BIGGER PICTURE Energy & climate
Why running more cooling equipment can sometimes use less energy
Data-centre cooling is a system of interacting pumps, towers and temperatures. Google's experiment searched for efficient combinations that fixed operating rules can miss.

- 2016 reported result
- Up to 40% less cooling energy
- Later control interval
- Every 5 minutes
- Boundary of the claim
- Cooling, not total computing energy
How the idea works
Read the plant's operating state
Sensors describe temperatures, pressures and equipment settings.
Compare possible settings
Models estimate energy and operating consequences.
Apply verified changes
Local controls check proposed actions before execution.
The difficult part is how the equipment interacts
Servers release heat that pumps, chillers and cooling towers must remove. Changing one setting affects other equipment, and outside weather changes the options again. Google's earlier engineering work used operational sensor histories to model overall facility efficiency. This made the plant's behaviour something that could be estimated from data, instead of relying exclusively on a collection of fixed rules.
Sources: [3]
A real cooling result, with a narrow denominator
In 2016, DeepMind reported reducing cooling energy by up to 40% in a live Google data-centre test. The team predicted future efficiency and checked temperature and pressure constraints before recommending actions. Forty percent referred to cooling, not the servers' electricity or the total environmental footprint of digital services. The distinction is essential when judging the size of the benefit.
Sources: [1]
A counter-intuitive operating choice
A 2018 follow-up described operators learning to spread cooling loads across more equipment rather than less. The system could evaluate combinations instead of assuming fewer running machines always meant lower consumption. It also progressed from recommendations to direct control. This is the appealing engineering detail: efficiency can depend on where machines operate, not merely on whether they are switched on.
Sources: [2]
Automation came with a local veto
The later controller proposed actions every five minutes and checked them against constraints; a local controller checked again, and operators could leave AI mode. Those safeguards are part of the mechanism, not an optional accessory. Both savings reports came from the developers. They demonstrate a documented operating approach, while leaving independent replication and whole-facility sustainability as separate questions.