Nvidia names Anthropic and OpenAI among first users of its Vera chip

At a glance:

• Nvidia unveiled Vera CPU, successor to Grace, featuring 88 Olympus cores and up to 1.2TB/s memory bandwidth.
• Early benchmarks show Vera outperforming Intel Xeon and AMD EPYC on agentic workloads.
• First users include Anthropic, OpenAI, SpaceX, and Oracle, with Oracle set to deploy at scale in H2 2026.

What is Vera and how it differs from Grace

Vera is presented as a ground‑up redesign of Nvidia’s data‑centre CPU line, succeeding the Grace processor that relied on off‑the‑shelf Arm Neoverse cores. The new chip integrates 88 of Nvidia’s proprietary “Olympus” cores, marking a shift from licensed Arm designs to an in‑house architecture. Nvidia says the processor is already in full production.

The architecture is tuned for the era of AI agents—software that plans and executes multi‑step tasks rather than merely responding to a prompt. To support those workloads, Vera couples its cores with up to 1.2 TB/s of memory bandwidth, a figure Nvidia highlights as critical for feeding large model inference and agent orchestration. The chip is also meant to be sold as part of the Vera Rubin platform, pairing the CPU with Nvidia’s GPUs as a matched pair.

• 88 Olympus cores (Nvidia‑designed, replacing Arm Neoverse used in Grace)
• Up to 1.2TB/s memory bandwidth
• Successor to Grace, positioned as a ground‑up redesign

Early performance and benchmark claims

Nvidia cites early independent benchmarks that place Vera ahead of Intel’s Xeon and AMD’s EPYC processors on several performance measures, particularly those that mimic agentic AI workloads. The company wants the comparison to be drawn because it positions Vera as a CPU that can keep up with, or exceed, the incumbent x86 offerings in the data centre.

The benchmarks focus on metrics such as instructions per cycle, memory latency, and throughput for multi‑threaded agent simulations, where Vera’s Olympus cores and high memory bandwidth give it an edge. Nvidia did not disclose the exact benchmark suite or the testing environment, but the claim is reinforced by the company’s own internal testing that reportedly shows similar gains.

• Intel Xeon (current generation)
• AMD EPYC (current generation)

Launch customers and deployment timeline

At Computex in Taipei, Jensen Huang read out a guest list that named Anthropic, OpenAI, SpaceX and Oracle as among the first big users of Vera. The mention of the two leading AI labs underscores Nvidia’s view that the chip is aimed at the compute‑intensive workloads driving the current AI build‑out.

According to Nvidia’s own account, the first Vera units were hand‑delivered in May, ahead of the public announcement in Taipei. Oracle Cloud Infrastructure is slated to be the first hyperscaler to deploy Vera at scale, with broader availability across the other major cloud providers expected in the second half of 2026.

• Anthropic

• OpenAI

• SpaceX

• Oracle

• First units hand‑delivered: May 2025 (according to Nvidia)

• Public announcement: Computex Taipei, June 2025

• Oracle Cloud Infrastructure: first hyperscale deployment (scale)

• Broader availability across other major clouds: second half of 2026

Business implications for Nvidia

For Nvidia, introducing its own data‑centre CPU removes a long‑standing reliance on third‑party suppliers for a core component of AI systems. By bundling Vera with its GPUs in the Vera Rubin platform, the company can sell a matched CPU‑GPU pair, potentially increasing attachment rates and margin on data‑centre sales.

Jensen Huang has repeatedly described Taiwan as the “epicentre” of this effort, highlighting the island’s role in the chip’s development and launch. While the announcement showcases full production and a marquee customer list, Nvidia did not disclose pricing or the exact number of units any of the named labs have committed to, leaving volume shipments to paying customers as the next hurdle to watch.

• Pricing for Vera units
• Commitment volumes from Anthropic, OpenAI, SpaceX, Oracle