AI data center cooling: how Google uses AI to cut energy use

Why data center cooling is hard

The amount of energy used by large data centers has always been a problem for technology companies. Keeping servers at the right temperature is such a challenge that Facebook even built one of its data centers in Luleå, in northern Sweden near the Arctic Circle. Google tried a different solution: it used its DeepMind unit and applied AI to manage the energy consumption of its data centers.

Google has cared about data center performance for as long as it has run them. Such centers need considerable energy for cooling, as well as constant adjustment of temperature, pressure and humidity, to operate as efficiently as possible. Early on, Google decided to design and build its own facilities from scratch, so it could keep piloting new cooling technologies and operating strategies. Its data centers use advanced cooling techniques, intelligent temperature and lighting controls, and power distribution redesigned to minimize energy losses. In 2018 its servers were also built to draw as little power as possible, without unnecessary components such as graphics cards. Today the picture has flipped: AI servers are built around power-hungry accelerators, and they have put cooling back at the top of the agenda.

How much energy does a data center use? The International Energy Agency estimates that data centers consumed about 415 TWh in 2024, around 1.5% of the world’s electricity, and expects that to roughly double to 945 TWh by 2030. Servers take around 60% of a modern facility’s electricity. Cooling ranges from about 7% in efficient hyperscale data centers to over 30% in less efficient enterprise ones. In the United States, data centers used 4.4% of all electricity in 2023 (176 TWh), and the Department of Energy expects 6.7% to 12% by 2028.

How DeepMind’s AI cut cooling energy by 40%

Google looked for further ways to reduce used energy and brought artificial intelligence into the process. It reported that it could cut the energy used for cooling by applying machine learning from DeepMind, the British AI lab it bought in 2014. DeepMind trained an ensemble of deep neural networks on historical data from thousands of sensors in the data center: temperatures, power, pump speeds and setpoints. The main model predicted the future PUE, and two more ensembles predicted temperature and pressure over the next hour, so that recommended actions stayed within operating limits.

In a dynamic environment such as a data center, it is difficult for people to see how all the variables interact: IT load, which rises with demand such as energy-consuming YouTube videos, outside air temperature, chillers and pumps. In DeepMind’s words, the equipment, how it is operated and the environment “interact with each other in complex, nonlinear ways”, and each data center has a unique architecture and environment. Computers are irreplaceable when it comes to this kind of historical data analysis.

The work started earlier. In 2014 Jim Gao, an engineer on Google’s data center team, took the information Google collects during daily operations and ran it through a model to help understand the complex interactions his team struggled with. Google said at the time that his models predicted PUE with 99.6% accuracy.

The results reported in 2016: DeepMind cut the energy used for cooling by up to 40%, which, after electrical losses and other overheads, equals a 15% reduction in overall PUE overhead, and the test site reached the lowest PUE it had ever seen. “The implications are significant for Google’s data centres, given its potential to greatly improve energy efficiency and reduce emissions overall,” wrote DeepMind’s Richard Evans and Google’s Jim Gao. “This will also help other companies who run on Google’s cloud to improve their own energy efficiency.”

From recommendations to autonomous control

Google also declared that it planned to use DeepMind artificial intelligence for other parts of the data center infrastructure, not only for cooling, and DeepMind named power plants, semiconductor manufacturing and industrial production as next candidates. The 2016 system only gave recommendations, which operators had to apply by hand. In August 2018 Google switched to direct control, which works like this:

• Every five minutes, a cloud-based AI takes a snapshot of the cooling system from thousands of sensors.
• Deep neural networks predict how different combinations of actions will change future energy use.
• The AI picks the actions that use the least energy while meeting a set of safety constraints.
• The actions are sent back to the data center, where the local control system checks them again before applying them.
• Operators can leave AI control mode at any time, and control then returns to the usual on-site rules.

After nine months the autonomous system was delivering energy savings of around 30% on average, up from about 12% at launch. In 2022 DeepMind and Google published live tests of the same approach on the cooling systems of two real-world facilities, run with the building systems company Trane, with energy savings of about 9% and 13%. In 2025 DeepMind’s AlphaEvolve found a better scheduling rule for Google’s cluster manager, Borg, which recovers on average 0.7% of Google’s worldwide compute resources: efficiency that comes from the work itself, not the cooling.

How AI data centers are cooled in 2026

AI hardware changed the cooling problem itself, because accelerators pack much more power into each chip. As Richard Whitmore, chief executive of the cooling company Motivair, put it on NVIDIA’s blog: “Once the watts per chip crossed a certain level, liquid cooling became mandatory.” That is the short answer to why AI data centers cannot be cooled with air alone. The main options today:

• Air cooling: fans push cold air through the servers, with hot and cold aisles kept apart. It is simple, but fan energy grows quickly with density, and evaporative cooling towers consume water.
• Direct-to-chip liquid cooling: cold plates sit on the processors and a coolant loop carries the heat away. NVIDIA’s Rubin generation takes coolant, 75% water and 25% propylene glycol, at up to 45°C, warm enough to be cooled by dry coolers outside, with chillers needed only a few days a year in favorable climates.
• Immersion cooling: whole servers sit in a non-conductive (dielectric) liquid, for the highest densities, at the cost of special equipment and handling.

Water is now part of the efficiency debate. NVIDIA puts a cooling tower system at roughly 2.6 million gallons of water per megawatt per year, against near zero for a closed liquid loop with dry coolers. Google, which now publishes its figures, reports that its data centers used 30.8 TWh of electricity in 2024, 27% more than a year earlier, and that its data centers and offices consumed about 8.1 billion gallons of water.

Data center energy efficiency: how it is measured

The standard metric is PUE (power usage effectiveness): the total energy a facility uses divided by the energy used by its IT equipment. A PUE of 1.0 would mean every kilowatt-hour goes into computing. Google reports a fleet-wide average of 1.09 for 2025, against a global average of 1.54 in the Uptime Institute’s 2025 survey, which by Google’s count means 83% less overhead energy. PUE does not show how much useful work the servers do, or how much water the cooling uses, so operators track water and carbon metrics next to it.

The most common ways to reduce energy consumption in a data center:

• Measure first: meter power at the facility, room and rack level, so you know where the overhead is.
• Raise temperature setpoints where the hardware allows. NVIDIA cites industry estimates that one degree warmer at the chiller plant saves about 4% of cooling energy costs.
• Separate hot and cold aisles, and use outside air (free cooling) when the climate allows.
• Move dense AI racks to liquid cooling instead of adding more air handling.
• Let machine learning tune chillers, pumps and fans continuously, within limits set by people.