What happened
Meta published an explainer on August 27 describing how it cools high-density AI servers. The company says its newest AI-optimized data centers primarily use closed-loop liquid cooling, while reinforcement learning is being used in simulations and pilots to optimize cooling operations.
Meta says the growing power and heat output of AI hardware is making conventional air cooling less efficient for dense server deployments. The company contrasts newer AI infrastructure with an earlier data-center example in Altoona, Iowa, where racks containing 16 Nvidia H100 GPUs were cooled entirely with air and used minimal water. According to Meta, newer hardware designs have increased the need for a different approach. The article’s central claim is that cooling is an important part of AI infrastructure, not merely a building-services detail.
The system Meta describes circulates a mixture of water and glycol directly through server hardware to carry heat away from the racks. Heat exchangers then transfer heat out of the coolant before the same mixture returns to the servers in a continuous loop. Meta says the liquid is not expelled from the facility and expects the coolant to remain in service for up to a decade without replacement. Where a building lacks integrated liquid-cooling infrastructure, the company says it uses air-assisted liquid cooling: distributed racks containing pumps and heat exchangers that provide the same basic closed-loop process on a smaller scale.
Meta also argues that liquid cooling can improve physical density. Its explainer says air cooling the same servers would require nearly twice the server-tray space because of the additional cooling equipment. Direct-to-chip liquid cooling, by contrast, can allow more GPUs to fit in the same rack, reducing the number of racks needed for a facility with a given computing capacity. Meta further says a typical AI-optimized data center using closed-loop cooling and dry coolers consumes less water annually than a couple of full-service restaurants, but it does not provide the underlying annual volumes or a precise basis for that comparison.
The article adds a software component. Meta says engineers built a physics-based data-center simulator that models weather, server load and cooling-equipment behavior. Reinforcement learning can test cooling decisions in that simulated environment instead of experimenting first on a live facility. Meta says the approach has been extended from an experiment to air-cooled data centers across its fleet and that a pilot at one site reduced energy used by air-cooling supply fans by an average of 20% while reducing water use by 4% across different weather conditions. The source does not identify the pilot site, its duration, its baseline consumption or the number of facilities involved in the broader rollout.
Read the source: about.fb.com ↗
Why it matters
Cooling is becoming a practical constraint as AI hardware generates more heat and requires denser deployments. Meta’s account suggests that liquid cooling and software-based optimization could affect the energy, water, space and infrastructure requirements of large AI computing facilities, although the reported results come from Meta and one pilot.
AI infrastructure is often discussed in terms of chips and electricity, but the source highlights a less visible physical bottleneck: removing heat from tightly packed computing equipment. If liquid cooling allows more GPUs in the same rack, operators may be able to increase computing capacity without expanding the building footprint at the same rate. That could influence the design and cost of facilities used to train and run AI systems.
Water use is a public concern around data centers, but the source correctly presents it as dependent on the cooling design rather than an automatic feature of AI computing. Meta’s closed-loop description indicates that water can function primarily as a recirculating coolant instead of being continually consumed or discharged. This distinction is practically important for communities evaluating proposed facilities, though the article’s restaurant comparison is not enough to establish the total water impact of Meta’s data-center fleet.
The reported reinforcement-learning pilot is relevant because it connects AI methods to the operation of AI infrastructure. The model is not described as controlling a general-purpose system or making unsupervised decisions about public services; it is used to test cooling strategies in a simulator that represents physical conditions. Meta says the resulting policy reduced fan energy by 20% on average and water use by 4% in the pilot. If replicated, such optimization could lower operating costs and resource demand without requiring a change to the underlying AI workload.
The source also describes an industry-sharing pathway. Meta says its Open Compute Project, founded in 2011, is intended to make data-center infrastructure more efficient, scalable and sustainable, and notes that the company announced the IcePack liquid-cooled network-rack platform in 2025 for free sharing through that project. That earlier announcement is background rather than new news in this article. The new, timely contribution is Meta’s detailed explanation of the cooling architecture and its account of reinforcement-learning results, not a new product launch or independently verified industry-wide standard.
What to watch next
The key questions are whether Meta’s reported savings hold across more sites, climates and workloads, and how much water and energy the systems use in absolute terms. Watch for independent validation, details about deployment scale, and evidence that the open designs and reinforcement-learning methods are useful beyond Meta’s own facilities.
The most important limitation is verification. All performance figures in the source are Meta’s own claims, and the article does not cite an independent audit, a peer-reviewed study or measurements from outside the company. The 20% fan-energy reduction and 4% water reduction therefore should be treated as reported pilot results, not established results for AI data centers generally. Future reporting should seek the pilot’s site, duration, weather range, workload profile, baseline measurements and statistical variation.
Absolute resource use remains unknown. Meta gives relative reductions and a comparison with two full-service restaurants, but does not state the data center’s annual water consumption, electricity use, cooling load or total number of GPUs. The comparison also does not explain whether it includes water used elsewhere in the facility or in electricity generation. Without those figures, readers cannot calculate the scale of the savings or compare the system with other cooling designs.
Deployment scope is another open question. Meta says the majority of its newest AI-optimized data centers use closed-loop liquid cooling and that the reinforcement-learning approach has been scaled to air-cooled data centers in its fleet, but it does not identify how many facilities are covered or whether the systems are operating continuously. The company also says coolant can last up to a decade, an expectation that may depend on maintenance, contamination controls, equipment design and local operating conditions that the source does not discuss.
The broader test will be whether the approach transfers beyond Meta. Closed-loop performance may vary with climate, building design, hardware generation, rack configuration and workload. Air-assisted systems may be useful for retrofits, but the source provides no comparative cost, reliability or maintenance data. Watch for independent measurements, disclosures from organizations using IcePack or similar designs, and evidence that simulation-trained cooling policies remain safe when conditions fall outside the scenarios modeled. It is also not clear whether future, more power-dense AI hardware will preserve the reported efficiency gains or create new cooling requirements.
