For most of the last two decades, data centre cooling meant moving cold air. Raised floors, containment aisles and CRAH units were refined year after year, and for racks drawing a few kilowatts each, that worked well.
AI and high-performance computing have changed maths. GPU racks now draw several times the power of a typical enterprise rack, and air simply cannot carry that much heat away from such a small footprint without enormous fan power and very cold supply air. That is why liquid cooling for data centres has moved from niche to mainstream planning, especially in Singapore, where operators also face some of the tightest efficiency expectations in the region.
This article explains where air cooling reaches its limits, the main types of liquid cooling, and what our engineers look at when a facility is deciding whether, and how, to make the move.
Why Air Cooling Hits a Limit
Air is a poor carrier of heat compared with water. To remove more heat with air, you either move more of it or make it colder, and both cost energy. As rack density rises, three problems appear together:
- Fan power climbs steeply. Moving enough air through a dense rack takes far more fan energy, which shows up directly in your PUE (power usage effectiveness, the ratio of total facility power to IT power).
- Hot spots become harder to control. Even with containment, high-density racks create local recirculation that can starve the top of a rack or the end of a row of cold air.
- Chilled water temperatures have to drop. Colder supply air means colder chilled water, which reduces chiller efficiency and the number of hours you can use free or economised cooling.
There is no single cut-off point, but as a working rule, many operators start finding air cooling difficult somewhere in the range of 30 to 40 kW per rack. Current AI training racks are widely reported to exceed 100 kW. At that level, liquid cooling is no longer an efficiency upgrade. It is the only practical option.
The Singapore Context
Singapore adds two pressures that make the decision more urgent.
First, the climate. High temperatures and humidity all year round limit the free cooling hours that data centres in temperate countries rely on. Every efficiency gain has to come from the cooling design itself.
Second, policy. IMDA’s Green Data Centre Roadmap sets out to bring all data centres in Singapore to a PUE of 1.3 or better at 100% IT load over the next ten years. The roadmap also committed to introducing standards for IT equipment energy efficiency and liquid cooling. New capacity awarded under the second Data Centre Call for Application (DC-CFA2) has been held to a stricter PUE requirement still. For operators planning AI capacity, liquid cooling is increasingly part of how those targets get met. (Our PUE optimisation page explains how we model a facility against these targets.)
The Main Types of Liquid Cooling
Rear-Door Heat Exchangers
A water-cooled coil replaces the rear door of the rack, so hot exhaust air is cooled as it leaves the rack. The servers themselves are unchanged and still air-cooled.
Best for: retrofitting existing halls with moderately dense racks, without changing IT hardware. It extends the life of an air-cooled hall rather than replacing it.
Direct-to-Chip (Cold Plate) Cooling
Liquid flows through cold plates mounted directly on the processors and GPUs, carrying away most of the heat at the source. A coolant distribution unit (CDU) connects the rack loop to the facility water system. The remaining heat from memory, power supplies and other components is still removed by air.
Best for: high-density AI and HPC racks. This is the approach most new GPU platforms are designed around.
Immersion Cooling
Servers are submerged in a non-conductive fluid that absorbs heat from every component. In single-phase systems the fluid stays liquid and is pumped through a heat exchanger. In two-phase systems it boils off the components and condenses again.
Best for: very high densities and specialised deployments. It removes nearly all server fans, but it changes how hardware is serviced and needs compatible equipment.
| Rear-door heat exchanger | Direct-to-chip | Immersion | |
|---|---|---|---|
| Changes to IT hardware | None | Cold plates on chips | Hardware rated for immersion |
| Heat still removed by air | All of it, at the rack door | A minority share | Almost none |
| Typical use | Retrofit, moderate density | New AI and HPC deployments | Highest densities, specialised sites |
| Main design challenge | Water at the rack, coil capacity | CDU sizing, residual air load | Serviceability, fluid handling |
Why Most Facilities End Up With a Hybrid Hall
In practice, very few halls go fully liquid overnight. A typical scenario is a hall that keeps its existing air-cooled racks while adding rows of direct-to-chip GPU racks. That hybrid setup is where many projects run into trouble, because:
- Direct-to-chip racks still reject a share of their heat to air, and that share is easy to underestimate when the rack’s total power is very high.
- The air system was sized for the old layout. New high-density rows can disturb airflow to neighbouring air-cooled racks.
- The facility water loop now serves two very different loads, often at different supply temperatures.
This is where simulation earns its keep. Using data centre CFD (computational fluid dynamics, a way of modelling airflow and heat in 3D), we model the residual air load from liquid-cooled racks alongside the existing air-cooled ones, and check that every rack still receives air at the right temperature. You can see an example in our hybrid cooling CFD case study.
Questions to Answer Before You Commit
- What will the rack densities actually be, and when? Plan for the hardware roadmap, not just the first deployment.
- How much heat will still go to air? This decides whether your existing CRAH units and containment can cope.
- What water temperatures can the system run at? Warmer water for liquid-cooled loops can improve chiller efficiency and increase economiser hours, but only if the rest of the plant is designed for it.
- How does the change affect annual PUE, not just peak? A design that looks good at full load can underperform at partial load. Hourly, whole-facility modelling shows the real annual picture.
- What happens if a CDU or pump fails? Liquid-cooled racks heat up much faster than air-cooled ones when cooling stops, so redundancy and response times need to be modelled.
How Megagenix Helps (Liquid Cooling For Data Centres)
We are an independent engineering consultancy. We don’t sell or install cooling equipment, so our recommendations are vendor-agnostic. For liquid and hybrid cooling projects, our engineers:
- model the hall with CFD to check air-side performance around new liquid-cooled racks;
- size liquid and hybrid cooling systems against your planned rack densities;
- run whole-facility energy simulation to show the effect on annual PUE;
- test failure scenarios before they happen in a live hall.
If you’re planning AI capacity or a hybrid retrofit, speak to our engineers before the cooling design is fixed. It’s far cheaper to change a model than a live hall.
Frequently Asked Questions
At what rack density do you need liquid cooling?
There is no fixed threshold, but many operators find air cooling becomes difficult and inefficient somewhere around 30 to 40 kW per rack. Current AI training racks can go well beyond 100 kW, where liquid cooling is effectively required.
What is the difference between direct-to-chip and immersion cooling?
Direct-to-chip cooling pumps liquid through cold plates attached to the main processors, while the rest of the server is still air-cooled. Immersion cooling submerges the whole server in a non-conductive fluid that absorbs heat from every component.
Can an existing air-cooled data centre add liquid cooling?
Yes. Rear-door heat exchangers can be added with no change to servers, and direct-to-chip racks can be introduced alongside air-cooled ones. The air system and facility water loop need to be checked for the new mix of loads, which is best done with CFD and energy modelling before installation.
Does liquid cooling improve PUE?
It usually does, because it reduces fan energy and can allow warmer water temperatures that improve chiller efficiency. The actual improvement depends on the whole plant design, so it should be modelled on an annual, hourly basis rather than assumed.
