The CPU Is No Longer Scaffolding

AI data centers are no longer GPU-only buildings with CPUs as scaffolding. The CPU has become a co-lead, and Nvidia is now selling it that way.

Wired's reporting on Nvidia's Vera Rubin preview, its next-generation AI superchip platform and the successor to Grace Blackwell, pins the shift in a single ratio: 36 Vera CPUs to 72 Rubin GPUs inside one liquid-cooled NVL72 rack. That is not a chip-by-chip spec. It is a strategic declaration. In the Grace Blackwell era, the CPU was bundled silicon, the helpful neighbor of the GPU. With Vera Rubin, Nvidia is selling the CPU as a named product line, reportedly pitched to Chinese customers as soon as August, and the company's own road map language has caught up. Ian Buck, the architect behind CUDA, told journalists Nvidia is now set to "crank out new architectures, not just GPUs but CPUs."

The demand driver is the workload. As AI systems move from answering prompts to planning, calling tools, and running multi-step tasks, the CPU's job of orchestrating data and networking grows in lockstep with the GPU's job of training and inference. A rack optimized for agentic AI is no longer a GPU with support silicon around it. It is a CPU-GPU pair where the orchestration half is strategic, not incidental.

Nvidia is moving now, previewing before AMD's Thursday product event, because the GPU-only pitch no longer matches the workload mix. The CPU is no longer scaffolding.