At Oracle OpenWorld CloudWorld AI World, Oracle announced a major expansion of its AI infrastructure on Oracle Cloud Infrastructure (OCI), unveiling plans to deploy a global AI supercomputing cluster powered by 50,000 AMD Instinct MI450 series GPUs starting in the third quarter of 2026. This marks a pivotal step in Oracle’s strategy to become a leading provider of AI compute, leveraging AMD’s latest “Altair” GPU architecture. The cluster will be built using AMD’s “Helios” double-wide rack-scale design, a high-density system aimed at competing with Nvidia’s Oberon racks. Each Helios rack is expected to house 72 GPUs, with future support for AMD’s upcoming “Venice” EPYC CPUs and Pensando “Vulcano” DPUs, creating a tightly integrated AI platform. The MI450 series includes two variants: the standalone MI450, targeted at eight-way nodes, and the MI450X, designed for Helios racks with 64 or 128 GPUs. The MI450 features 2nm process technology, 432 GB of HBM4 memory, and 19.6 TB/sec memory bandwidth per GPU. In an eight-way system, it delivers 3.2 exaflops at FP4 precision. The MI450X, used in the Helios racks, is expected to scale to 128 GPUs, with each GPU capable of 50 petaflops at FP4. The full rack is projected to deliver 1.45 exaflops at FP8 and 2.9 exaflops at FP4, with 31 TB of total HBM4 memory and 1.4 PB/sec aggregate bandwidth. Oracle’s deployment will use the MI450 variant, likely customized for maximum HBM capacity, suggesting a focus on memory-intensive AI workloads. The cluster will also integrate AMD’s UALink over Ethernet (UALoE), enabling Infinity Fabric to run over Ethernet for GPU memory sharing across the cluster—potentially using Pensando DPUs as switches, bypassing traditional networking vendors like Broadcom or Cisco. This infrastructure will be part of OCI’s general public cloud, not exclusive to Oracle’s partnerships with model builders like OpenAI. Customers will be able to rent time on the MI450 cluster, similar to how they currently access the MI355X cluster, which became generally available this week. The MI355X, a previous AMD GPU, already delivers 74 petaflops at FP6/FP4 and 2.3 TB of HBM3E memory per eight-way node. The project is expected to span approximately 700 Helios racks, costing between $3.5 billion and $4 billion all-in, including storage and networking. Given the extreme scarcity of high-end GPUs and the massive demand, Oracle is unlikely to receive discounts on MI450 chips or top-tier CPUs and DPUs. Oracle’s Acceleron network architecture, which uses DPUs as integrated switches to reduce network layers, will likely be deployed across this new cluster, enhancing scalability and performance. This approach aligns with broader industry trends toward disaggregated, DPU-driven AI infrastructure. The announcement underscores Oracle’s aggressive push into the AI infrastructure market, positioning itself as a viable alternative to Nvidia-dominated ecosystems. By combining AMD’s cutting-edge GPUs with its own networking innovations, Oracle aims to capture a significant share of the enterprise AI market—especially as HBM allocations stabilize across vendors like Nvidia, AMD, and custom XPU makers. This move could reshape how enterprises access and scale AI workloads in the coming years.