NVIDIA shipped its first Vera CPUs directly to Anthropic, OpenAI, SpaceX AI, and Oracle, with Ian Buck, the company’s vice president of hyperscale and high-performance computing, personally overseeing delivery. The company reported the milestone on May 18, 2026, via its official blog.
Vera is NVIDIA’s first in-house-designed CPU, built around 88 custom Olympus cores. The chip delivers 1.2 TB/s of memory bandwidth and, according to NVIDIA, offers roughly 50 percent better per-core performance compared to prior host processors used in its AI systems. Those numbers matter because host CPU performance has historically been a bottleneck in large-scale GPU clusters, where data staging and coordination tasks can limit how efficiently the accelerators operate.
The chip is designed to serve as the host processor for the Vera Rubin NVL72, NVIDIA’s next-generation AI computing platform. It connects to a pair of Rubin GPUs via second-generation NVLink-C2C, NVIDIA’s high-bandwidth, low-latency chip interconnect. NVLink-C2C allows the CPU and GPU to share a unified memory space, which removes a class of data-copy overhead that conventional PCIe-linked server architectures carry. The practical effect is tighter CPU-GPU integration than anything NVIDIA has shipped before.
Until Vera, NVIDIA built its AI platforms around CPUs from Arm (for the Grace series) or Intel and AMD processors in standard server configurations. Vera represents a decision to control the full silicon stack, from the accelerator to the host. The strategic logic mirrors what Apple did when it moved from Intel to its own chips: owning the CPU allows you to optimize memory subsystems and interconnects specifically for the workload, rather than inheriting design choices made for general-purpose computing.
The decision to hand-deliver the first units, rather than ship through standard logistics, signals that these are pre-production or early qualification samples going to partners who will validate the hardware before broader rollout. Anthropic and OpenAI running validation cycles on Vera now means feedback loops that could shape the final production configuration of Vera Rubin systems. That is meaningful influence for the labs and meaningful risk reduction for NVIDIA before it scales manufacturing.
NVIDIA has not disclosed a general availability date for the Vera Rubin NVL72, nor has it published third-party benchmark results for Vera. The performance claims in the announcement are NVIDIA’s own. Independent validation from the labs receiving these units will not be public until those organizations choose to share it, and they rarely do on NVIDIA’s timeline.
For teams currently planning infrastructure procurement for late 2026 or early 2027, the arrival of Vera CPUs at top labs is a signal that Vera Rubin systems are on track and that qualification cycles have begun. Any organization building or budgeting for large-scale NVIDIA GPU clusters should factor Vera Rubin NVL72 into their planning horizon rather than committing solely to current Blackwell-based configurations.
Reported by NVIDIA on the NVIDIA official blog, published May 18, 2026.
