Nvidia Delivers Vera CPUs to Anthropic and OpenAI

Nvidia has started placing its new Vera CPU systems with major AI organisations. PCMag reported that Nvidia vice president Ian Buck "hand-delivered" first Vera systems to Anthropic, OpenAI, SpaceXAI, and Oracle Cloud Infrastructure. Datacenter.news and Yahoo Finance also reported the initial rollouts. Nvidia and Datacenter.news state the processor contains 88 Olympus cores, delivers up to 1.2 TB/s of memory bandwidth, and offers per-core performance Nvidia says is 50% faster under full load than earlier designs used in comparable systems. VentureBeat and Nvidia described Vera Rubin as a seven-chip platform that combines Vera CPUs with Rubin GPUs and other components; VentureBeat quoted CEO Jensen Huang calling the platform "a generational leap." Datacenter.news quoted Anthropic compute lead James Bradbury on interest in how Vera could fit broader infrastructure, while VentureBeat quoted Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman on the platform's potential impact.

The reporting frames Vera as a purpose-built host CPU for what Nvidia describes as "agentic AI" workloads, where models run background tasks, tool calls, orchestration, or reinforcement learning. Industry-pattern observations: CPUs with higher sustained memory bandwidth and many moderate-performance cores are commonly used to host I/O, orchestration, and tool-invocation layers in heterogeneous AI stacks. For practitioners, the key hardware signals to watch are per-core sustained throughput and memory subsystem design, because those determine how well a host CPU can feed accelerators and manage concurrent system services.

Public coverage places this rollout in the broader push toward heterogeneous, co-designed AI racks that pair custom CPUs, GPUs, DPUs, and networking. VentureBeat's reporting highlights Nvidia positioning the Vera Rubin stack as an end-to-end infrastructure play, with Rubin GPUs claimed to deliver up to 10x more inference throughput per watt and one-tenth the cost per token versus recent Blackwell systems, per VentureBeat's coverage of Nvidia statements. Yahoo Finance and PCMag both reported Oracle Cloud Infrastructure as an early hyperscaler partner that intends large-scale Vera deployments, with Yahoo Finance saying Oracle plans to roll out "hundreds of thousands" of Vera CPUs beginning in 2026.

• •Benchmark and telemetry data showing sustained host throughput and system-level token cost comparisons versus Blackwell-derived stacks. These will clarify the asserted efficiency gains.
• •Software and stack integration, including kernel, scheduler, networking, and NVLink-C2C interoperability with Rubin GPUs; VentureBeat notes Vera acts as the host processor for the Vera Rubin NVL72 node where it links to two Rubin GPUs.
• •Cloud availability and pricing details from OCI and other providers. Reports that Oracle intends hyperscale deployment make its rollout cadence a practical leading indicator of broader adoption.

Companies adopting new, co-designed CPU+GPU platforms typically face a multi-quarter phase of integration work: kernel tuning, orchestration changes, and validation of sustained throughput at scale. For practitioners, evaluating the Vera stack will require system-level tests under representative agentic and RL workloads rather than isolated microbenchmarks.

• •PCMag reported Ian Buck said he "hand-delivered the first-ever Nvidia Vera CPUs" to partners.
• •VentureBeat quoted Nvidia CEO Jensen Huang calling Vera Rubin "a generational leap."
• •Datacenter.news quoted Anthropic compute lead James Bradbury expressing interest in how Vera could fit broader infrastructure.

This account synthesizes PCMag, Datacenter.news, Yahoo Finance and VentureBeat reporting on the initial Vera deliveries and the technical claims Nvidia has publicly presented.