Google’s AI-controlled data centre cooling

Cooling is one of the largest sources of energy use in a data centre, and it is hard to optimise. Equipment, weather and computing load interact in complex, non-linear ways, and each data centre is different. In 2016, Google and its AI research company DeepMind set out to see whether machine learning could do better than the engineers’ rules and experience.

This case is valuable less for the headline saving than for how Google handed control to the AI: gradually, and with safety built in.

Stage one: recommendations

DeepMind trained ensembles of deep neural networks on historical data already collected by thousands of sensors in the data centre, including temperatures, power, pump speeds and setpoints. The system predicted future energy efficiency, and further models predicted temperature and pressure over the next hour to check that recommended actions would stay within operating limits.¹

In July 2016, DeepMind reported that the system had achieved a 40% reduction in the energy used for cooling, equivalent to a 15% reduction in overall power usage effectiveness (PUE) overhead once electrical losses and other inefficiencies were taken into account.¹ At this stage, the AI made recommendations, and people decided whether to act on them.

Stage two: autonomous control

In August 2018, DeepMind announced that the AI was now controlling cooling directly. Every five minutes, a cloud-based AI takes a snapshot of the cooling system from thousands of sensors, uses deep neural networks to predict how different combinations of actions will affect future energy use, and identifies the actions that minimise energy use within safety constraints. Those actions are sent back to the data centre, where they are verified by the local control system before being implemented.²

DeepMind reported that the autonomous system was delivering consistent energy savings of around 30% on average.² One Google data centre operator, Dan Fuenffinger, described watching the AI “learn to take advantage of winter conditions and produce colder than normal water that reduces the energy required for cooling within the data centre,” adding: “Rules don’t get better over time, but AI does.”²

Safety by design

The 2018 announcement set out eight safety mechanisms. They included continuous monitoring to stop the AI breaching safety constraints, automatic failover to a safe state, two layers of verification for every action, uncertainty estimation so that only high-confidence actions are taken, and a fall-back to established rules. Above all: “Human override is always available and will supersede any AI actions.”²

Why it worked

The problem suited the technology. Cooling is measurable, data-rich and repetitive, with a clear objective and fast feedback. That is where machine learning is strongest.

Trust was earned in stages. Google ran the system in recommendation mode first, and only moved to autonomous control once operators had seen it perform. The figures also show the trade-off honestly: the autonomous system’s reported average saving was lower than the headline figure from the recommendation phase.

Control stayed with people. Constraints, verification and override were designed in, not added afterwards. The AI was trusted to optimise within limits that humans set and could always enforce.

Lessons for leaders

1. Choose problems where the feedback is fast and the objective is clear. The faster you can see whether an AI decision was right, the faster you can safely extend its authority.
2. Move from advice to action in deliberate stages. Recommendation first, autonomy later, with evidence at each step.
3. Specify the safety case before the savings case. Decide how the system fails safely, who can override it, and how, before you decide how much it should save.

Questions to ask in your organisation

1. Which of our operational processes are measurable, data-rich and repetitive enough to benefit from AI optimisation?
2. For any AI that takes actions rather than giving advice, what are its hard limits, and who can override it?
3. What evidence would we need to see before moving an AI system from recommending to deciding?