Guest lecture · Part 2 · §2.8 · Slide 21

Data-centre cooling with RL — how the 40% figure was earned

Google's DeepMind reported a 40% cut in cooling energy at their data centres. The number is real. The tool that delivered it was not reinforcement learning turning valves — it was a learned optimiser choosing setpoints, running advisory-first, with hard safety constraints and operator override. This is what that looks like.