Oracle, AMD to Launch AI Supercluster with 50,000 GPUs in 2026

Oracle and AMD are expanding their partnership to create a publicly available AI supercluster. The collaboration will feature an initial deployment of 50,000 AMD Instinct MI450 Series GPUs, with availability scheduled to begin in 3Q26.

The partnership aims to provide customers with the necessary infrastructure to scale their AI initiatives as next-generation models outgrow the limits of existing AI clusters. "With our AMD Instinct GPUs, EPYC CPUs, and advanced AMD Pensando networking, Oracle customers gain powerful new capabilities for training, fine-tuning and deploying the next generation of AI," says Forrest Norrod, Executive Vice President and General Manager of the Data Center Solutions Business Group, AMD.

Demand for large-scale AI capacity is accelerating. To train and operate these workloads, customers require flexible and open compute solutions engineered for extreme scale and efficiency. This collaboration directly addresses that need by combining AMD’s latest processor and networking technology with Oracle Cloud Infrastructure (OCI). The resulting platform is designed to provide the performance and scalability required for the most demanding AI applications.

The new OCI AI superclusters will be powered by the AMD Helios rack design. This vertically-optimized, rack-scale architecture is engineered to deliver maximum performance, scalability, and energy efficiency for large-scale AI training and inference. The design integrates GPUs, CPUs, and advanced networking into a cohesive system.

The initial deployment of 50,000 GPUs is scheduled to begin in 3Q26, with further expansion planned for 2027 and beyond. The Helios rack design features AMD Instinct MI450 Series GPUs, next-generation AMD EPYC CPUs codenamed Venice, and next-generation AMD Pensando advanced networking codenamed Vulcano.

Each AMD Instinct MI450 Series GPU will provide up to 432GB of HBM4 memory and 20TB/s of memory bandwidth. This allows customers to train and infer models that are 50% larger than previous generations entirely in-memory, which can reduce the need for model partitioning. The system uses dense, liquid-cooled, 72-GPU racks to optimize performance density and energy efficiency.

The architecture uses an open, high-speed interconnect standard called UALink, which is transported over a UALoE fabric. This technology enables direct, hardware-coherent networking and memory sharing among GPUs within a rack to reduce latency. For scale-out networking between racks and pods, the system uses an open fabric aligned with the Ultra Ethernet Consortium (UEC) standard.

The platform will run on the open-source AMD ROCm software stack, which provides a flexible programming environment and supports popular frameworks and libraries. OCI also announced the general availability of OCI Compute with AMD Instinct MI355X GPUs, which are available in the zettascale OCI Supercluster that can scale to 131,072 GPUs.