NVIDIA markets Vera CPU for agents

NVIDIA used its March 16 GTC event to argue that the next bottleneck in AI is not only the GPU, but the CPU that runs code, tools and data workflows around the model. Jensen Huang, the company’s chief executive, said the new Arm-based Vera CPU is designed for “agentic AI and reinforcement learning,” a category NVIDIA defines as systems that plan tasks, call tools, run code, interact with data and validate results. Huang framed that change in unusually direct terms. “The CPU is no longer simply supporting the model; it’s driving it,” he said in NVIDIA’s March 16 launch release for Vera. NVIDIA said Vera delivers results with twice the efficiency and 50% faster than traditional rack-scale CPUs, and pitched the chip as part of a broader “AI factory” buildout. (nvidianews.nvidia.com) NVIDIA’s public materials tie Vera to a larger rack-scale system rather than a standalone server CPU story. The company’s GTC keynote described AI factories as a new platform category, while its Rubin and Vera product pages present CPUs, GPUs, networking and storage as a co-designed stack for AI reasoning and deployment. ### Why is NVIDIA talking about CPUs in an AI launch? (nvidianews.nvidia.com) NVIDIA’s March 16 announcement said agentic AI shifts more work onto the infrastructure around the model, including task planning, tool execution, code running and result checking. Vera is aimed at that layer: NVIDIA said the CPU will handle compilers, runtime engines, analytics pipelines, agentic tooling and orchestration services. (nvidia.com) The Vera product page makes the same case in plainer system terms. NVIDIA said Vera acts as the host CPU in accelerated systems, directing data movement, managing memory and orchestrating system control so GPU-heavy AI pipelines keep running at speed. ### What exactly is Vera? NVIDIA said Vera has 88 custom Olympus Arm cores and full Armv9.2 compatibility. (nvidianews.nvidia.com) The company said the chip is its first CPU to support FP8 precision and that it doubles the performance of its predecessor while targeting high energy efficiency. NVIDIA’s technical blog added more detail on the chip’s role in AI factories. (nvidia.com) The company said Vera offers 1.2 terabytes per second of memory bandwidth and up to 50% faster “agentic sandbox” performance for reinforcement learning post-training, agentic inference and analytics workloads. ### How does Vera fit with Blackwell and Rubin systems? NVIDIA is selling Vera as one component in a rack-scale architecture. Its Vera CPU rack page says the system is built on the MGX modular design and is meant for “modern AI factories,” where CPUs execute the code, tools and data workflows that drive agentic systems. The March 16 launch release said a Vera CPU rack integrates 256 liquid-cooled Vera CPUs and can sustain more than 22,500 concurrent CPU environments, each running independently at full performance. (developer.nvidia.com) NVIDIA said those racks are intended to scale agentic tools and instances into the tens of thousands. (nvidia.com) NVIDIA’s Rubin platform materials extend that pitch to the full stack. The company said the Vera Rubin platform combines the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch, with systems configured for pretraining, post-training, test-time scaling and agentic inference. (nvidianews.nvidia.com) ### Who is signed up to use it? NVIDIA said customers collaborating to deploy Vera include Alibaba Cloud, ByteDance, Meta, Oracle Cloud Infrastructure, CoreWeave, Lambda, Nebius and Nscale. It also said manufacturing partners adopting the CPU include Dell Technologies, HPE, Lenovo and Supermicro, alongside Taiwanese server makers and integrators. (nvidianews.nvidia.com) CoreWeave gave one of the clearest timelines attached to the broader platform. NVIDIA said CoreWeave will begin integrating Rubin-based systems into its AI cloud in the second half of 2026, giving customers an early route to the CPU-plus-GPU stack NVIDIA is promoting. ### What is NVIDIA trying to get customers to buy? (nvidianews.nvidia.com) NVIDIA’s product pages and launch materials describe Vera less as a replacement for general-purpose enterprise CPUs than as the control layer inside an integrated AI system. The company’s wording centers on paired deployment with NVIDIA GPUs, rack-scale interconnects and factory-style infrastructure rather than on drop-in competition with conventional x86 server stacks. (nvidianews.nvidia.com) That is an inference from NVIDIA’s packaging and product descriptions, not a direct company quote. NVIDIA’s next concrete milestone is commercial rollout across the Rubin platform. CoreWeave said Rubin-based systems will start joining its cloud in the second half of 2026, and NVIDIA said Vera collaborators already include Meta, Oracle Cloud Infrastructure and Alibaba Cloud. (nvidianews.nvidia.com) (nvidia.com)