Nvidia Vera CPU vs EPYC 9996: 5,470 SPEC vs 1,459 Est. [2026] - shattered.io

The competition in AI processors intensified this fall as AMD made its 6th Gen EPYC SP7 processors generally available on November 10, 2026, while Nvidia's Vera CPU became available on Oracle Cloud Infrastructure on October 27. Intel's Xeon 6 family remains the incumbent x86 option and serves as the host CPU in Nvidia's DGX Rubin NVL8 systems, creating a three-way battle for dominance in AI agent infrastructure.

AMD has published list prices and SPEC CPU2017 scores for its EPYC 9996 flagship processor and Intel's Xeon 6980P, claiming the EPYC delivers 5,470 SPEC points compared to 2,510 for the Xeon. However, AMD estimates Nvidia's Vera CPU at only 1,459 SPEC points—a figure Nvidia disputes. The chipmaker has published neither official SPEC scores nor pricing for Vera, instead making workload-specific claims of up to 1.8x faster single-threaded performance and 80% faster agentic task completion against x86 processors.

Availability and Deployment

Nvidia positions Vera as "the first CPU built for AI agents," offering it both as a host CPU for Rubin GPUs and as a standalone processor for agent orchestration. The first Vera systems went to Anthropic, OpenAI, Oracle Cloud Infrastructure, and SpaceXAI. AMD's EPYC 9006 Series systems began shipping to OEMs and ODMs by November 10-12, though cloud instances on Microsoft Azure and Oracle Cloud remain "coming soon." Intel maintains its position with Xeon 6, which continues to power Google Cloud TPU 8tx infrastructure and serves as the host in Nvidia's own DGX systems.

Performance Claims

AMD's benchmark shows the EPYC 9996 achieving 5,470 SPEC points in a two-socket configuration, compared to 2,510 for the Xeon 6980P—roughly a 2.2x performance advantage. Against Intel's Xeon 6990E+ with 288 E-cores, the EPYC still leads by 1.79x despite having fewer cores. The per-core performance tells a different story: the Xeon 6990E+ achieves only 5.30 points per core versus EPYC's 10.68, suggesting AMD's architecture delivers more performance per core despite having fewer total cores.

The Vera Question

Nvidia's Vera CPU presents a challenge for direct comparison. The company has not published official SPEC CPU results, instead providing workload-specific claims. AMD's estimate of 1,459 SPEC points for Vera remains unverified, and Nvidia's own claims lack specific test conditions or comparison chips. The Vera CPU uses Nvidia's Olympus cores and is designed specifically for AI agent workloads, suggesting it may perform differently in different scenarios than general-purpose benchmarks can capture.

For general-purpose and high-core-count work today, the EPYC 9996 has the strongest published case with verified benchmarks and shipping systems. For GPU-attached agent infrastructure inside Vera Rubin racks, Nvidia's CPU is the default choice. The true competitive picture will depend on independent tests that have not yet been released, leaving buyers to navigate vendor-chosen workloads that may not reflect real-world performance.