Meta is utilizing closed-loop liquid cooling to enhance the efficiency of its AI infrastructure, as explained by Tom Shaw. This approach addresses the engineering challenge of cooling servers, which has intensified with the increasing power and heat generation of AI hardware.
Traditional data centers often rely on air cooling. However, newer AI hardware designs necessitate more optimal cooling methods. Meta's approach involves recirculating a liquid coolant within a sealed loop, significantly reducing ongoing water usage.
Key Points
- Meta is using closed-loop liquid cooling to power AI more efficiently.
- The system recirculates a liquid coolant (water and glycol mix) through server hardware to remove heat.
- The coolant is pumped through heat exchangers to dissipate heat and is then returned to the servers in a continuous loop.
- Meta anticipates using these coolants for up to a decade without replacement.
- For facilities without built-in liquid cooling infrastructure, Meta uses Air-Assisted Liquid Cooling, a smaller, distributed closed-loop system.
- This method allows for fitting more GPUs into the same server rack, reducing the number of racks required for a given capacity.
- Meta is sharing these advances through the Open Compute Project, including IcePack, a liquid-cooled network rack platform announced in 2025.
- Reinforcement learning is being used to optimize the design and operations of Meta's cooling systems, including air-cooled data centers.
Context
According to Meta, the increasing power of AI hardware generates more heat, making traditional air cooling less efficient. While air cooling was sufficient for Nvidia H100s in some data centers a few years ago, current AI hardware demands a more advanced solution. Meta's closed-loop system is designed to be resource-efficient, using less water annually than a couple of full-service restaurants.
Why It Matters
This shift in cooling technology indicates a change in how builders and operators must design and manage data center infrastructure for AI. The adoption of closed-loop liquid cooling and the sharing of designs through the Open Compute Project could influence industry standards for efficiency and scalability in AI deployments.
What To Do
- Note the efficiency gains of closed-loop liquid cooling for high-density AI hardware.
- Watch for further developments and shared designs from Meta's Open Compute Project regarding cooling solutions.
- Consider the implications of reinforcement learning for optimizing data center operations.
- Compare the resource usage of closed-loop liquid cooling against traditional air cooling methods for new AI deployments.