Nvidia Launches Vera CPU to Power Agentic AI Orchestration
The Armv9.2-based processor features 88 Olympus cores and massive memory bandwidth to eliminate GPU bottlenecks.
Nvidia has launched the Vera CPU, a custom Armv9.2-based processor designed to orchestrate the next generation of agentic AI. The chip marks a strategic shift toward specialized silicon capable of managing complex tool-calling and long-context state management for AI agents.
Built on the TSMC N3 process using CoWoS-R packaging, the Vera CPU features 88 custom-designed "Olympus" cores and 176 threads utilizing spatial multithreading. The chip contains approximately 227 billion transistors. To handle massive data throughput, Vera supports up to 1.5 TB of SOCAMM LPDDR5X system memory with 1.2 TB/s of memory bandwidth. Connectivity is driven by NVLink-C2C, providing 1.8 TB/s of coherent bandwidth—roughly seven times the speed of PCIe Gen 6.
The Shift to AI Factories
Vera is a central component of the broader Vera Rubin platform, representing Nvidia's transition toward "AI factories." In this model, the CPU evolves from a simple host manager into an active orchestrator for reinforcement learning (RL) and complex reasoning tasks. By replacing generic x86 architectures from Intel and AMD with custom Arm-based silicon, Nvidia is optimizing the CPU-to-GPU ratio to better support the high-frequency state changes required by autonomous AI agents.
Challenging the Datacenter Status Quo
This move represents Nvidia's most aggressive entry into the standalone CPU market. By integrating extreme memory capacity and NVLink bandwidth directly into the processor, Nvidia is removing the traditional communication bottlenecks that have long hindered the efficiency of GPU clusters. This architecture is critical for agentic AI, which requires rapid data movement to execute multi-step reasoning and tool-based interactions in real-time.
Early Deployment and Adoption
The Vera CPU is being deployed both as a standalone product and as part of the Vera Rubin NVL72 platform. Nvidia has already delivered early units to a group of high-profile AI labs and infrastructure providers, including OpenAI, Anthropic, SpaceX, and Oracle. Industry observers will now be watching for performance benchmarks to see if this specialized hardware significantly reduces latency in agentic workflows compared to traditional server CPUs.