Nvidia unveils Vera CPU Arm server chip

Nvidia’s Vera CPU is the company’s latest attempt to control more of the AI server stack itself. The chip is a custom Arm-based server processor built by Nvidia to sit alongside its GPUs inside rack-scale systems for AI training, inference, reinforcement learning and analytics. Nvidia disclosed the processor in more detail at GTC in March and positioned it as the host CPU for its next-generation Vera Rubin platform. The timing matters because Nvidia has said Vera Rubin systems are due from partners in the second half of 2026, putting Vera at the center of the company’s next data-center product cycle. ### So what exactly is Vera? Nvidia said Vera is an 88-core CPU built on custom Arm-compatible “Olympus” cores, rather than a general-purpose off-the-shelf server design. The company describes it as a processor built for “agentic AI” and reinforcement learning workloads, where the CPU handles code execution, tool use, memory management and orchestration around the GPU-heavy parts of the system. (nvidianews.nvidia.com) Nvidia’s product page says Vera is the host CPU in accelerated systems, pairing with Nvidia GPUs to direct data movement and system control. Its technical blog says the chip uses Nvidia Spatial Multithreading and a second-generation Scalable Coherency Fabric. ### Why is Nvidia building its own Arm server CPU now? Nvidia has been moving from selling standalone GPUs toward selling full AI systems. (nvidianews.nvidia.com) The Vera Rubin platform combines the Vera CPU with Rubin GPUs, NVLink switching, networking and other components in a tightly integrated rack design. Nvidia said in January that the Rubin platform was built through “extreme codesign” across six chips, and in March said the platform had expanded to seven chips in full production. (nvidia.com) Arm said Nvidia’s Rubin announcement was a validation of converged AI data centers built with Arm technology at the core. That framing matches Nvidia’s broader push to make the CPU another in-house element of its AI factory architecture rather than relying only on external server processors. ### What performance claims is Nvidia making? (investor.nvidia.com) Nvidia said Vera delivers “twice the efficiency” and is 50% faster than traditional rack-scale CPUs for the targeted workloads it highlighted at launch. In a separate technical post, Nvidia said Vera offers up to 50% faster “agentic sandbox” performance and 1.2 TB/s of memory bandwidth. Those figures come from Nvidia and were presented in the context of reinforcement learning, agentic inference and real-time analytics. (newsroom.arm.com) TechSpot, citing Nvidia’s GTC disclosures, reported that Nvidia also outlined dense rack designs using large numbers of Vera CPUs. Nvidia’s own materials focus less on standalone CPU sales and more on how Vera fits into full accelerated systems. ### How does Vera fit into Vera Rubin? (nvidianews.nvidia.com) Nvidia said Vera is the CPU foundation for Vera Rubin NVL72 systems and related rack configurations. The company’s March newsroom release said the Vera Rubin platform includes Vera CPU racks and Vera Rubin NVL72 GPU racks among the building blocks for AI factories. (techspot.com) Third-party coverage ahead of Nvidia’s earnings, including Stocktwits and TechTimes, linked Vera to expected deployments at Amazon Web Services, Google Cloud and Microsoft later in 2026. Nvidia’s official releases reviewed here say partner availability is expected in the second half of 2026, but they do not, in the materials cited, specifically confirm those three cloud companies for Vera CPU deployments. (nvidianews.nvidia.com) ### What should readers watch next? Nvidia’s next milestone is partner rollout in the second half of 2026 for Vera Rubin systems, according to the company’s March release. Investors and customers will likely look for more detail in Nvidia product updates, customer announcements and future earnings materials on which server makers and cloud providers are taking Vera-based systems first. (techtimes.com) (investor.nvidia.com)