Data centres: Engineering methods

Air, free cooling, cold plates or tanks of fluid

What this tab covers The engineering methods that can do this conversion, compared on efficiency, scale, economics, limitations and reliability.

There is one basic conversion, electricity to computation to heat, but several ways to build and cool the facility.

Comparison

Method	Efficiency	Scale	Economics	Limitations	Reliability
Enterprise or legacy air-cooled data centre²	Older and smaller sites tend to sit above the 1.54 average PUE	Under a few MW	Low utilisation of servers and cooling raises cost per unit of computing	Cannot cool high-density AI racks	Depends on redundancy level; power problems caused 45% of impactful outages in the 2025 survey
Hyperscale cloud campus with free cooling¹³	Google fleet PUE 1.09 (2025)	Hundreds of MW to over 1 GW per campus; a typical AI data centre uses as much electricity as 100,000 households, the largest under construction 20 times more	Lowest cost per unit through scale and utilisation	Needs very large grid connections and land; evaporative cooling uses water	Redundancy spread across sites as well as within them
AI data centre with direct-to-chip liquid cooling¹¹¹²	GB200 NVL72 rack about 2.7 TFLOPS per watt (FP16 with sparsity, calculated); less fan power than air cooling	Racks of up to about 132 kW	Higher capital cost per MW; much more computing per square metre	Plumbing in the white space, leak risk, new skills	Pumps and quick-disconnect fittings add failure points
Immersion cooling	Removes almost all fan power; not quantified in the sources used	Niche, mostly pilots and crypto-mining	Specialised hardware and fluids	Maintenance is harder; some fluids are fluorinated chemicals under regulatory pressure	Limited long-term data
Colocation²	Close to industry average PUE	From a few MW to hundreds of MW	Customers pay per kW of power reserved	Operator controls the building, tenants control the IT, so efficiency incentives are split	Sold on uptime guarantees

Efficiency gains in the building (PUE falling from 2.5 in 2007 to about 1.5) have largely been taken; the average has hardly moved since 2018, though the best hyperscalers reach 1.1. The bigger lever is now computing per watt inside the server and rack, and liquid cooling allows racks around 15 times denser than the average (calculated). The constraint has moved from efficiency to the sheer size of grid connections, with demand projected to more than double by 2030.

How each method works

Enterprise or legacy air-cooled data centre

Computer room air handlers blow chilled air into a raised floor or aisles; chillers make the cold water. Often small and run by the company itself.

Hyperscale cloud campus with free cooling

Very large, standardised halls use outside air or evaporative cooling for most of the year, high-efficiency power chains and custom servers, with software moving workloads between sites.

AI data centre with direct-to-chip liquid cooling

Cold plates on GPUs and CPUs carry warm water or coolant to coolant distribution units; the remaining heat in the air is removed conventionally.

Immersion cooling

Servers sit in tanks of dielectric (electrically insulating) fluid that carries heat away by circulation or by boiling and condensing.

Colocation

A developer builds and runs the facility and rents space, power and cooling to many customers who install their own servers.

Sources

Energy and AI (World Energy Outlook special report) IEA, 2025
Uptime Institute Global Data Center Survey 2025 Uptime Institute, 2025
Data centers: efficiency (power usage effectiveness) Google, 2026
NVIDIA GB200 NVL72 (specifications) Nvidia, 2025
Vertiv and Nvidia collaborate on 7MW reference architecture for GB200 NVL72 DatacenterDynamics, 2024