Jensen Huang dropped a number on Wednesday's earnings call that made Wall Street sit up: $200 billion in new total addressable market, opened by Nvidia's Vera CPU. After posting $81.6B in quarterly revenue and guiding $91B next quarter, Huang positioned Vera not as a GPU accessory but as a standalone compute platform for agentic AI workloads.

Here's the architectural argument. Traditional cloud CPUs optimize for multi-instance parallelism—lots of cores juggling many apps. Vera flips that: it's designed to process tokens as fast as possible, trading core count for single-thread token throughput. Huang claims this makes it "the world's first CPU purpose-built for agentic AI," where agents spend most of their time running CPU-bound inference and tool execution rather than GPU-heavy training.

Nvidia already reports $20B in standalone Vera CPU sales this year, which is notable because the product only launched in March. Every major hyperscaler and system integrator is signed up as a partner, according to Huang. The bundling strategy with Rubin GPU gives customers a unified memory fabric, reducing data movement overhead between CPU and GPU for agent loops.

The timing matters. AWS just announced a massive Meta contract for its own Trainium CPUs, signaling that cloud providers are serious about homegrown silicon. But Huang counters that Nvidia's advantage isn't just hardware—it's the CUDA ecosystem and the fact that agentic AI pipelines already run on Nvidia stacks. Rebuilding that toolchain for a competitor's CPU carries real engineering cost.

Whether Vera captures that $200B depends on whether agentic AI scales to billions of agents as Huang predicts. Each agent needs a CPU-driven "PC" for tool use, he argues. That's a lot of silicon. For data engineers, the takeaway is architectural: the bottleneck in agentic systems may shift from GPU memory to CPU token processing speed. Vera is Nvidia's bet on that transition.

Source: TechCrunch