On July 20, 2016, DeepMind and Google announced that machine learning had cut the energy used to cool Google’s data centers by up to 40 percent, which they framed as roughly a 15 percent reduction in overall Power Usage Effectiveness (PUE) overhead after accounting for electrical losses and other non-cooling inefficiencies.
The work, led by Richard Evans and Jim Gao, trained an ensemble of deep neural networks on years of sensor data - temperatures, power, pump speeds, setpoints - to predict the data center’s future PUE, plus additional networks to predict temperature and pressure so the system stayed within safe operating limits. Cooling systems interact in complex, nonlinear ways that hand-tuned engineering rules capture poorly, which is exactly the regime where a learned model can find better operating points.
The result became one of the most cited early examples of AI improving the energy efficiency of physical infrastructure, and it foreshadowed a tension that grew sharper as AI itself became a major driver of data center electricity demand.
Why business readers should care: the same modeling approach that optimizes a data center generalizes to any complex facility with sensors and controllable setpoints, where small percentage efficiency gains compound into large recurring savings.