Often, yes, but it depends on how the heat is rejected. The 2025 Nature study found blue water use falls by 31% to 52% versus air cooling. A closed loop with dry coolers avoids evaporation, while a cooling tower does not.
Is liquid cooling more energy efficient?
The same study found energy demand falls by 15% to 20% versus air cooling for the cases it modeled. Real savings depend on the facility, climate, and how well the system runs.
How common is liquid cooling?
It is growing but not yet the majority. Uptime Institute's 2025 cooling survey found 22% of respondents using direct liquid cooling and 75% using perimeter air cooling.
Do AI chips run at a lower temperature with liquid cooling?
Liquid can hold chips at steadier temperatures and support higher power per chip. Google says its Ironwood TPU, which is liquid cooled, can sustain up to twice the performance of standard air cooling under continuous heavy workloads.