Nvidia Vera CPU Hits Oracle Cloud as AMD Counters With 2.24x Claim

Nvidia's Vera CPU became a rentable product on October 27, 2026, when Nvidia and Oracle declared the processor and the Vera Rubin NVL72 rack-scale system generally available on Oracle Cloud Infrastructure. The launch ignited a benchmark dispute: Nvidia claims 1.8x faster single-threaded performance, while AMD's own testing reports 2.24x platform-level performance over a Vera-based platform. No independent lab has confirmed either figure.

Founder and CEO Jensen Huang framed the launch as production, not slides. "We are bringing Vera into production on OCI to support these workloads at scale," Huang said in the October 27 announcement. OpenAI is among the early-access customers on the platform.

What does the Vera CPU actually contain?

Nvidia unveiled Vera on May 31, 2026 as a host processor for AI factories, built for tool calls, orchestration and sandboxed code execution by AI agents. Published specifications include:

• A monolithic 88-core die with unified cache and the NVIDIA Scalable Coherency Fabric; Nvidia says this cuts peak loaded latency by 40% versus x86 CPUs.
• LPDDR5X memory delivering up to 1.2 TB/s total bandwidth and 14 GB/s per core — claimed at more than 3x per-core bandwidth of traditional data center CPUs at under half the power.
• Capacity up to 1.5 TB via detachable SOCAMM modules; the second-generation coherency fabric adds 3.4 TB/s of bisection bandwidth.
• Up to 256 Vera CPUs per standalone rack.

How do the benchmark claims conflict?

Nvidia's headline 1.8x single-thread figure carries a caveat: the published wording names no baseline chip. A companion claim says 85% of reinforcement learning evaluations finish inside one training window versus 45% on a baseline CPU. Nvidia's own Vera Rubin platform page separately cites 50% higher single-core performance — two different multipliers for the same trait, likely measured against different baselines.

AMD answered on November 3 with 1.2x per-core and 2.24x platform-level performance in its own testing against a Vera-based platform, without publishing the system configuration behind the larger number. The named Venice part, EPYC 9996, carries 256 cores and 512 threads. A per-core edge of 1.2x cannot alone yield 2.24x at platform level; core count and sockets must close the gap, making AMD's a throughput argument against Nvidia's single-thread case. Both may be measuring different things truthfully.

Independent data is thin. Tom's Hardware benchmarked a standalone Vera system against Intel's Granite Rapids-AP and AMD's EPYC Turin, with one summary reporting Vera matching or exceeding those systems in selected workloads. Turin predates Venice, so the result says little about the sixth-generation part now shipping. AMD said on its Q3 2026 earnings call that Venice revenue shipments began in the quarter, with first OEM systems due in Q4. Vera systems from Cisco, Dell, HPE, Lenovo and Supermicro are slated for the second half of 2026 — meaning the first credible head-to-head tests should follow in the same window.

Who is deploying, and at what price?

Oracle anchors the Vera side, committing to multi-gigawatt Vera Rubin superclusters. Meta intends to expand Vera Rubin use alongside Spectrum-X and BlueField-4 networking, and Nvidia said it will invest $3 billion in Meta's equity offering. Google Cloud plans up to 960,000 Rubin GPUs across multiple sites, with A5X instances reportedly reaching general availability in November.

AMD's rack-level answer, Helios, pairs 72 Instinct MI455X GPUs with 18 EPYC Venice CPUs. On November 3, AMD said Helios systems run workloads at Anthropic, Meta, Microsoft, OpenAI and OCI, and claims up to 30% more inference tokens per dollar — another unverified vendor figure. Its Helion blueprint, introduced October 12, supports CPU-only racks, letting AMD compete for agent serving without a single GPU.

Intel stays in the budget fight through the GPU D270, an inference card with 160 GB of LPDDR5X (partners can build up to 480 GB), sampling with customers. Intel claims single-card support for 300-billion-parameter models and over 1 trillion across eight GPUs.

Six products across the three vendors — Vera CPU, NVL72, HGX Rubin NVL8, EPYC 9006, Helios and the D270 — carry zero published prices as of November 5. AMD reported record Q3 Data Center revenue of $8.5 billion, up 97% year over year. Nvidia reports fiscal Q3 2027 results on November 17 at 2 p.m. PT, the first realistic chance to put a number on Vera CPU revenue.