Oracle Says It Will Deploy Hundreds of Thousands of NVIDIA Vera CPUs

Oracle Cloud Infrastructure says it will deploy hundreds of thousands of NVIDIA Vera CPU units beginning in 2026 and offer them as bare-metal and virtual instances. The commitment puts large cloud capacity behind a processor designed for AI-agent support work, even though its production power and cost profile is not yet public.

Vera is an 88-core, 176-thread Arm CPU based on NVIDIA’s Olympus cores. Each socket supports up to 1.5 TB of LPDDR5X memory and 1.2 TB/s of memory bandwidth, giving it a large pool of fast-access memory for data-heavy work.

A CPU for the steps between model responses

OCI is pitching Vera for work that happens around a model’s response: coordinating agents, serving tools, running isolated code, searching data and managing GPU fleets. These tasks can depend on quickly moving and checking many small pieces of information while many jobs run at once.

• Agent orchestration and tool serving, which coordinate sessions and calls to outside software.
• Sandbox execution for agent-written code and reinforcement-learning environments.
• Vector search, retrieval and in-memory databases that keep data close to the processor.
• Control planes that schedule, route and monitor GPU fleets.

Promising tests, incomplete buying evidence

OCI says it received an early Vera system in May alongside Anthropic, OpenAI and SpaceX. OpenAI CTO of Compute Uday Ruddarraju called Vera’s memory bandwidth and per-core performance on OCI promising for emerging agentic AI workloads.

OCI cites Prime Intellect testing in which one Vera socket ran 176 virtual machines and delivered roughly 30% more throughput per CPU than AMD Zen 5 on specified reinforcement-learning sandbox workloads. It also cites Phoronix results placing a single Vera socket about 1.5 times ahead of a current 128-core x86 system across the tested workloads.

Those tests used pre-production silicon in an open chassis, according to OCI. That leaves production thermal behavior, power use, price performance and performance per watt unverified. NVIDIA separately describes Vera as built for loaded, sequential work in agentic AI and reinforcement learning; those performance claims are NVIDIA’s own.

Oracle’s plan is a direct bet that the CPU layer can become a limiting part of AI-agent systems. The unresolved question is whether Vera’s early benchmark lead carries over to the power, cooling and cost conditions of production cloud use.