Accelerating liquid cooling adoption in data centers: necessity, challenges and solutions

Overview

The paper states that cooling systems in air-cooled data centres consume around 40% of total data-centre energy, and that optimising cooling strategies is crucial to the world's carbon footprint with the explosive growth of AI. Liquid cooling is described as a must-have when GPU/AI accelerator thermal design power exceeds 1,200 W (for example Nvidia GB300 and beyond). After about 4 years of operation, TCO with liquid cooling can be up to 20% lower than with air cooling (a cited vendor figure).

Challenges: for a typical deployment where 80% of rack heat is removed by liquid, initial cost of liquid cooling is about 1.9-2.2 times that of air cooling (citing a Schneider Electric analysis). Some polymers and elastomers absorb corrosion inhibitors from the coolant, lowering inhibitor concentration and accelerating corrosion of wetted materials. Immersion liquids should meet their property specifications under operating conditions for 10 years. Leak risk, lack of standards and a skilled-workforce shortage are also listed.

The proposed fixes are modular component-level liquid-cooled design, better cold plates (additive-manufactured and two-phase), negative-pressure CDUs, AI-driven pump, valve and temperature control, and interoperability programmes such as standardised universal quick disconnects (UQDs). Electrochemical additive-manufactured cold plates showed 21% lower thermal resistance and 201% better thermal uniformity than a straight-microchannel baseline (a cited study). A negative-pressure (vacuum) CDU makes any leak draw air into the loop instead of spilling coolant.

Limits: the paper is a short perspective rather than an experimental study; its numbers are cited from other sources and include vendor figures. The conclusion is that liquid cooling is critical for the rising power consumption and carbon emissions of GPUs and AI accelerators, but cost, leakage risk, standardisation and workforce shortages hinder adoption.

Key points

1. Cooling systems in air-cooled data centres consume around 40% of total data-centre energy.
2. Liquid cooling is described as a must-have when GPU/AI accelerator thermal design power exceeds 1,200 W (for example Nvidia GB300 and beyond).
3. After about 4 years of operation, TCO with liquid cooling can be up to 20% lower than with air cooling (cited vendor figure).
4. For a typical deployment where 80% of rack heat is removed by liquid, the initial cost of liquid cooling is about 1.9-2.2 times that of air cooling (citing a Schneider Electric analysis).
5. Some polymers and elastomers absorb corrosion inhibitors from the coolant, lowering inhibitor concentration and accelerating corrosion of wetted materials.
6. Immersion liquids should meet their property specifications under operating conditions for 10 years.
7. Electrochemical additive-manufactured cold plates showed 21% lower thermal resistance and 201% better thermal uniformity than a straight-microchannel baseline (cited study).
8. A negative-pressure (vacuum) CDU makes any leak draw air into the loop instead of spilling coolant.

Using it in design

• For racks where GPU/accelerator thermal design power exceeds 1,200 W, take liquid cooling as a mandatory design input, and weigh the initial cost (about 1.9-2.2 times air cooling) against TCO (up to 20% lower after 4 years, vendor figure).
• Add a negative-pressure CDU for leak risk and wetted-material compatibility (polymers absorbing corrosion inhibitor) to the design specification.
• Consider standardisation and interoperability (UQD) requirements in supplier selection for quick disconnects and coolant.

Who should read it

Designers and investment decision makers deciding the move from air to liquid cooling for AI racks.

Document contents

1. 1. Necessity of transitioning from air cooling to liquid cooling
2. 2. Challenges of deploying liquid cooling
3. 3. Solutions (3.1 Advance new liquid cooling design, 3.2 Utilize AI-assisted liquid cooling, 3.3 Build data center liquid cooling ecosystem)
4. 4. Conclusion

Related standards

• OCP ACS cold plate requirements
• OCP immersion fluid base specification
• ASHRAE TC 9.9

Read the document

Publication details

Authors

David Zhou; Liguang Du; Ran Xu; Jiqiang Song; Na Chen; Shanshan Zheng; Yuehong Fan

Publisher

Technology Review for Carbon Neutrality (Tsinghua University Press)

Year

2025

Pages

9550014

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.26599/trcn.2025.9550014

Licence

CC BY 4.0 — share with attribution

Licence evidence

doi.org

Citation