Oracle and NVIDIA are expanding full stack collaboration with the General Availability of NVIDIA Vera Rubin NVL72, and NVIDIA Vera CPUs on Oracle Cloud Infrastructure (OCI); and introducing NVIDIA Rubin NVL8 GPUs on OCI, as well as multiple PaaS and SaaS capabilities that bring accelerated computing to enterprise AI applications and Oracle Database.
Announcing General Availability of Vera Rubin NVL72, Vera CPUs, and introducing Rubin on OCI
Oracle is announcing general availability of NVIDIA Vera Rubin NVL72 on OCI, expanding the configurations available to customers building their frontier models on NVIDIA’s newest AI infrastructure on Oracle. Vera Rubin delivers up to 30 times higher inference throughput per watt and a 4x leap in training performance compared to the prior NVIDIA Blackwell generation. By combining ultra-fast HBM4 memory with an advanced NVIDIA NVLink 6 scale-up fabric, Vera Rubin NVL72 effectively eliminates data-movement bottlenecks to drastically lower the operational cost of running complex, multi-step agentic AI and large-scale reasoning models.
At this scale, data center design is fundamental to enabling computing platform capability. 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. The infrastructure is designed to remain productive as customer needs change. Capacity can move between frontier training and production inference, and the campuses can accommodate successive NVIDIA accelerated computing platforms. This helps customers sustain compute utilization as models, hardware, and agentic workloads evolve.
“As our frontier models advance, we need infrastructure that can keep pace,” said Sreekanth Pothanis, engineering leader, compute infrastructure at OpenAI. “NVIDIA Vera Rubin on OCI gives us the computing performance and scale to accelerate model training and serve increasingly demanding AI workloads.”
OCI also plans to introduce NVIDIA Rubin NVL8, delivering eight Rubin GPUs connected through NVLink in an x86-based server configuration. The planned platform will combine Intel Xeon 6 processors with high-speed RDMA networking to deliver the performance and flexibility needed for AI training, fine-tuning, and inference Rubin NVL8 enables customers to adopt next-generation Rubin GPU technology in a smaller eight-GPU configuration while maintaining compatibility with x86-based applications.
OpenAI provides an example of of scale in practice. GPT-6 Astra, the latest and most performance OpenAI model available today, was trained and served on an OCI Supercluster with more than 100,000 NVIDIA Blackwell GPUs in Abilene, TX, as announced by OpenAI President Greg Brockman at GPT-6 Astra’s launch announcement in September. NVIDIA’s CEO, Jensen Huang, took to X after this, stating that “AGI has arrived.”
Announcing NVIDIA Vera CPUs on OCI: max single-threaded capability for agentic workloads
Oracle is also announcing general availability of NVIDIA Vera CPUs on OCI, with an initial deployment of 150,000 CPUs to support the growing demand for Agentic AI workloads. This expansion complements OCI’s GPU infrastructure with high-performance CPU capacity designed to accelerate agentic AI applications and improve overall AI system efficiency.
Oracle is the first hyperscaler to bring Vera CPUs to market.
Vera is NVIDIA’s first fully custom CPU, with a ground-up core called NVIDIA Olympus, built around the agent loop rather than around raw floating point throughput. Each CPU carries 88 cores and 176 threads through NVIDIA’s spatial multithreading, up to 1.2TB/s of LPDDR5X memory, and 3.4 TB/s per second of bandwidth.
Vera’s unique characteristics make it particularly well suited for multiple inference and agentic AI workloads, including:
- Agentic orchestration and tool serving: This is the workload Vera was designed for, and it is the one AI native teams keep raising with Oracle first.
- Sandbox execution at density: This is the workload that has generated the most direct pull from AI native companies.
- Control plane and orchestration for GPU fleets: This is a use case that comes up in almost every conversation once the obvious ones are covered.
- Vector search and retrieval: For a retrieval-augmented agent, this pairs naturally with the orchestration workload; where the same socket running the agent brain can hold the retrieval index for agent queries – co-location that retrieval-side optimizations people reach for first.
- In-memory databases and columnar analytics: The database result is where the independent benchmarks are strongest for Vera. When customers run operational analytics next to their AI workloads, an in-memory database, a feature store, a real-time aggregation layer, Vera running on the same Arm platform as their orchestration simplifies their fleet.
- Operational applications and general services: Not everything in an AI native company is an agent. Consolidating general services onto the same high-density Arm platform as the agentic workloads is a real operational simplification, one fewer architecture to run and worry about.
Learn more about ideal workloads in real-world environments through this more detailed blog.
Multiple customers are harnessing the benefits on Vera CPUs on OCI through early access to some of the first production systems available.
“NVIDIA Vera CPUs on OCI deliver the world-class loaded performance needed to execute code and run tools across our rapidly growing agentic sandbox workloads,” said Uday Ruddarraju, chief technology officer of compute at OpenAI. “We are bringing Vera into production on OCI to support these workloads at scale.”
Scaling out with Oracle Acceleron
NVIDIA Vera Rubin NVL72 and Vera CPUs, and Rubin NVL8 all run on Oracle Acceleron networking for all deployments across OCI. Oracle Acceleron is built on 102.4 Tb/s NVIDIA Spectrum-X Ethernet networking switches with advanced congestion control and zero jitter communication, ConnectX-9 SuperNICs and NVIDIA Bluefield-4 DPUs to deliver powerful computing and accelerations for networking, data storage, and cybersecurity.
Oracle Acceleron RDMA over Converged Ethernet (RoCE), allows customers to scale beyond a single NVLink domain. Its wide, shallow, multiplanar fabric uses the GPU network interface card as a mini-switch across physically and logically isolated network planes. Fewer network tiers help provide more uniform GPU-to-GPU latency. When a network plane becomes congested or unstable, traffic can shift to another plane, helping training jobs continue without costly checkpoint restarts. This architecture combines tightly connected GPUs within each pod with the communication fabric needed to build much larger clusters.
Instead of relying on a single network plane, Oracle Acceleron uses a multi-planar network fabric to distribute traffic across multiple independent planes. This provides greater aggregate bandwidth, isolates failures, and allows the network to continue operating efficiently even as workloads become larger and more dynamic. For customers, the benefit is easier to express than the topology: less waiting on communication, greater flexibility, security, and control; and a system that can continue doing useful work even if part of the network has a problem. This is critical to enable a network that keeps delivering predictable performance as OCI Superclusters have grown from thousands to hundreds of thousands – and now to millions of GPUs.
Oracle Acceleron leverages Multipath Reliable Connection (MRC) as a transport layer protocol leverages across sites. The key idea behind MRC is simple: instead of sending a GPU-to-GPU transfer down a single network path, MRC can spread packets across hundreds of paths simultaneously. This is accomplished through an eight-plane architecture, where every server can connect to all 8 planes at the same time, and any two servers have 8 independent paths between them. This enables three critical capabilities:
- Performance and predictable GPU utilization: By dynamically balancing traffic across the fabric, MRC minimizes congestion and tail latency, helping keep expensive GPUs productive rather than waiting on the network. MRC “sprays” packets from a single connection across hundreds of available network paths rather than pinning the flow to one path. This distributes load, avoids hotspots, reduces tail latency, and enables high throughput across a multiplanar fabric.
- Built-in resiliency with near-instant failure recovery: If a link or path becomes congested or fails, MRC can detect it and route around it in microseconds, allowing large training jobs to keep running through failures and maintenance events.
- Scale and simplicity: Combined with Oracle Acceleron Multiplanar Networking and SRv6 source routing, MRC enables two-tier networks supporting more than 131,000 GPUs, with fewer components, lower power consumption, and a simpler network control plane. With MRC, we can do this across only two tiers of ethernet switches (vs. 3-4 in conventional designs). And we can route intelligence toward the NIC/edge, reducing core routing complexity while also lowering component count, power consumption, and cost.
Extending accelerated computing across the Oracle stack
The collaboration extends from infrastructure into platform services and enterprise software. These capabilities bring NVIDIA technologies to the environments where customers deploy inference workloads, build database indexes, and develop AI applications.
- NVIDIA Dynamo brings disaggregated inference to Oracle Kubernetes Engine (OKE), extending the infrastructure story to containerized AI workloads. This adds an inference software capability alongside the compute and networking available on OCI.
- Oracle Database Index Service v2 uses NVIDIA cuVS to accelerate vector search and index builds, enabling 5x faster index builds compared to CPU-offloading.
- Oracle AI Accelerator Packs with NVIDIA cuOpt extend the collaboration into optimization. NVIDIA cuOpt is an open-source GPU-accelerated optimization engine delivering near real-time solutions for complex problems.
- NVIDIA Nemotron is powering Oracle Health Data Intelligence’s cohorting agent, enabling healthcare teams to query patient data in natural language. Oracle plans to deploy the agent in EMEA, with a wider deployment planned Q1 next year.
- Oracle and NVIDIA are jointly optimizing Oracle’s AI Media Models for image generation using NVIDIA GB300, with initial optimizations integrated and further improvements under evaluation. The Oracle AI Media Models are currently available in the U.S. only.
NVIDIA & Oracle enable the next generation of agentic AI
Oracle and NVIDIA continue to expand a multi-year collaboration that connects AI infrastructure with the software customers use to put it to work – with a full stack of offerings across IaaS, PaaS, and SaaS. The latest, fastest, and most powerful NVIDIA AI Platforms, Larger NVLink domains, Oracle Acceleron networking, and state-of-the-art OCI Superclusters custom-engineered for AI support the next wave of training and inference. Vera CPUs and accelerated software extend those capabilities into agentic workloads and enterprise applications, giving customers more ways to build and run AI on OCI.
Read the Press Release accompanying this blog, watch this short video about Oracle and NVIDIA’s history of co-innovation, and learn more about our collaboration.
Future Product Disclaimer: The preceding is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, timing, and pricing of any features or functionality described for Oracle’s products may change and remains at the sole discretion of Oracle Corporation.
Prathiba Srinivasan is a Senior Product Marketing Manager at NVIDIA, focused on Accelerated Computing products, with experience in GPU product marketing and enterprise AI go-to-market.