Cutting data-centre cooling energy with machine-learning control
The problem
Cooling can consume roughly a third of a data centre’s energy. Operators run conservative, static setpoints — safe but wasteful — because the plant is too complex and safety-critical to tune by hand in real time.
The AI approach
A neural-network model trained on thousands of sensor readings (temperatures, pump speeds, loads) learns how the facility behaves and recommends cooling parameters that minimise energy use within safe operating limits. Later iterations moved from recommendation to direct, bounded control.
Evidence it works
DeepMind reported a ~40% reduction in energy used for cooling and a ~15% improvement in overall power-usage effectiveness (PUE).
What “good” looks like
Sustained energy reduction with zero safety incidents; the system stays inside hard safety bounds and hands back to standard control when uncertain.
Feasibility & cost shape
Significant data and instrumentation prerequisites and meaningful integration effort; best suited to large facilities where small percentage gains are financially material.
High value for large operators — but only once the data foundation and safety envelope are in place. We’d start with a shadow-mode evaluation before any closed-loop control.
Based on publicly reported information about the Google / DeepMind work.
This is an industry example included for illustration. It is not a Leia Intelligence project, and no client of ours is implied. Figures are as publicly reported by the original parties.
Sources: DeepMind/Google ("DeepMind AI reduces Google data centre cooling bill by 40%") · Helena (Phaidra profile)