Oracle adds NVIDIA’s Vera Rubin to its cloud, with multi-gigawatt growth planned

Oracle adds NVIDIA Vera CPUs for AI agents, with 150,000 planned initially

AI agents’ tool calls and code execution are getting dedicated computing capacity on Oracle’s cloud. Oracle announced NVIDIA Vera-powered Compute instances on October 27, 2026, with an initial 150,000 CPUs planned across its regions. The announcement calls the offering generally available, but also says access will begin in select regions this quarter.

Oracle’s two availability statements leave an important distinction: the general-availability announcement is not a promise of immediate access in every region. Its stated rollout begins in select regions during the current quarter, while the 150,000-CPU figure describes the initial capacity it intends to make available across Oracle regions.

Customers can choose bare-metal servers or Flex Virtual Machine instances. Oracle also positions Vera as a standalone CPU platform, allowing agent services to run independently of GPU infrastructure. It can support GPU-based environments where needed, rather than requiring those surrounding services to be deployed as part of a GPU system.

The pitch addresses a specific delay in agent workflows. A GPU can generate a model response, but the next model step may have to wait for a CPU to execute code, run a tool or retrieve information. Oracle says faster completion of those jobs can shorten the wait between model interactions and improve an agent’s overall response time.

Oracle identifies several kinds of work for the new instances. They extend beyond producing model responses to the software that coordinates agents, prepares their data and manages the computing resources around them:

Coordinating agent sessions, executing tools and processing their results.

Running isolated code, tests and simulations, including environments used for reinforcement learning.

Searching and retrieving data, including vector search and in-memory analytics.

Scheduling GPU workloads, routing requests and supporting monitoring and data pipelines.

Reinforcement learning adds another reason to care about CPU speed. In that process, a system acts in an environment and receives feedback. Oracle says faster environment execution can accelerate those feedback loops and increase training throughput. That is a different benefit from making an individual model response arrive faster.

Oracle says multiple customers have used early-access production systems. One named customer is OpenAI: Uday Ruddarraju, its chief technology officer of compute, said the company is bringing Vera into production on OCI to support its growing agentic sandbox workloads, where agents execute code and run tools.

Each Vera processor uses 88 custom Arm-compatible Olympus cores and supports up to 1.2 TB per second of memory bandwidth. NVIDIA’s design combines a single compute die with a shared system for keeping data consistent across the processor, which Oracle says is intended to support many concurrent jobs. The OCI configuration also includes 7.68 TB of local NVMe storage for fast data access.

The largest performance figure needs a separate qualification. NVIDIA reports up to 80% faster agentic CPU performance than leading x86 CPUs. Those results come from NVIDIA platform benchmarks, not OCI instance-level testing, and actual performance varies with the workload and configuration. The figure therefore is not a measured speedup for an agent running on Oracle’s new instances.

Oracle also argues that shortening CPU-bound steps can help GPUs retain cached context—the stored information from earlier model processing. Preserving that context can reduce repeated computation, letting GPUs spend more time generating responses rather than waiting for dependent tasks. That remains a stated benefit of the design, rather than a quantified OCI performance result.