Meta Says A Cooling Pilot Cut Fan Energy 20 Percent
Meta published a newsroom explainer on August 27, 2026 about how it cools the hardware inside its AI data centers, and the piece includes a reinforcement learning result that has not been widely reported.
Transcript
Meta published new cooling numbers today. A reinforcement learning pilot cut fan energy twenty percent.
Meta says AI hardware generates so much heat that air cooling stops being efficient. Liquid coolant carries heat to exchangers, then returns to the racks.
Meta says that coolant is water and glycol, reused for up to a decade. Air alone would need nearly double the server tray.
Meta trained reinforcement learning in a simulator. Meta reports the pilot cut fan energy twenty percent and water four percent, then scaled it.
These are Meta's own figures, published on Meta's newsroom by an outside creator, not independently verified. Meta shares IcePack free through the Open Compute Project.
Inboxsmith builds AI receptionists for small businesses. Please like and subscribe for more news.
Sources
Every claim in this video comes from the top ranking coverage of this topic. The claims and where each one came from:
- The more powerful the AI hardware becomes, the more heat it generates during operation. And at a certain point, simply blowing more air through a server rack stops being a particularly efficient method of cooling.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- A liquid coolant (a mix of water and glycol) is passed through the server hardware to move heat away from the server racks. But instead of that liquid being expelled from the facility, it is pumped through a series of heat exchangers, which are used to dissipate and transfer the heat away from the liquid. Once the liquid has cooled down, it is sent back around to the server racks in a continuous looping process.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- Meta expects to use these coolants for up to a decade without needing to replace them.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- If you were to attempt to cool these same servers with air, you'd likely need nearly double the size of the server tray in order to add in the required air cooling equipment.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- Meta's engineering team has been experimenting with reinforcement learning to help inform the design and operations of their cooling systems. The simulator can model variables such as weather conditions, server load, and the behavior of the cooling equipment.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- In a pilot at one of Meta's data centers, this reinforcement learning-based approach reduced the amount of energy consumed by the air cooling supply fans by an average of 20% while also reducing water usage by 4% across different weather conditions. This reinforcement learning-based approach has since been scaled to the air-cooled data centers in Meta's fleet.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)
- By Tom Shaw, Software Developer and Content Creator. Consistent with Meta's Open Compute Project legacy, these advances are being shared with the industry. In 2025, Meta announced IcePack, a liquid-cooled network rack platform that's being shared openly for free via the Open Compute Project.(Meta's official announcement, Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI, published August 27, 2026)