Nvidia delivers first Vera CPUs to Anthropic & OpenAI

NVIDIA has delivered its first Vera CPU systems to Anthropic, OpenAI, SpaceXAI and Oracle Cloud Infrastructure, marking the chip's first customer rollout.

Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, personally handed over the systems at the four organisations' California sites as the company begins placing the processor with major AI users and cloud operators.

Vera is NVIDIA's first custom CPU. It is aimed at what NVIDIA describes as agentic AI workloads, in which systems do more than generate responses and instead run software tools, manage orchestration layers, retrieve information and handle other background computing tasks.

The launch offers a clearer view of NVIDIA's push beyond graphics processors and networking products. Chief executive Jensen Huang introduced Vera earlier this year and described it as the foundation of a major new business.

NVIDIA says the processor has 88 custom Olympus cores, 1.2 TB/s of memory bandwidth and per-core performance that is 50% faster under full load than existing designs used for comparable tasks. It is designed for sustained, concurrent workloads that arise when AI systems call external tools, manage long context windows and run reinforcement learning jobs.

Early users

At Anthropic's San Francisco office, James Bradbury, head of compute, received one of the first systems. He said the lab is interested in how the chip could fit into the broader infrastructure used to train and run models.

"Scaling compute is an important accelerant for the growth of models," Bradbury said. "We're excited to see Vera emerge as a promising part of the ecosystem when solving for agentic workloads."

OpenAI also took delivery of a Vera system at its Mission Bay headquarters. NVIDIA said Sachin Katti, head of compute infrastructure at OpenAI, accepted the handoff as Buck walked through the machine's design, though OpenAI did not provide a direct quote.

SpaceXAI received a system at its Palo Alto office, where Elon Musk examined the hardware and asked questions about the chip's cores, memory layout and cooling, according to NVIDIA. The company said SpaceXAI is assessing Vera for reinforcement learning work and simulation pipelines used in training.

Oracle Cloud Infrastructure, or OCI, appears set to be the biggest early deployment partner among the first recipients. At Oracle's AI Customer Excellence Centre in Santa Clara, Buck demonstrated the system to a team that included Karan Batta, who leads overall product management, and Gary Miller, chief customer and partner success officer.

Buck said Vera addresses the growing amount of CPU work created by AI systems that generate code and execute multi-step tasks rather than returning a simple answer.

"When AI models are posed a question, the answer, often, isn't already prepped and ready to go. The models actually have to generate some Python code to arrive at the correct answer," Buck said.

He added: "That's why we are seeing the demand for CPUs skyrocket."

Cloud push

Oracle signalled plans for a large rollout, saying it intends to deploy hundreds of thousands of Vera CPUs from 2026. That would make it the first cloud provider to commit to the processor at scale.

"OCI plans to deploy hundreds of thousands of NVIDIA Vera CPUs beginning in 2026 because agentic AI demands sustained performance at massive scale," Batta said. "Vera's architecture is purpose-built for high-throughput reasoning workloads, delivering the efficiency, density and footprint OCI needs to power the next generation of enterprise AI."

The commitment matters because cloud providers are likely to determine whether a new server processor becomes a mainstream part of AI infrastructure or remains a niche product used only in specialist systems. NVIDIA already has a strong position in AI accelerators, but CPUs still play a central role in orchestration, memory management, data preparation and software execution around those accelerators.

Vera is also intended to serve as the host processor in NVIDIA's Vera Rubin NVL72 systems, paired with Rubin GPUs through the company's NVLink-C2C interconnect. NVIDIA says this allows the CPU and GPU to share a unified memory architecture, with the CPU managing the control and data-movement tasks needed to keep the GPUs occupied.

That points to a broader strategy in which NVIDIA sells an increasingly complete AI system, combining CPU, GPU, networking, interconnects and rack architecture rather than relying on standalone chips. For AI labs and cloud providers, the appeal is tightly integrated infrastructure from a single supplier.

Oracle executives said they expect customer interest once the systems are available for testing in the company's demonstration centre. "I am really looking forward to the reaction of people who come through here, and working together to get the most from Vera," Miller said.