Vera Rubin racks and Vera CPUs now generally available on OCI

Oracle Cloud Infrastructure has made Nvidia Vera Rubin NVL72 rack systems and Vera CPUs generally available; each rack connects 72 GPUs. The announcement came at Oracle AI World.

Oracle will also add Rubin systems to its multi-gigawatt-scale Superclusters offering. At a later date, OCI will add Rubin HGX NVL8 systems with eight Rubin GPUs connected through NVLink in an x86 server configuration.

What does the OCI rollout include?

Each Supercluster combines NVL72 racks with Oracle Acceleron networking and Nvidia Spectrum-X Ethernet featuring Spectrum-6 switches and BlueField-4 DPUs. Gigawatt-class power and cooling support training and serving the largest AI models.

NVLink connects 72 GPUs within each rack. OCI's resilient, multiplanar network scales across hundreds of thousands of GPUs.

The Rubin rollout builds on Grace Blackwell compute that Oracle already offers. OpenAI used an OCI Supercluster of more than 100,000 Blackwell GPUs to train GPT-6 Astra in Abilene. Oracle says standalone Vera CPU deployments target agentic AI, combining high single-thread performance with memory bandwidth to complete CPU-intensive tasks faster.

“As our frontier models advance, we need infrastructure that can keep pace,” said Sreekanth Pothanis, engineering leader, compute infra at OpenAI. “Nvidia Vera Rubin on OCI gives us the computing performance and scale to accelerate model training and serve increasingly demanding AI workloads.”

“Oracle Cloud Infrastructure gives customers the performance, security and scale to advance frontier AI,” said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure. He said the collaboration extends those capabilities across training, inference and agentic AI. Oracle says combining Rubin GPUs, Vera CPUs, Acceleron networking and AI-optimized data centers helps keep GPUs productive, secure and scalable as customers' needs grow.

Why do host CPUs matter for agents?

Oracle's multi-year collaboration with Intel will expand Xeon-based capacity on OCI, covering current Xeon platforms and planned OCI X14 offerings based on Intel's next-generation server processor, codenamed Diamond Rapids. The partnership increases CPU capacity and gives customers familiar tooling and software environments as AI agents are rapidly adopted.

Intel Data Center Group executive vice president and general manager Kevork Kechichian said agents need processors that run software, move and secure data, and accelerate inference. He said Xeon combines performance with hardware-based confidential computing, giving organizations a trusted, cost-effective foundation for AI at scale.

The expanded collaboration is intended to combine Intel's multigenerational roadmap with OCI technology to put AI to work faster and more cost-effectively. Diamond Rapids-based OCI X14 offerings will target AI and general-purpose computing. Intel Advanced Matrix Extensions will support FP8, FP16 and BF16 for inference, while Advanced Performance Extensions will help applications execute with less instruction overhead.

Oracle also nears bringing E7 instances online with sixth-generation AMD Epyc 9006 server processors. AMD Pensando DPUs will support Oracle Acceleron NICs on the AMD-powered instances, which target CPU work surrounding AI models.

Each Epyc 9006 socket provides up to 256 cores and 512 threads, up to 1,024 MB, or 1 GB, of L3 cache, 16 memory channels with transfer rates up to 8,000 MT/s, and PCIe 6.0 connectivity. AMD says the gains over fifth-generation Epyc run enterprise workloads faster and more efficiently, allowing more agents per instance and per watt.

What other chip options and timelines are in play?

Cerebras will supply wafer-scale compute for high-speed AI inference. Its processor integrates compute and memory on a single wafer-scale processor, reducing data movement and inter-chip communication.

Cerebras co-founder and CEO Andrew Feldman said every agent step requires inference for reasoning, checking work or taking action. Faster inference lets agents do more without making users wait, he said, and OCI will give AI labs, enterprises and developers the speed to build agents that do more. Cerebras Inference is planned for availability in calendar 2027.

Oracle will also bring SambaNova's SN50 online, although partnership details remain sparse and no timeline has been set. Positron's Atlas Inference Servers are generally available on OCI, purpose-built for large language models and designed for low-latency work on popular open-weight models.

The servers use Oracle Acceleron networking. Positron's first-generation Atlas chip was fabricated by Intel in the United States and is shipping to customers.

Positron claims Atlas delivers three times the compute per watt of Nvidia H100 GPUs and supports models with up to 500 billion parameters. Oracle also revealed that Google TPUs will deploy on OCI under a multi-year, multi-gigawatt contract.

For cloud buyers, OCI now offers generally available Rubin racks, Vera CPUs and Positron Atlas, with Supercluster-scale Rubin and Google capacity slated for later deployment. Cerebras Inference targets calendar 2027, while SambaNova's SN50 has no announced date, making hardware availability uneven across the portfolio.