NVIDIA ships Vera CPUs to labs

Nvidia said on May 18 that it had started delivering its new Vera CPU systems to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI, moving a product it unveiled in March from launch-stage claims to early deployment. The company described Vera as its first CPU built specifically for “the age of agentic AI,” a category Nvidia has used to describe workloads that plan tasks, call tools, run code and validate results. Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, handed over the first systems, according to an Nvidia blog post. Nvidia did not disclose shipment volumes, pricing or production timelines for broader availability. ### Which part of the system is Vera supposed to handle? Nvidia said on March 16 that Vera is designed for the host-compute side of AI systems rather than as a replacement for its accelerators. The company said the chip is aimed at reinforcement learning, agentic inference, analytics, sandboxing and data-heavy orchestration tasks that sit around GPU clusters. (blogs.nvidia.com) The Nvidia developer blog said Vera uses 88 custom “Olympus” cores and delivers 1.2 terabytes per second of memory bandwidth. Nvidia also said the CPU connects to Rubin GPUs through NVLink-C2C, which the company’s product page lists at 1.8 terabytes per second of coherent bandwidth. (investor.nvidia.com) ### Why are Anthropic, OpenAI, Oracle and SpaceXAI the first named recipients? Nvidia named Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI as the first organizations receiving Vera systems. The company’s May 18 blog post framed those deliveries as early production deployments at labs and infrastructure operators already running large AI workloads. (developer.nvidia.com) Anthropic was identified in Nvidia’s post through James Bradbury, the company’s head of compute. Oracle was named through Oracle Cloud Infrastructure, giving Nvidia a cloud distribution partner as well as model labs among the first recipients. Nvidia did not publish customer quotes from OpenAI or SpaceXAI in the materials surfaced here. ### How does Vera relate to Grace and Rubin? (blogs.nvidia.com) Nvidia launched Vera alongside the Rubin platform at GTC on March 16. The company said the broader Vera Rubin system combines new CPUs, GPUs, networking and rack components for what it calls AI factories. The company’s technical materials present Vera as the CPU paired with Rubin GPUs in that next-generation stack. (blogs.nvidia.com) Nvidia has not described Vera in the official materials reviewed here as a simple one-for-one replacement for Grace, but it has presented Vera as the CPU element of its next platform after Grace-based systems. That is an inference from Nvidia’s March launch materials and product positioning. (nvidianews.nvidia.com) ### What does Nvidia say the performance gain is? Nvidia said at launch that Vera delivers results “twice the efficiency” and “50% faster” than traditional rack-scale CPUs. In a separate developer post, the company said Vera provides up to 50% faster agentic sandbox performance under full load. Nvidia did not specify in the materials reviewed here which third-party CPUs were used as the baseline in those headline comparisons. (investor.nvidia.com) The product page also says Vera is intended to work as part of a unified memory system with Rubin GPUs, allowing CPU and GPU resources to share data for large AI and high-performance computing workloads. That language places the chip inside Nvidia’s larger effort to sell integrated systems rather than standalone accelerators. ### What comes next, and what is still not public? (investor.nvidia.com) Nvidia’s next scheduled public update is its quarterly earnings report on May 20, when investors are expected to press the company on demand durability and its broader AI systems strategy, according to Reuters. The company has not yet disclosed how many Vera systems it plans to ship in 2026, when wider customer rollouts begin, or whether the first deliveries are pilots or revenue-generating production deployments. (nvidia.com) (blogs.nvidia.com)