Key Highlights
- Nvidia’s latest Vera CPU has transitioned into production, with initial systems reaching leading AI organizations such as Anthropic, OpenAI, Oracle Cloud Infrastructure, and xAI.
- Ian Buck, Nvidia’s VP, conducted personal deliveries of these systems, beginning with Anthropic’s headquarters in San Francisco.
- The Vera processor boasts 88 specialized Olympus cores, delivers 1.2 TB/s memory bandwidth, and achieves 50% superior per-core performance versus conventional CPUs.
- Oracle Cloud Infrastructure will deploy hundreds of thousands of Vera processors starting in 2026, becoming the inaugural hyperscale cloud provider to implement Vera.
- xAI is conducting Vera evaluations for reinforcement learning applications, with Elon Musk receiving a detailed system walkthrough.
Nvidia (NVDA) has successfully transitioned its Vera CPU from development to active manufacturing, with initial deployments reaching several prominent artificial intelligence organizations.
Ian Buck, Nvidia’s Vice President of Hyperscale and HPC, conducted in-person deliveries of inaugural units to Anthropic, OpenAI, Oracle Cloud Infrastructure, and xAI during a two-day period last week.
The delivery journey commenced at Anthropic’s South of Market headquarters in San Francisco, where James Bradbury, the company’s compute division head, accepted the hardware. Buck accompanied the delivery with an exposed Vera CPU motherboard to demonstrate the processor’s distinctive features.
“Expanding computational capacity represents a critical catalyst for model advancement,” Bradbury stated. “Vera’s arrival as a valuable ecosystem component for agentic workload solutions is particularly encouraging.”
The subsequent delivery proceeded to OpenAI’s Mission Bay campus, where Sachin Katti, responsible for compute infrastructure, met Buck on an exterior terrace. During their discussion, Buck used a screwdriver to open the chassis and display the system’s internal components.
Musk Receives Detailed Demonstration
The day’s concluding delivery arrived at xAI’s Palo Alto location. Nvidia’s delegation provided Elon Musk with a comprehensive architecture overview. Musk inquired about core configuration, memory organization, and thermal management.
xAI is assessing Vera for reinforcement learning applications and agent-based simulation workflows within its training infrastructure.
The following Monday, deliveries continued at Oracle’s AI Customer Excellence Center in Santa Clara, where OCI product management and customer success teams examined the unpackaged system.
“OCI will begin deploying hundreds of thousands of Nvidia Vera CPUs in 2026 because agentic AI requires consistent performance across enormous scale,” stated Karan Batta, who oversees comprehensive product management at OCI.
OCI has become the first cloud infrastructure provider to commit to hyperscale Vera implementation.
Understanding Vera’s Purpose
Vera represents Nvidia’s inaugural custom CPU design, engineered explicitly for agentic AI workloads — scenarios where models execute actions, run code, invoke tools, and maintain extended context state rather than simply providing responses.
While GPUs manage intensive computation, the surrounding orchestration infrastructure — encompassing tool invocations, data transfers, sandboxing, and retrieval operations — requires CPU processing. Vera was designed specifically for these tasks.
The processor incorporates 88 custom Olympus cores, provides 1.2 TB/s memory bandwidth, and achieves 50% enhanced per-core performance under maximum load compared to conventional architectures.
“When AI models receive questions, appropriate responses aren’t always pre-generated,” Buck explained during the Oracle visit. “Models must frequently create Python code to determine correct answers. This explains the surging CPU demand we’re witnessing.”
Vera additionally functions as the host processor in the Vera Rubin NVL72 configuration, connecting with Rubin GPUs through second-generation NVLink-C2C in a unified memory framework. Nvidia claims this arrangement operates at twice the energy efficiency of conventional infrastructure.
Jensen Huang initially unveiled Vera at GTC San Jose in March, characterizing it as Nvidia’s upcoming multi-billion dollar business segment.
