As the demand for artificial intelligence intensifies, finding efficient ways to cool servers has become a major engineering challenge. Enhanced AI hardware generates more heat, rendering traditional air-cooling methods less effective. This shift was evident during a visit to Meta’s AI infrastructure in Texas, where the focus was on a different kind of technology: the plumbing system behind it.
The shift in cooling strategies
Data centers have typically used air cooling to maintain hardware temperatures, suitable for tasks like social media interactions. Even recently, air cooling was sufficient for AI hardware. A visit to a data center in Altoona, Iowa, demonstrated this, where racks of 16 Nvidia H100s were cooled entirely with air, using minimal water for cooling air during warmer periods.
However, newer AI hardware designs have necessitated a more effective cooling method, leading to the adoption of closed-loop liquid cooling. Despite misconceptions, AI data centers are not huge water consumers. Meta’s data centers use a closed-loop system, which reuses water continuously, thus reducing water use significantly.
How closed-loop cooling works
Closed-loop liquid cooling is straightforward. A coolant, made of water and glycol, circulates through server hardware to absorb heat. Instead of expelling the heated liquid, it travels through heat exchangers that dissipate heat before being recirculated. This system allows Meta to use the same coolant for up to a decade without replacement.
In locations lacking built-in liquid cooling infrastructure, Meta employs Air-Assisted Liquid Cooling. This involves racks equipped with pumps and heat exchangers, mimicking the large-scale system on a smaller scale. This method is resource-efficient, as AI-optimized data centers with closed-loop systems use less water annually than a couple of full-service restaurants.
Efficiency and capacity optimization
Beyond saving water, this cooling system effectively utilizes space within server racks. Air cooling the same servers would require nearly double the space, leading to diminishing returns. Closed-loop liquid cooling allows more GPUs to fit in a given rack, reducing the overall number of racks needed and enabling capacity scaling without expanding the facility size.
Meta’s open-source cooling initiatives
Meta not only designs its own systems across its infrastructure but also shares these advancements through the Open Compute Project, an initiative aimed at improving data center efficiency. In 2025, Meta released IcePack, a liquid-cooled network rack platform, freely accessible through this project.
AI-driven cooling optimization
Meta’s engineering teams have embraced reinforcement learning to optimize cooling infrastructure. This approach, initially experimental, has been effectively scaled to Meta’s air-cooled data centers. By using a physics-based simulator, the team tested variables like weather and server load, reducing cooling energy consumption by 20% and water usage by 4% in a pilot center. These significant reductions enhance efficiency across their data center operations.