Nvidia Hand-Delivers First Vera CPUs to OpenAI, Anthropic, and SpaceXAI

Nvidia Steps Into the CPU Arena

Nvidia has officially shipped the first units of its long-anticipated Vera CPU, marking the company's formal entry into the data centre processor market. In a strikingly personal gesture, Nvidia's Vice President of Hyperscale and High-Performance Computing, Ian Buck, hand-delivered the inaugural chips to leading AI organisations over the weekend. Anthropic in San Francisco, OpenAI in Mission Bay, and SpaceXAI in Palo Alto each received their units on Saturday, 16 May, with Oracle Cloud Infrastructure in Santa Clara following on Monday.

Elon Musk personally met with Buck to receive a Vera CPU for SpaceXAI, the AI division formed after SpaceX absorbed xAI earlier this year. Oracle has signalled enormous ambition, indicating plans to deploy hundreds of thousands of the chips in 2026.

Built for the Age of Agentic AI

The Vera CPU is described by Nvidia as its first fully custom data centre processor, purpose-built for what the company calls the age of agentic AI. Under the hood, it features 88 Olympus cores with Spatial Multithreading, enabling 176 total threads. The design delivers 1.2 TB/s of memory bandwidth through a 1024-bit LPDDR5X interface — roughly three times the per-core bandwidth of competing Intel and AMD designs.

According to Nvidia's developer documentation, the chip offers up to 50% faster sandbox performance compared to rival platforms. It targets the unique demands of agentic AI workloads, which require high single-core performance and deterministic execution with minimal tail latencies. At CES 2026, Nvidia claimed the broader Rubin platform would deliver inference at one-tenth the token cost of prior-generation Blackwell systems.

Full-Stack Ambition

These early deliveries arrive ahead of broader production shipments expected in the second half of 2026, when cloud providers including CoreWeave plan to deploy Vera CPU racks at scale. The chip is designed to pair with Nvidia's Rubin GPUs in the NVL72 server architecture, aggregating 72 GPUs connected by NVLink 6.

The move positions Nvidia in direct competition with AMD and Intel in the data centre CPU market, leveraging a custom Arm v9.2 architecture rather than off-the-shelf cores. As Buck noted at GTC 2026, the world needs a really fast CPU that can generate training data on the fly so the GPU never goes idle. With Vera, Nvidia is no longer just the GPU king — it is becoming a full-stack AI infrastructure provider.