Nvidia Says Its Vera CPU Runs AI Agents Faster and Denser Than x86 Rivals

The pitch goes to the part of the AI build-out that sits outside the GPU. Nvidia argues that agents spend much of a task off the accelerator: searching for information, querying databases, writing and running code, and taking actions. Each of those steps runs on a CPU. Because AI data centres are limited by how much power they can draw, the company says operators need processors that both finish each agent's steps quickly and fit as many agents as possible into every watt. Today, Nvidia says, providers often mix fast x86 parts with dense, efficient ones to strike that balance.

In its 100-megawatt comparison, Nvidia said Vera supported 1.6 times as many agents per watt as the fastest x86 CPU while matching or beating it on agent speed, and delivered twice the agent speed of the most energy-efficient x86 CPU. The blog post does not name the x86 processors used, describe the workloads, or set out the assumptions behind the model, so the figures cannot be checked independently from what Nvidia has published.

The partner results are more specific. Perplexity, the AI search company, said agentic coding tasks ran 1.5 times faster on Vera than on traditional CPUs, and Nvidia said Perplexity plans to use Vera in an upcoming production system. Inference provider DeepInfra reported up to 2.2 times the speed of the best x86 processor on agent orchestration for its production workload, the fastest of the four architectures it tested.

Analytics database company ClickHouse said its standard software build, with no tuning for Vera, scaled almost linearly across 176 cores and took first place on the ClickBench hardware leaderboard. DuckLabs reported about 1.5 times the median performance of an Intel Xeon 6 baseline on the TPC-H SF1000 database benchmark. Daytona, which runs sandboxes for AI agents, brought 1,000 sandboxes online in 10.8 seconds on Vera against 88.4 seconds on x86, and completed 12.7 jobs per second with 2,000 sandboxes running at once. Modal reported up to twice the sandbox throughput per core of comparable x86 hardware, with the gap widening under heavy load.

Vera is Nvidia's first data centre CPU built around its own cores, rather than licensed Arm designs. It has 88 custom Olympus cores, each able to run two tasks, with LPDDR5X memory delivering up to 1.2 terabytes per second of bandwidth. Nvidia said in March that the chip had entered full production and that, for the first time, it would sell the processor on its own as well as paired with GPUs in its Vera Rubin systems. At the time, chief executive Jensen Huang said: "The CPU is no longer simply supporting the model; it's driving it."

That standalone strategy puts Nvidia in direct competition with Intel's Xeon and AMD's Epyc lines in general-purpose servers, a market the two x86 makers have long dominated. Nvidia said in March that standalone Vera servers would ship in the second half of 2026 from manufacturers including Dell Technologies, HPE, Lenovo and Supermicro, with CoreWeave planned as the first cloud to offer it on its own. Meta has also signed a deal covering deliveries of Vera CPUs.

The Thursday post gives no new shipping date or pricing. Perplexity's planned production deployment is the next public test of whether the gains Nvidia describes hold up outside benchmark conditions.