Meta's AI data centers are being constructed in tent structures, with Zuckerberg revealing ambitious Blueprints for 'numerous multi-gigawatt clusters,' such as Hyperion, which is nearly as expansive as Manhattan.
In a bold move to stay ahead in the global AI race, Meta, the tech giant behind Facebook, has announced an ambitious infrastructure plan for AI training and inference. Inspired by the principles of rapid iteration and aggressive scaling from Elon Musk's xAI, this strategy aims to dramatically cut construction time and operational costs while maintaining tight control over power and compute resources.
At the heart of this plan is the rapid deployment of AI data centers using unconventional data center designs. Instead of traditional roofed buildings, Meta is constructing AI data centers in "tents," a approach that significantly speeds up construction and deployment, allowing Meta to scale its infrastructure rapidly to meet AI training demands.
Another key aspect of Meta's strategy is the creation of enormous AI training clusters. The "Prometheus" cluster in Ohio, set to come online in 2026, will boast a 1GW capacity, providing over three billion TFLOPS of compute power. Meta also plans to build another cluster, named "Hyperion," which will require up to 5GW of power. These clusters will be built on a hybrid infrastructure strategy, combining self-built campuses, leased facilities, AI-optimized designs, multi-datacenter campus setups, and high-bandwidth networking to maximize efficiency and flexibility while cutting costs and time.
To ensure reliability and cost control, Meta is also building on-site natural gas power plants, totalling hundreds of megawatts, with equipment drawn from suppliers like Solar Turbines and Siemens. This reduces dependency on external power grids, improves energy cost control, and enhances operational reliability.
Meta's new data centers will draw power from on-site meta substations and employ smart workload management tools to maximize power available to each datacenter. The new Meta sites will also do away with traditional backup diesel generators, instead relying on ultra-high-bandwidth networks powered by advanced switches, such as the Arista 7808 with Broadcom ASICs.
Mark Zuckerberg, Meta's CEO, stated that Meta Superintelligence Labs will have industry-leading levels of compute, and the company is prioritizing building data centers with AI in mind. The new datacenter strategy will see Meta accelerating its investment in large-scale data centers for artificial intelligence training and inference, with each of Meta's "titan clusters" potentially as large as a sizeable portion of Manhattan.
With this infrastructure plan, Meta aims to outpace rivals by owning and optimizing its entire AI compute stack, avoiding third-party cloud costs, and deploying capacity at a scale comparable to or exceeding the largest hyperscalers. This strategic shift underscores Meta's commitment to leading the way in AI technology and innovation.
Data-and-cloud-computing technology plays a vital role in Meta's infrastructure plan for AI training and inference, as the company invests heavily in large-scale data centers with AI-optimized designs. The strategy also includes the use of advanced switches, such as the Arista 7808 with Broadcom ASICs, to power ultra-high-bandwidth networks and eliminate traditional backup diesel generators.
Meta's new AI data centers, equipped with industry-leading compute resources, will be built on a hybrid infrastructure strategy that combines self-built campuses, leased facilities, AI-optimized designs, multi-datacenter campus setups, and high-bandwidth networking to maximize efficiency and flexibility while cutting costs and time.