Nvidia is positioning its Vera CPU, an 88-core Arm processor now in full production, around a metric the company rarely leads on: single-threaded performance, which it argues is critical for the “think fast, verify, and orchestrate” workloads of agentic AI.
Announced at GTC in March and shipping in systems from Dell, HPE, Lenovo, and Supermicro in the second half of 2026, Vera uses Nvidia’s custom Olympus cores built on Arm v9.2. Each core employs Nvidia’s Spatial Multithreading technology, running two hardware threads by physically partitioning resources rather than time-slicing, delivering 176 threads across the chip.
Why single-thread matters for agents. Nvidia’s argument is that agentic AI systems, which plan tasks, execute code, query databases, call tools, and validate results, are fundamentally different from traditional batch AI workloads. They require fast, low-latency responses for each step in a multi-step reasoning chain, which depends heavily on per-core throughput rather than raw core count. Benchmarks published by Phoronix show Vera outperforming Intel and AMD on AI-adjacent CPU workloads, particularly in memory-bandwidth-sensitive tasks.
Vera delivers 2.4x the memory bandwidth of its predecessor Grace, with 1.5 TB of LPDDR5X memory and a second-generation Scalable Coherency Fabric that sustains over 90% of peak bandwidth under load. Nvidia says a single rack integrating 256 liquid-cooled Vera CPUs can sustain more than 22,500 concurrent agent environments.
Ecosystem adoption. Early adopters include Meta, Oracle Cloud Infrastructure, Alibaba Cloud, and CoreWeave. Research institutions such as Los Alamos National Laboratory, Lawrence Berkeley National Lab, and the Texas Advanced Computing Center are also planning deployments.
The Vera CPU anchors Nvidia’s six-component Vera Rubin platform, which powers the NVL72 rack rated at 14.4 exaFLOPS of FP4 performance. Bank of America projects the data center CPU market will double from $27 billion to $60 billion by 2030, driven by agentic orchestration demands.
Source: NVIDIA Newsroom, Phoronix, WinBuzzer