NVIDIA unveils Vera, an 88-core Arm CPU designed for agentic AI workloads

On May 31, 2026, NVIDIA announced Vera, its first standalone CPU positioned for the agentic-AI era. Vera pairs 88 NVIDIA-designed Olympus cores with high-bandwidth LPDDR5X memory and is being positioned as a drop-in alternative to x86 server CPUs across hyperscaler, AI lab, and enterprise workloads. Anthropic, OpenAI, ByteDance, CoreWeave, Oracle, and the New York Stock Exchange are named adopters; Dell, HPE, Lenovo, and Supermicro are the first OEMs taking it to market.

What's new

Vera is the CPU half of the Vera Rubin platform that NVIDIA also moved into full production on the same day, but the May 31 press release calls out Vera as its own product line — sold standalone as well as integrated into Vera Rubin NVL72 racks.

• Architecture. Per NVIDIA: Vera is "featuring 88 Olympus cores, Spatial Multithreading and a LPDDR5X memory subsystem that delivers up to 1.2TB/s of bandwidth." The Olympus cores are NVIDIA's own Arm-based core design.
• Performance positioning. NVIDIA claims "1.8x faster task completion compared with x86 CPUs to drive diverse workloads across industries — including agentic AI, reinforcement learning and data processing."
• Named customers. NVIDIA discloses that "customers exploring the Vera CPU include finance leader NYSE, global AI labs Anthropic, OpenAI and SpaceXAI, and hyperscalers ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure."
• OEM availability. Standalone Vera servers will be offered by "Dell, HPE, Lenovo and Supermicro," alongside ASUS, Compal, Foxconn, GIGABYTE, Pegatron, Quanta Cloud Technology, Wistron, and Wiwynn.

Jensen Huang framed the launch this way: "Vera is the first CPU designed for that future — built to run agentic AI at hyperscale with extraordinary performance, efficiency and programmability."

Context

NVIDIA's prior server CPU, Grace, shipped in volume as part of Grace Hopper and Grace Blackwell modules — primarily tightly coupled to a GPU rather than sold standalone. Vera changes that posture: NVIDIA is now selling an Arm CPU that hyperscalers and enterprises can deploy as a general server CPU, not only as a co-located accelerator host. This is a direct challenge to x86 dominance in the AI-factory tier, where the CPU-side workload is increasingly orchestration, vector pre/post-processing, scheduling, and data movement around long-running agent jobs.

The named-customer list is unusually candid. Both Anthropic and OpenAI publicly evaluating an NVIDIA CPU is a meaningful break from the assumption that frontier labs default to x86 head nodes for their training and inference racks. NYSE in the lineup signals NVIDIA also wants Vera in low-latency finance workloads, not just AI-only servers.

Why it matters

The substantive AI question is whether agentic inference is, in fact, CPU-limited in a way Vera meaningfully addresses. NVIDIA's pitch is that long-running agents — "one prompt can launch a thousand-step journey of reasoning, retrieval, tool use and response generation," as Huang said in the parallel Vera Rubin release — push CPUs hard on orchestration, KV-cache management, and tool-call dispatch. A CPU with 1.2TB/s of memory bandwidth and 88 cores designed to overlap heavy threading is a credible answer if the 1.8x x86 claim holds up on representative agent workloads.

The second-order effect is supply concentration. If Vera lands well, NVIDIA captures more of the AI rack BOM than it does today — GPU plus NVLink fabric plus Ethernet (Spectrum) plus DPU (BlueField) plus CPU. The procurement teams at hyperscalers and frontier labs will read the named-customer list as cover to take the bake-off seriously, which is exactly NVIDIA's intent.