Air or liquid: which is the better way to cool a data centre? For most facilities today, the answer is neither on its own. Air cooling is cheap, mature and still runs the majority of the world's data centres, but it cannot keep up with AI racks. Liquid cooling removes far more heat, yet it costs more and is harder to retrofit. That is why hybrid cooling, with liquid taking the hottest components and air handling the rest, has become the mainstream design for new high-density data centres. This guide compares air vs liquid cooling and explains how to choose.
Why Data Centre Cooling Matters More Than Ever
Electricity is one of the largest operating costs in any data centre, and cooling is one of its biggest consumers. Industry analysts estimate that cooling can account for up to 40% of a data centre's electricity use.[1]
AI is raising the stakes. A single NVIDIA B200 GPU draws around 1,000 W, up from about 700 W for the previous-generation H100,[2] and AI racks are expected to exceed 150 kW.[1] Cooling is no longer a background facility system. It now decides how much compute a data centre can actually host.
Air Cooling: Mature, Low-Cost, but Reaching Its Limits
Air cooling moves heat away from IT equipment by circulating air. Fans in the servers pull cool air through the equipment, and computer room air conditioning (CRAC) or air handling units remove the heat from the room.
It remains the default choice for good reasons:
- Low upfront cost: fans and air conditioners are mass-produced and widely available.
- Mature technology: designs, suppliers and maintenance skills are well established.
- Fast to deploy: no coolant loops, manifolds or leak detection to install.
The problem is density. Air has a low heat capacity, so moving more heat means moving much more air. In practice, air cooling becomes difficult beyond roughly 20–40 kW per rack.[3] Mixing high-density servers with low-density equipment in the same hall also creates hot spots and uneven heat distribution.
Rack Pools and Rack Containers
To get more out of air cooling, the industry developed the rack pool (also called an integrated micro data centre). It combines cabinets, UPS, power distribution, cooling, monitoring and cabling into one system, and separates hot and cold airflows so that cooling air is not wasted.
The rack container takes this one step further. Cabinets, power distribution, UPS, cooling, monitoring, fire protection and structured cabling are built into a single container in the factory. A traditional construction project becomes a prefabricated, plug-and-play unit that can be deployed quickly and replicated as demand grows.
Liquid Cooling: Built for High-Density and AI Workloads
Liquids carry far more heat than air, so liquid cooling removes heat faster and with less energy. It allows much higher rack densities, gives more precise temperature control, and can lower Power Usage Effectiveness (PUE) significantly. Air-cooled facilities typically run at a PUE of around 1.4–1.8, while facilities using liquid cooling commonly reach about 1.05–1.2.[2][3]
Indirect and Cold Plate Liquid Cooling
In indirect liquid cooling, a coolant picks up heat through a heat exchanger and carries it to external cooling equipment such as cooling towers or chillers. The liquid never touches the electronic components, which keeps the risk to the equipment low.
Cold plate (direct-to-chip) liquid cooling places a metal plate with internal coolant channels directly on the hottest components, usually CPUs and GPUs. It is the most widely adopted form of liquid cooling because it combines high efficiency with a familiar server layout. For high-power GPUs such as the NVIDIA GB200, cold plate cooling is essential: air alone cannot remove the heat these chips produce under sustained AI workloads.
Immersion Liquid Cooling
Immersion cooling submerges entire servers or compute nodes in a dielectric (non-conductive) liquid. The fluid touches every component directly, so it removes heat very efficiently and also keeps dust and airborne contaminants away from the hardware. Immersion delivers the best thermal performance, but its higher cost, heavier tanks and specialised maintenance mean it is still used mainly in extreme-density or specialist deployments. Every component in the tank, including optical modules designed for immersion cooling, must be compatible with the coolant.
Beyond cold plates and immersion, liquid cooling can also be distributed across a facility through purpose-built heat exchange networks, designed around the equipment layout and needs of each data centre.
Air vs Liquid Cooling Compared
The table below compares the two approaches on the factors that matter most when planning a data centre.
| Factor | Air cooling | Liquid cooling |
|---|---|---|
| Heat removal | Moderate; limited by air's low heat capacity | High; heat is removed at the chip or by full immersion |
| Practical rack density | Up to roughly 20–40 kW per rack | 100 kW+ per rack |
| Typical PUE | Around 1.4–1.8 | Around 1.05–1.2 |
| Upfront cost | Lower | Higher (coolant loops, CDUs, manifolds, leak detection) |
| Deployment | Quick; mature, widely available | More complex, especially as a retrofit |
| Maintenance skills | Standard facility skills | Specialist skills for coolant and fluid handling |
| Best for | Low and medium-density racks, network and storage equipment | AI training, HPC and high-power GPU servers |
Figures are typical industry ranges and vary with climate, design and equipment.[2][3]
Hybrid Cooling: Today's Mainstream Approach
Liquid cooling is growing fast, but it has not replaced air. According to an Uptime Institute survey, around one in five data centres (19%) had deployed liquid cooling, and another 36% planned to within the next 12 to 24 months.[2] Even inside liquid-cooled racks, air still has a job to do.
Hybrid cooling combines both. Cold plates are fitted to the components that produce the most heat, such as CPUs, GPUs and sometimes memory, while fans and room air handle everything else: power supplies, storage, network switches and other lower-power parts. In a direct-to-chip AI rack, liquid typically removes around 70–80% of the heat, and air takes the rest.[1] Across a whole hall that also contains conventional racks, the liquid share is lower, because those racks remain fully air-cooled.
Typical heat split in a direct-to-chip AI rack.[1] The exact share varies by equipment.[4]
Hybrid has become the mainstream design for several practical reasons:
- Targeted investment: liquid is used only where the heat density demands it, instead of rebuilding the entire facility.
- Lower cost: air handles the remaining load more cheaply than extending liquid to every component.
- Easier retrofits: existing air-cooled halls can add liquid-cooled rows step by step.
- Manageable skills gap: operations teams can build liquid cooling experience gradually.
Rear-door heat exchangers add another layer. Fitted to the back of a cabinet, they use liquid to cool hot exhaust air before it re-enters the room. This reduces the load on room air conditioning and can support higher room temperatures, moving the facility closer to a fully liquid-cooled design over time.
How to Choose the Right Cooling Approach
The right choice depends on equipment type, rack density, operating environment and budget. As a starting point:
| Scenario | Recommended approach | Why |
|---|---|---|
| Enterprise IT, network and storage rooms | Air cooling with hot/cold aisle containment | Low density, lowest cost, simplest to operate |
| Edge sites and fast deployments | Rack pools or prefabricated rack containers | Factory-built, plug-and-play, easy to replicate |
| Mixed halls adding AI or GPU servers | Hybrid: cold plate liquid cooling plus air, with rear-door heat exchangers where needed | Liquid where heat is concentrated, air for everything else |
| Dedicated AI training clusters | Direct-to-chip liquid cooling as the primary system | Only liquid can remove the heat from high-power GPUs |
| Extreme density or harsh environments | Immersion cooling | Highest thermal performance and protection from dust |
What Hybrid Cooling Means for Racks, Power and Cabling
Cooling choices shape the physical infrastructure around them. Liquid-cooled racks need space for manifolds and coolant lines, so cable routing has to be planned to keep power and fibre pathways clear of fluid connections. Higher rack densities also call for higher-capacity power distribution and more fibre per rack to connect GPU clusters.
- Racks and cabinets: deep, well-organised data centre racks leave room for cooling hardware and clean cable management.
- Power distribution: high-density rack PDUs support the increased load of AI servers.
- High-speed connectivity: optical transceivers and pre-terminated fibre keep dense GPU interconnects fast to deploy and easy to maintain.
Frequently Asked Questions
Is liquid cooling better than air cooling?
For high-density AI and HPC racks, yes: liquid removes far more heat and lowers PUE. For low and medium-density equipment, air cooling is still cheaper and simpler. Most new high-density facilities therefore use both.
What is hybrid cooling in a data centre?
Hybrid cooling uses liquid cooling, usually cold plates, for the hottest components such as CPUs and GPUs, and air cooling for the rest of the equipment. It is the most common way to add liquid cooling today.
Is cold plate or immersion cooling more common?
Cold plate (direct-to-chip) cooling is far more widely used. Immersion offers higher thermal performance but costs more and requires specialist handling, so it is mainly used in extreme-density deployments.
References
- Verdantix: Hybrid data centre cooling emerges as the near-term solution for AI factories
- mGrid: Data centre PUE gains stall as liquid cooling adoption hits 19% (citing the Uptime Institute global survey)
- Introl: Liquid cooling vs air cooling for AI data centers
- Chatsworth Products: How to build a hybrid cooling strategy with air and liquid cooling