Why Air Cooling Is Still More Popular Than Liquid Cooling (and Why That Is Not Stupid)
Read enough AI infrastructure news and you would think air cooling died years ago. Liquid is more efficient. Liquid handles the hottest chips. Singapore even has a national standard for it now.
And yet most data centres in Singapore still cool their servers with air. That is not ignorance or inertia. For most workloads, air is still the rational choice, and it is worth understanding why before anyone talks you into plumbing.
Air vs liquid at a glance
| Compare | Air cooling | Liquid cooling |
|---|---|---|
| 🏢 Works in the building you already have | 🟢 Almost always |
🔴 Rarely without new pipework |
| 🧰 Skills needed | 🟢 Every IT and facilities team has them | 🟡 Specialist, still scarce |
| 🔧 Swapping a server | 🟢 Unplug cables, slide out | 🟡 Also disconnect coolant lines |
| 💦 Leak risk | 🟢 None | 🟡 Managed, but real |
| ⚡ Energy efficiency | 🟡 Baseline |
🟢 Can cut data-centre energy use by more than 30% |
| 🔥 Highest rack density | 🔴 Limited |
🟢 130 kW and beyond |
| 🔊 Noise | 🔴 Loud at high density | 🟢 Much quieter |
Sources: Enterprise Singapore on SS 726:2026 for the energy figure, HPE GB300 NVL72 for rack density. Bars are illustrative.
Seven reasons air is still on top
1. The buildings are already built
A data centre is a twenty-year asset. Almost every hall operating today was designed around chilled air: raised floors, air handlers, hot and cold aisles. Legal commentary on Singapore's new liquid cooling standard notes that most data centres here currently use air cooling. Changing that means pipes, pumps and heat exchangers in a live facility. It is possible. It is not quick.
Source: Rajah & Tann Asia
2. Most workloads simply do not need it
The headline AI racks draw 130 kW or more. The overwhelming majority of business computing does nothing of the sort. File servers, databases, virtual machines, and even a good share of AI inference sit comfortably inside what air can remove. Liquid solves a density problem that most racks do not have.
3. Air has no failure mode that ruins the hardware
When a fan fails, a server runs hot and throttles or shuts down. When a coolant fitting fails, liquid can reach electronics. The industry has good defences, as we cover in What Happens When a Liquid-Cooled AI Rack Leaks? But "very unlikely" is still a different conversation with a risk committee than "cannot happen".
4. Everyone knows how to run it
Air cooling needs no coolant chemistry, no leak runbook and no specialist contractor. Singapore's standard was written partly because the industry's experience with data-centre liquid cooling is still limited. Skills follow deployments, and deployments are only now arriving in volume.
5. The upfront bill is smaller
Liquid cooling adds equipment that air does not need: coolant distribution units, manifolds, piping, sensors and commissioning. The energy savings are real, but they accrue to whoever pays the power bill over years. In colocation, that is often the operator, while the tenant pays for the liquid-ready hardware. Split incentives slow adoption.
6. Vendors keep making air-cooled options, because customers keep buying them
If air were finished, manufacturers would stop building for it. They have not. HPE lists the Compute XD690 with eight air-cooled Blackwell Ultra GPUs, and its mainstream AI inference servers such as the ProLiant Compute DL380a Gen12 and DL384 Gen12 run on air. That is a commercial judgement about where demand is.
Source: HPE AI Servers
7. Flexibility
An air-cooled server can go into nearly any rack in nearly any data centre in the world. A liquid-cooled one can only go where the plumbing is. For organisations that move workloads between sites, or rent space from different operators, that portability has value.
Where air stops working
None of this means air wins everywhere. Its weakness is physics. Air carries heat poorly compared with liquid, so as power per rack rises, the volume of air needed becomes impractical. Industry guides put the comfortable limit for conventional air cooling in the region of 20 to 25 kW per rack, with careful engineering stretching it further.
| Rack density | Cooling that fits | Typical contents |
|---|---|---|
| 🟢 Low |
Air | General servers, storage, networking |
| 🟡 Medium |
Air Well engineered; rear-door units help | GPU inference servers, one or two 8-GPU servers |
| 🟠 High |
Liquid Recommended | Racks full of 8-GPU servers |
| 🔴 Extreme |
Liquid Only | NVL72-class racks at 130 kW and above |
Source for the air-cooling range: AMPCOM data centre cooling guide. Bands are indicative; the exact limits depend on the hall.
The fair criticism of air
Popular is not the same as best. Air cooling uses more energy for the same heat, which matters in a country where data-centre power is tightly rationed. It is noisy. And at the top of the market its time is running out: the next generation of flagship AI servers appears to have no air-only option, as we discuss in The Last Air-Cooled Flagship?
So the accurate picture is a split market. Air remains the sensible default for most of what a data centre does. Liquid is becoming mandatory for a small, fast-growing, very power-hungry slice.
Which side of the line are you on?
| If you are planning… | Likely answer |
|---|---|
| 💬 An internal AI assistant or RAG system | Air |
| 👁️ Vision, speech or analytics inference | Air |
| 🔧 One or two 8-GPU servers for fine-tuning | Air If the hall has the power and airflow |
| 🏋️ A dense cluster of 8-GPU servers | Liquid Worth costing |
| 🗄️ A rack-scale system such as an NVL72 | Liquid No alternative |
The practical advice is unfashionable: do not adopt liquid cooling because it is modern, and do not avoid it because it is unfamiliar. Work out your rack density first. The cooling choice usually follows from that single number.




