NVIDIA Vera CPU: New Standard for Agentic AI Tasks in Factories

NVIDIA Developer Blog published in May 2026 a piece about the Vera processor, which the company positions as a new standard for computing in agent AI tasks within so-called "AI factories" — dense data centers built specifically for training and inference of models. The article's starting point is the idea that each new wave of AI generates its own scaling law: if pretraining scaled model intelligence through increased datasets, more parameters, and massively parallel GPU systems, then for the next phase — the phase of autonomous agents — a different type of computational core is needed.

How AI Scaling Laws Have Changed

During the pretraining phase, the industry followed a simple formula: more data, more parameters, more GPUs connected into parallel clusters — and the model becomes more capable. This scaling law defined data center architecture for the past decade, where the primary computational load fell on graphics processors, while the central processor played a supporting role in orchestration and data preparation for GPUs. NVIDIA's article describes this era as the first wave, followed by logic specific not to training, but to using models as agents working autonomously over extended periods.

NVIDIA has maintained the practice over recent years of naming its chip generations after famous scientists — a tradition that has already given names to the Grace processor and the Blackwell line of graphics accelerators. Vera continues this tradition for the company's new generation of central processors, and the fact that the Developer Blog dedicates a separate article to the CPU role rather than GPU is itself telling: for a long time, the GPU remained the undisputed hero of NVIDIA's marketing and technical materials, while the processor was viewed as an auxiliary platform component.

What the Vera Processor Brings New

Vera is a processor that NVIDIA is preparing as the successor to the Grace line in its platform architecture, oriented toward CPU-GPU integration for AI tasks. The idea is that agent workloads — with their long-lived sessions, orchestration of multiple parallel tool calls, and need to quickly switch between tasks — place demands on the central processor that differ from classical inference: not only memory throughput to the GPU matters, but also the CPU's ability to efficiently manage numerous simultaneous lightweight processes of agent logic. It is for this workload that NVIDIA formulates a new scaling law, positioning Vera at the center of next-generation "AI factory" architecture.

Key facts:

• Article published in NVIDIA Developer Blog in May 2026
• Processor called Vera and positioned for agent AI tasks
• Term "AI factories" used by NVIDIA for dense data centers for training and inference
• Article thesis: pretraining scaled through data, parameters, and parallel GPU systems; agent stage requires a new approach

What "AI Factory" Means for Agent Tasks

The concept of "AI factory" reflects a shift in metaphor: instead of a data center as a warehouse of computational resources, NVIDIA describes it as a production line, where data and requests enter on one end and agent solutions and actions come out on the other, produced at predictable speed and cost. In such a factory, a processor like Vera takes on the role not of a secondary GPU helper, but of a full-fledged production line element responsible for orchestrating thousands of simultaneous agent sessions.

For data center operators and developers of agent platforms, this means that when planning infrastructure for large-scale agent systems, CPU characteristics of the new generation must be built in as a distinct, not secondary parameter — on par with the number and type of installed graphics accelerators. The shift in focus to the CPU in the context of agent tasks can be viewed as an acknowledgment that further growth in AI infrastructure performance is no longer reducible to simply adding new generations of graphics accelerators: balancing the entire platform, including the processor, memory, and network connections between nodes, becomes no less significant a factor in efficiency than the raw computational power of an individual GPU.