How AI took control of cooling systems and reduced their energy use

In August 2018, Google described a concrete change in responsibility inside its data centres. An AI system that had recommended cooling adjustments was now carrying them out. Operators no longer had to implement each suggestion: every five minutes, the system assessed sensor readings, selected actions and sent them to local controls for checking and execution.[1]

That change explains an important form of AI autonomy. The work continues because the system has a standing objective and a continuing stream of measurements. Its next decision is triggered by the control cycle, with no fresh conversation required.

The whole cooling system matters

Cooling equipment operates as an interconnected system. Google’s earlier 2016 account described neural networks trained on temperatures, pump speeds, power readings and other operating data. Additional models forecast temperature and pressure so proposed efficiency improvements could be assessed against equipment limits. The purpose was to learn interactions that were difficult to capture with a fixed collection of operating rules.[4]

A machine using less electricity is not always enough: an adjustment can shift work elsewhere. The commercial-building study published in 2022 gives a useful example. Making a cooling tower work harder can deliver colder water to a chiller, reducing the electricity the chiller needs. Its BCOOLER agent learned to balance those demands as conditions changed.[2]

Testing a controller while a building stays open

BCOOLER was tested at a university site and a building containing shops, restaurants, apartments and a clinic. Three-month comparisons alternated AI and conventional control on successive days. The analysis removed two hours after each handover and compared periods with similar weather and cooling demand. This helped separate controller performance from changes in the job it had to do.[2]

The operators first enabled AI control through the building-management system. They could withdraw it, and automated checks could also return control to the existing rules.[2]

Learning from a production data centre’s operating history

A separate ICLR 2025 paper described experiments conducted during 2024 at one commercial data centre in China. The controller learned from historical operating records, then used current measurements to adjust fan speeds and cold-water valves. Its model represented connections between air-cooling units and temperature sensors, helping it account for how an adjustment affected different parts of the room.[3]

Across more than 2,000 hours of experiments, the researchers reported no temperature-safety violations. They expanded control from selected units to all units in the two rooms. The energy comparison used periods with similar server loads. That is evidence of functioning control at the tested facility; it does not establish the same savings across every data-centre design.[3]

What remains with the operator

Google’s 2018 design kept two rounds of checks: the AI assessed proposed actions against operating constraints, and local systems checked the commands again. Operators retained the ability to exit AI mode and hand control back to established rules. Uncertain actions could be rejected before reaching equipment.[1]

These examples put the benefit of autonomy in a concrete place. A learned controller repeatedly coordinates equipment choices while the responsible team sets the operating limits and watches the results. Evaluating it requires both sides of that arrangement: how much energy the controlled equipment saved, and how reliably the required cooling was delivered. A convincing percentage becomes useful when it is attached to a clearly defined system, baseline and operating record.

Report an error or suggest a correction