NVIDIA 與 Oracle 計劃將 OCI 上的 Vera Rubin NVL72 Superclusters 擴展至多吉瓦規模

NVIDIA and Oracle announced on October 27, 2026, plans to expand Oracle Cloud Infrastructure (OCI) Superclusters with NVIDIA Vera Rubin NVL72 systems to multi-gigawatt scale, citing growing customer demand for AI training, inference and agentic applications. The companies also announced general availability of Vera Rubin NVL72 and NVIDIA Vera CPUs on OCI, and said OCI plans to introduce NVIDIA HGX Rubin NVL8 systems.

According to the announcement, the expanded collaboration gives customers infrastructure for both the GPU-intensive work of training and serving frontier models and the CPU-intensive tasks agents perform between model calls, including executing tools, testing code and retrieving data. Customers can scale CPU and GPU resources independently to match their workloads, which the companies said helps reduce idle capacity.

OCI Supercluster Architecture and Vera Rubin Performance

OCI Superclusters combine NVIDIA NVL72 rack-scale systems, Oracle Acceleron networking with NVIDIA Spectrum-X Ethernet featuring Spectrum-6 switches and BlueField-4 DPUs, and gigawatt-class power and cooling. Each rack connects 72 GPUs through NVIDIA NVLink scale-up networking, while OCI’s multiplanar network scales across hundreds of thousands of GPUs.

The planned Vera Rubin expansion builds on OCI’s existing NVIDIA GB200 and GB300 NVL72 infrastructure, with NVIDIA’s common CUDA software foundation and continued software optimization supporting workloads across hardware generations.

In a companion Oracle post, the company said Vera Rubin delivers up to 30 times higher inference throughput per watt and a 4x training-performance gain over the prior Blackwell generation. The post states that OCI Superclusters run in gigawatt-class campuses optimized for density within a two-kilometer radius, with power delivery and liquid cooling engineered for systems that span multiple racks, and that capacity can move between frontier training and production inference.

Oracle Acceleron is built on 102.4 Tb/s Spectrum-X Ethernet switches with ConnectX-9 SuperNICs and BlueField-4 DPUs, according to the post. Its Multipath Reliable Connection protocol spreads packets from a single connection across hundreds of paths through an eight-plane architecture, and combined with Oracle’s multiplanar networking and SRv6 source routing enables two-tier networks supporting more than 131,000 GPUs.

Planned HGX Rubin NVL8 Systems

OCI plans to introduce NVIDIA HGX Rubin NVL8, delivering eight Rubin GPUs connected through NVLink in an x86-based server configuration. The planned platform features two 64-core Intel Xeon 6 CPUs, 4TB of DDR5 memory and 6.4 Tbps of high-speed RDMA networking, providing a deployment option for AI training, fine-tuning and inference. According to the announcement, HGX Rubin NVL8 enables customers to adopt next-generation Rubin GPU technology in a smaller eight-GPU deployment while maintaining compatibility with x86-based applications.

NVIDIA Vera CPUs on OCI

Oracle also announced general availability of OCI Compute with NVIDIA Vera CPUs, with an initial deployment of 150,000 CPUs across Oracle’s regions. A separate Oracle availability post states the CPUs are offered through Bare Metal and Flex Virtual Machine instances and will be available in select regions in the fourth quarter of 2026. Oracle described itself as the first hyperscaler to bring Vera CPUs to market.

Each Vera processor carries 88 custom Arm-compatible NVIDIA Olympus cores and 176 threads enabled by NVIDIA Spatial Multithreading, with up to 1.2 TB/s of memory bandwidth and a second-generation NVIDIA Scalable Coherency Fabric providing 3.4 TB/s of bisectional bandwidth. OCI’s dual-processor instances provide 176 cores and 352 threads across two Vera processors, 1TB of LPDDR5X system memory, 7.68TB of local NVMe storage and 200 Gbps Acceleron networking.

NVIDIA reports up to 80% faster agentic CPU performance compared to leading x86 CPUs. Oracle’s post notes those results are based on NVIDIA platform benchmarks rather than OCI instance-level testing, and that actual performance varies by workload and configuration.

The announcement frames the CPUs around the work behind AI agents: agents call tools, execute code and retrieve information between model calls, and slow execution of those CPU-intensive operations can delay results and leave GPUs waiting. Within that model, OCI Kubernetes Engine enables customers to scale CPU and GPU resources independently, while Oracle Acceleron provides hardware-based isolation for workloads running untrusted, AI-generated code.

OpenAI Deployment and Executive Statements

OpenAI has trained and served GPT-6 Astra on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs in Abilene, Texas.

“NVIDIA Vera Rubin on OCI gives us the computing performance and scale to accelerate model training and serve increasingly demanding AI workloads,” said Sreekanth Pothanis, engineering leader for compute infrastructure at OpenAI.

Uday Ruddarraju, chief technology officer of compute at OpenAI, said Vera CPUs on OCI deliver the performance needed to execute code and run tools across the company’s rapidly growing agentic sandbox workloads. “We are bringing Vera into production on OCI to support these workloads at scale,” he said.

Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, said OCI gives customers the performance, security and scale to advance frontier AI, and that the NVIDIA collaboration extends those capabilities across training, inference and agentic AI. Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, said the companies are bringing Vera Rubin and Vera CPUs to OCI to help customers put more of their infrastructure to productive use and sustain its value as AI evolves.

Software Across the Oracle Stack

The collaboration extends from infrastructure into platform services and enterprise software, according to Oracle’s post. NVIDIA Dynamo brings disaggregated inference to Oracle Kubernetes Engine, adding an inference software capability alongside OCI’s compute and networking. Oracle Database Index Service v2 uses NVIDIA cuVS to accelerate vector search and index builds, which Oracle said enables 5x faster index builds compared to CPU-offloading. Oracle AI Accelerator Packs incorporate NVIDIA cuOpt, an open-source GPU-accelerated optimization engine.

Oracle and NVIDIA are jointly optimizing Oracle AI Media Models for image generation using NVIDIA GB300, with initial optimizations integrated and the models currently available in the United States only. NVIDIA Nemotron powers Oracle Health Data Intelligence’s cohorting agent, which lets healthcare teams query patient data in natural language; Oracle plans to deploy the agent in EMEA, with wider deployment planned for the first quarter of 2027.