In material published on August 27, 2026, Meta explained how most of its new AI-optimized data centers rely on closed-loop liquid cooling, with reinforcement learning used to improve the operation of cooling systems. This approach comes as AI servers generate more heat, making air-only cooling less efficient as hardware capabilities and GPU density increase.
The material was written by Tom Shaw, whom Meta identified as a Software Developer and Content Creator, following a visit to the company’s AI infrastructure in Texas. The material compares this design with an earlier trial at a data center in Altoona, Iowa, where racks containing 16 Nvidia H100 units were air-cooled with limited water use, without sending water directly to the hardware.
How Does Closed-Loop Cooling Work?
A mixture of water and glycol passes through the hardware to draw heat from the server racks, after which the fluid moves to heat exchangers that transfer the heat away from it. Once cooled, the fluid returns to the racks in a continuous cycle instead of being discharged outside the facility. Meta says it expects to use these fluids for up to a decade without needing to replace them.
The method used to dissipate heat varies according to the data center’s location and environment. When liquid-cooling equipment is installed in facilities that lack integrated liquid infrastructure, Meta uses a system called Air-Assisted Liquid Cooling, which includes racks containing pumps and heat exchangers that perform the same function on a smaller and more distributed scale.
Water and Space Gains
According to Meta, a typical AI-focused data center that uses closed-loop liquid cooling with dry coolers can consume less water annually than two full-service restaurants. The company also believes the design is not limited to reducing resource consumption; air-cooling the servers would have required, by its estimate, an area nearly twice the size of the server tray to add the necessary cooling equipment.
Directly cooling chips through a closed liquid circuit makes it possible to place more GPUs in the same rack, which may reduce the number of racks required to accommodate equivalent computing capacity. This means that a facility with a fixed footprint can increase computing density without expanding its space by the same amount.
Open Ecosystem and Reinforcement-Learning Control
Meta says it shares some infrastructure designs with the industry through the Open Compute Project, an open-source hardware and software initiative founded in 2011. In 2025, the company announced the IcePack platform for liquid-cooled networking racks and made its design available for free through the project.
Engineering teams also built a physics-based simulator to test cooling decisions before applying them in live data centers. The system simulates weather, server loads, and cooling-equipment behavior, after which the reinforcement-learning model tests ways to reduce cooling while keeping servers within optimal operating conditions. In an initial trial at one of Meta’s data centers, the approach reduced the energy consumption of air-cooling supply fans by an average of 20% and reduced water consumption by 4% across varying weather conditions.
Why Does This News Matter?
The actual development here is not merely replacing air with liquid, but combining high-density direct cooling, fluid recirculation, and software-based operational optimization. This matters to data-center operators facing power, water, and space constraints, but it does not eliminate the engineering limitations associated with differences between facility locations; Meta itself notes that cooling design depends on the environment and available infrastructure. The energy- and water-reduction results also remain figures provided by the company from a specific trial and do not by themselves prove that they can be generalized to all data centers.