Closed-Loop Cooling Explained: The Plumbing Behind Meta's AI

Cooling AI servers has become a critical challenge as hardware power increases, making traditional air cooling less efficient. Meta’s newest AI data centers in Texas use closed-loop liquid cooling, where a water-glycol mix circulates through server hardware, absorbing heat before passing through heat exchangers to dissipate it. The coolant is then recirculated, reducing ongoing water use and extending its lifespan to up to a decade without replacement.

Meta employs two liquid cooling systems: direct-to-chip closed-loop cooling for AI-optimized facilities and Air-Assisted Liquid Cooling for retrofitting existing buildings. Both systems minimize water consumption, with annual usage comparable to just a few restaurants. The approach also improves space efficiency, allowing more GPUs per rack compared to air-cooled alternatives, which require larger trays and more infrastructure.

The company designs its cooling systems in-house, from custom chips to power and thermal management. In 2025, Meta introduced IcePack, a liquid-cooled network rack platform, and released it openly via the Open Compute Project to promote industry-wide adoption of efficient cooling solutions.

Meta’s engineering teams also use reinforcement learning to optimize cooling operations, testing decisions in a physics-based simulator before deployment. This approach has been scaled to air-cooled data centers, helping reduce energy use while maintaining optimal server temperatures across varying conditions.