Nvidia is positioning its next chip platform, Vera Rubin, around a metric it calls "intelligence per dollar" — the value an AI system delivers relative to what it costs to run.
According to Nvidia's own blog, Vera Rubin is designed to maximize that intelligence per dollar specifically for "post-training" workloads, which the company frames as a key metric for so-called agentic AI — AI systems that carry out multi-step tasks with a degree of autonomy. Nvidia attributes the gains to what it describes as "extreme codesign," a tight matching of hardware and software that it says yields the lowest cost per token. In AI systems, a token is a small chunk of text, and cost per token is a common way to measure how expensive it is to generate output at scale.
The platform pairs a central processor Nvidia calls Vera with its Rubin design. That CPU is already moving toward real-world trials: HPCwire reports that Los Alamos National Laboratory is set to begin testing the Nvidia Vera CPU for its Mission and Vision supercomputers — a signal that the technology is being evaluated for high-end scientific and government computing, not just commercial data centers.
The sources here are largely drawn from Nvidia's own materials, so the performance framing reflects the company's claims rather than independent benchmarks.
Why it matters: as AI shifts from one-off answers to agents that run many steps, the cost of each token adds up fast — so a chip that lowers that cost could shape how affordable, and how widespread, the next generation of AI becomes.