- Type
- Data centre thermal management
- Main variants
- Direct-to-chip, immersion, rear-door heat exchanger
- Heat captured
- Roughly 70–90% of server heat in direct-to-chip systems
- Adopted for
- High-density AI training and inference racks
- Emerged
- Mainstream from 2024–2025
- Related
- AI data centres, GPU cluster, high-bandwidth memory
- Type
- Data centre thermal management
- Main variants
- Direct-to-chip, immersion, rear-door heat exchanger
- Heat captured
- Roughly 70–90% of server heat in direct-to-chip systems
- Adopted for
- High-density AI training and inference racks
- Emerged
- Mainstream from 2024–2025
- Related
- AI data centres, GPU cluster, high-bandwidth memory
Overview
A conventional data centre cools IT equipment by pushing chilled air through raised floors and hot-aisle/cold-aisle containment. That works for racks drawing five to ten kilowatts, but modern AI racks draw far more, and air — a poor heat conductor — cannot keep up without enormous fans and chilled-water plants.
Liquid cooling comes in three broad forms. Direct-to-chip (D2C) cooling mounts a metal cold plate directly on the processor, GPU or memory and pumps coolant through it; a coolant distribution unit (CDU) transfers the heat to the facility water loop while isolating and regulating the IT loop. Immersion cooling submerges the entire server in a non-conductive dielectric fluid that absorbs heat from every surface at once. Rear-door heat exchangers (RDHx) fit a cooling coil into the rack door as a transitional option.[2]
In practice most AI deployments are hybrid: liquid captures the 70–90% of heat generated by the hottest components — CPUs, GPUs and memory — while air still handles power supplies and networking gear.[2]
Key Benefits
Liquid cooling offers several advantages over air. It removes heat at the source, allowing rack densities that would otherwise require dozens of racks spread across a hall. It raises energy efficiency, because pumps move heat with less power than server fans and chilled-air plants, lowering power usage effectiveness (PUE). It cuts acoustic noise and improves hardware reliability by holding component temperatures steady, and in many designs it enables heat reuse, where captured heat feeds district heating or industrial processes.[1][3]
The engineering costs are real: facilities need CDUs, leak detection, quick-disconnect fittings for servicing, filtered coolant loops and operators trained in fluid handling. Coolant quality and circulation reliability are safety-critical — a pump failure stalls a training run as surely as a power cut.[3]
Applications
Liquid cooling is deployed for AI training clusters, where thousands of accelerators run at full load for weeks, for high-density inference racks serving large language models, and for high-performance computing in weather modelling, drug discovery and engineering simulation. Cloud providers including CoreWeave build dense GPU clusters with liquid cooling as the baseline, and colocation operators now design new halls for direct-to-chip readiness even when the first tenant still uses air.[4]
>See Also
🇲🇾 Malaysia's tropical climate raises the cooling load of every data centre in the country: warm, humid ambient air gives air-cooled systems less headroom, which makes liquid cooling comparatively more attractive than it would be in a temperate market. At the same time, water availability is a genuine constraint — Malaysia's data-centre pipeline in Johor and Kuala Lumpur has intensified scrutiny of water consumption, and water-based or evaporative strategies compete with households and agriculture for supply.[5]
The country's Energy Efficiency Management Act 2024 brings large electricity consumers such as data centres under mandatory energy management obligations, which favours technologies that measurably reduce power draw. Local operators and multinational builders therefore specify direct-to-chip cooling for high-density AI halls while retaining air cooling for general colocation, and several operators trialled or adopted immersion systems for ultra-dense deployments. For Malaysian engineers, this creates demand for skills in coolant loop design, CDU operations and thermal monitoring — an area MDEC's digital skills programmes and local technical universities are beginning to address, and a concrete career track alongside the country's growing AI infrastructure workforce.
References
- ↑STL Partners. Data center liquid cooling in 2025: Evidence, economics and the road to mainstream adoption. https://stlpartners.com/articles/data-centres/data-center-liquid-cooling/
- ↑Data Center Knowledge. (2026). Liquid Cooling Options: RDHx, Direct-to-Chip, Immersion. https://www.datacenterknowledge.com/cooling/liquid-cooling-options-rdhx-direct-to-chip-immersion
- ↑IEEE Spectrum. Data Center Liquid Cooling: The AI Heat Solution. https://spectrum.ieee.org/data-center-liquid-cooling
- ↑Macquarie Data Centres. Liquid Cooling for AI Infrastructure. https://www.macquariedatacentres.com/solutions/liquid-cooling/
- ↑Data Center Knowledge. (2026). APAC Data Center Pipeline Surges as AI Pre-Leasing Tightens Supply. https://www.datacenterknowledge.com/data-center-construction/apac-data-center-pipeline-surges-as-ai-pre-leasing-tightens-supply