NVIDIA Vera CPU: A New Standard for Agentic AI Workloads in AI Factories

In May 2026, the NVIDIA Developer Blog published a piece on the Vera processor, which the company positions as a new computing standard for agentic AI workloads in so-called "AI factories": dense data centers built specifically for training and running models. The article's starting point is that each new wave of AI produces its own scaling law. Pretraining scaled model intelligence by growing datasets, parameter counts, and massively parallel GPU systems, but the next stage, that of autonomous agents, needs a different kind of compute core.

How AI scaling laws have changed

During pretraining, the industry followed a simple formula: more data, more parameters, more GPUs joined into parallel clusters, and the model becomes more capable. This scaling law shaped data center architecture over the past decade, with GPUs carrying most of the compute load while the CPU played a supporting role, orchestrating work and preparing data for the GPUs. NVIDIA's article describes that era as the first wave. It was followed by a logic specific not to training but to using models as agents that run autonomously for extended periods.

Over the past several years, NVIDIA has named its chip generations after famous scientists, a tradition that has already given us the Grace processor and the Blackwell line of GPU accelerators. Vera continues that tradition for the company's next generation of CPUs. The fact that the Developer Blog devotes a separate piece to the role of the CPU rather than the GPU is telling in itself: for a long time the GPU was the undisputed hero of NVIDIA's marketing and technical material, while the processor was treated as a supporting component of the platform.

What the Vera processor brings

Vera is a processor that NVIDIA is preparing as the successor to the Grace line in its platform architecture, which pairs CPUs with GPUs for AI workloads. The idea is that agentic workloads, with their long-lived sessions, orchestration of many parallel tool calls, and need to switch quickly between tasks, place demands on the CPU that differ from classic inference. Memory bandwidth for the GPU matters, but so does the CPU's ability to efficiently manage many simultaneous lightweight processes of agent logic. It is for this load that NVIDIA formulates its new scaling law, placing Vera at the center of the architecture of next-generation "AI factories."

Key facts:

• The article was published on the NVIDIA Developer Blog in May 2026
• The processor is called Vera and is positioned for agentic AI workloads
• NVIDIA uses the term "AI factories" for dense data centers built for training and inference
• The article's thesis: pretraining scaled through data, parameters, and parallel GPU systems; the agentic stage requires a new approach

What an "AI factory" means for agentic workloads

The concept of the "AI factory" reflects a change of metaphor. Instead of a data center as a warehouse of compute, NVIDIA describes an assembly line: data and requests go in, and decisions and agent actions come out, produced at predictable speed and cost. In such a factory, a processor like Vera is not a secondary helper to the GPU but a full element of the production line, responsible for orchestrating thousands of simultaneous agent sessions.

For data center operators and developers of agentic platforms, this means that when planning infrastructure for large-scale agentic systems, next-generation CPU characteristics should be treated as a separate, primary parameter, on par with the number and type of installed GPU accelerators. The shift of focus to the CPU in agentic workloads can be read as an acknowledgment that further growth in AI infrastructure performance no longer comes down to simply adding new generations of GPU accelerators. Balancing the whole platform, including the processor, memory, and network links between nodes, is becoming as important to efficiency as the raw compute power of an individual GPU.