Meta builds Prometheus supercluster

Meta completes the first phase of Prometheus, a gigawatt supercluster with hundreds of thousands of GPUs in Ohio. The infrastructure uses closed-loop liquid cooling and two NVIDIA platforms.

The Clemente platform offers 288 GB of HBM3e per GPU and more than 1 exaFLOP of FP4. Prometheus has dedicated networks for GPU-GPU communication, with high bandwidth and low latency. Microsoft unveiled the Maia 300, the third generation of its AI accelerator for Azure, expected in 2027. The Maia 300 is based on the TSMC N2P (2 nm) process and has 288 GB of HBM4. It delivers 13.8 petaOPS in FP4 and 6.9 petaOPS in FP8. It will be Microsoft's main inference system, promising 40% more tokens per megawatt than GPU-based systems.

Meta builds Prometheus supercluster with liquid cooling — Image: engineering.fb.com

Meta is expanding its AI infrastructure with Prometheus, while Microsoft develops the Maia 300, a 2 nm AI chip that promises greater efficiency in tokens per megawatt. ASUS has launched an Intel GPU for professional use, completing the trio of professional GPUs with NVIDIA and AMD. Closed-loop liquid cooling also reduces the risk of failures and extends the lifespan of components. The combination of Catalina and Clemente, NVIDIA platforms, provides flexibility for different types of workloads, from training to inference. The scale of Prometheus is a significant shift.

Meta and rivals advance in AI with new technologies

Meta and rivals advance in AI with new technologies. Meta is moving forward with Prometheus, its first gigawatt supercluster, while Microsoft and ASUS are also positioning themselves in the AI computing market. Meta's expansion of its AI infrastructure shows that the competition for efficiency and scalability is accelerating. By turning to an Intel model, ASUS demonstrates that demand for AI processing is becoming universal. Microsoft, with the Maia 300, seeks to outperform NVIDIA's GPUs in performance per unit of energy. Prometheus, with its closed-loop liquid cooling and hundreds of thousands of GPUs, represents a leap in the company's processing capacity.

The Catalina platform uses the GB200 NVL72, while Clemente uses the GB300 NVL72, with dedicated networks for GPU-GPU communication. The infrastructure is built to handle massive AI demands, enabling scalability and efficiency at large scale.

Dedicated networks speed up parallel processes

Meta is building the Prometheus supercluster with liquid cooling and next-generation NVIDIA GPUs, redefining scale and efficiency in AI. Closed-loop cooling allows more GPUs in a single space without thermal overload, maintaining stability in gigawatt-scale clusters.