NVIDIA ships Vera to OpenAI and Anthropic as agents get their own CPU

NVIDIA has delivered the first Vera CPU systems to some of the most important actors in the AI race. According to the company, the first systems arrived at Anthropic in San Francisco, OpenAI in Mission Bay and SpaceXAI in Palo Alto on Friday, followed by a delivery to Oracle Cloud Infrastructure in Santa Clara on Monday.

That may sound narrow. It is not. Vera is NVIDIA’s first custom CPU, built for agentic AI. The point is simple: AI agents do not run on GPUs alone. They launch sandboxes, make tool calls, retrieve context, run code, steer workflows and keep state over time. Much of that is CPU work.

When the largest model labs receive the hardware first, it says something about where AI infrastructure is heading. The capacity fight is no longer only about access to enough H100, Blackwell or Rubin accelerators. It is about the full system that moves agents from demo to production.

NVIDIA says Vera has 88 custom Olympus cores, 1.2 TB/s of memory bandwidth and 50 percent faster per-core performance under full load. The CPU is meant both for standalone CPU systems and as the host processor in Vera Rubin NVL72, where it connects to Rubin GPUs through second-generation NVLink-C2C. NVIDIA says Vera and Rubin use a unified memory architecture to keep accelerated compute better utilized.

For CIOs and technology leaders, this is the practical part: agent cost is not just price per token. It becomes price per workflow. Every sandbox, test run, internal API call, file analysis and long-context operation uses infrastructure around the model. CPU, memory, networking, logging and orchestration decide whether the agent becomes fast, expensive, slow or risky.

From model choice to system choice

NVIDIA frames Vera as a new CPU moment for the AI factory. Ian Buck, NVIDIA’s vice president of hyperscale and high-performance computing, says agentic AI creates new demand because models are moving from answering to acting.

That is also the right framing for enterprises. Many AI strategies are still written as if the central choice is the model name: GPT, Claude, Gemini, Llama, Mistral or DeepSeek. That phase is too simple. Once agents are allowed to do things, the company is choosing a system: where code runs, which tools the agent may use, where logs go, how secrets and permissions are isolated, and who pays when an agent loop runs too long.

Vera points directly into that discussion. NVIDIA says the CPU is designed for orchestration, tool-calling, reinforcement learning workloads, data analytics, agent sandboxing and long-context state management. That is not decoration around the model. It is the control surface for AI in production.

The Anthropic quote in NVIDIA’s post is sober. James Bradbury, Anthropic’s head of compute, says scaling compute is an important accelerant for model growth, and that Vera may become part of the ecosystem for agentic workloads. OpenAI is mentioned through a delivery to compute infrastructure head Sachin Katti. SpaceXAI is evaluating Vera for reinforcement learning and agent-based simulation pipelines.

The Oracle part may matter most for enterprise buyers. OCI says it plans to deploy hundreds of thousands of NVIDIA Vera CPUs beginning in 2026 because agentic AI demands sustained performance at massive scale. NVIDIA describes Oracle as the first cloud provider to deploy Vera at hyperscale.

That means Vera is not just lab hardware. It is moving into commercial cloud infrastructure. The decision is therefore moving closer to procurement, architecture and vendor governance.

What leaders should take from this

First, capacity. If AI agents are going to perform real work in software, finance, customer operations, security or analysis, the company needs capacity that can handle parallel workflows. A model contract and a clean user interface are not enough.

Second, dependency. When OpenAI, Anthropic, SpaceXAI and Oracle receive early access to new NVIDIA infrastructure, the supply chain gets even tighter. Model provider, chip provider and cloud provider are shaping the same operating layer. For a CIO, contracts must cover more than model pricing. They must cover capacity, portability, data location, auditability, exit options and what happens when one part of the stack becomes scarce or expensive.

Third, security. Agentic AI makes the CPU layer more sensitive because much of the action happens there: code runs, files are opened, APIs are called and tools receive permissions. If the agent platform lacks clear boundaries for sandboxing, secrets, network access and logging, faster infrastructure simply means faster failure.

Fourth, measurement. Leaders should stop asking only how good the model is. They should ask what a complete agent workflow costs, how often it fails, where it stops, how much human control it requires, and which parts of the infrastructure become bottlenecks. The Vera news makes that question more concrete.

This is why the story deserves main placement. NVIDIA is not just delivering a component. It is showing that agentic AI is becoming a dedicated operating architecture. When that architecture first reaches OpenAI, Anthropic, SpaceXAI and Oracle, leaders should read it as an early warning: AI budgets are moving from experimentation into production machinery.

Sources and media

• Primary source: NVIDIA Blog, “Vera Arrives: NVIDIA’s First CPU Built for Agents Lands at Top AI Labs”, published May 18, 2026. https://blogs.nvidia.com/blog/vera-cpu-delivery/
• NVIDIA states in the article that Vera systems were delivered to Anthropic, OpenAI, SpaceXAI and Oracle Cloud Infrastructure, and that OCI plans to deploy hundreds of thousands of Vera CPUs beginning in 2026.
• Official image/video material on the NVIDIA page was used only as source context. Thumbnail: OpenAI Image 2 / hogby.ai.