Meta details closed-loop liquid cooling for AI data centers, highlighting lower water use and tighter GPU packing

Meta is devoting more engineering effort to the “plumbing” behind its AI hardware as server heat loads rise, according to a new explanation published by the company’s Newsroom today. In the piece, Tom Shaw, who visited Meta’s AI infrastructure in Texas, argues that traditional air cooling increasingly reaches its limits with newer AI-optimized designs, making closed-loop liquid cooling a central part of how Meta keeps GPU servers within operating temperatures.

The post contrasts two approaches. In conventional data centers, air cooling is used to regulate temperatures for compute equipment that supports everyday tasks such as searching for creators on Instagram or liking posts on Facebook. Meta says air cooling can still work for some earlier-generation AI hardware, pointing to an example in Altoona, Iowa where racks of 16 Nvidia H100 GPUs were cooled using air with minimal water use, and no water was sent directly to the hardware.

Meta’s argument is that the cooling problem has changed more recently. The company says a common misconception is that AI data centers are automatically heavy water users. Its position, as described in the post, depends on cooling design, and it characterizes its newest AI-optimized facilities as using closed-loop systems that recirculate water in a sealed loop and require only very little ongoing make-up water.

At the core is what Meta calls closed-loop liquid cooling. The company says a liquid coolant, described as a mixture of water and glycol, is passed through the server hardware to carry heat away from GPU racks. Instead of expelling warmed liquid outside the facility, Meta says the coolant is pumped through heat exchangers that transfer heat away from the fluid. After cooling, the same coolant circulates back to the servers in a continuous loop.

Meta also says it expects to use its coolant for up to a decade without needing replacement. In cases where liquid-cooled equipment must be placed into facilities that do not already have built-in liquid cooling infrastructure, the company describes an “Air-Assisted Liquid Cooling” approach. Under that design, it says racks include pumps and heat exchangers that function similarly to a larger building system, while still relying on a closed-loop cooling concept at a smaller, more distributed scale.

The company links its cooling choice to both resource and space efficiency. Meta says liquid cooling is resource efficient because a typical AI-optimized data center using closed-loop liquid cooling with dry coolers uses less water annually than “a couple of full-service restaurants,” framing the point by comparison to real-world usage. It also says liquid cooling can use rack space more effectively, because attempting to cool the same servers with air would require nearly double the server tray area to accommodate the necessary air cooling equipment, creating diminishing returns as air-cooled setups scale.

On capacity and scaling, Meta states that direct-to-chip closed-loop liquid cooling allows more GPUs to fit in a given server rack. The company says that reduces the number of racks required, so a facility with the same physical footprint can scale compute capacity without needing proportionally more space for cooling hardware. Meta also positions these designs as part of a broader effort across its infrastructure stack, saying it designs and develops its own systems for items including chips, power infrastructure, and cooling systems.

Beyond mechanical design, Meta says it is applying reinforcement learning, a machine learning technique that tries to improve decisions through feedback, to optimize cooling operations. The post says the company experimented with reinforcement learning in a way that avoided trial-and-error on live data centers, instead building a physics-based simulator that can model weather conditions, server load, and how cooling equipment behaves. The simulator, Meta says, provides a safer environment for learning how to reduce cooling requirements while keeping servers within optimal operating conditions.

Meta then describes the approach as moving from experiment to deployment. In a pilot at one of its data centers, it says the reinforcement-learning method reduced energy consumed by air cooling supply fans by an average of 20% while also reducing water usage by 4% across different weather conditions. Meta frames these results as meaningful, and says applying similar percentage reductions across a broader data center fleet could translate into substantial efficiency gains.

The post does not provide additional specifics that readers might expect for a full technical assessment, such as the exact coolant flow rates, how long before any service refresh is needed in practice, which specific AI server generations are covered, or what proportion of Meta’s broader compute footprint is already using closed-loop cooling. It also does not disclose whether the reinforcement learning is now controlling fan speeds in real time across all air-cooled sites or whether its current scope is limited to particular regions and equipment layouts. What is clear is the company is making efficiency improvements that target both energy and water at the infrastructure level.

Looking ahead, the main question for the industry is how quickly closed-loop liquid cooling expands across AI deployments as new server designs increase heat density. Meta’s post suggests it sees liquid cooling and operational optimization as intertwined, with learning-based control complementing hardware-level plumbing. Observers will likely watch for more details on the roll-out scale, any expansion of open sharing through Meta’s infrastructure programs, and whether the same optimization approach can deliver comparable savings across different climate conditions and facility designs.