Nvidia's Vera Rubin claims quarter the GPUs of Blackwell to train

Nvidia has a new metric it wants you to price on, and the framing is worth taking seriously even if the specifics are self-reported. In a July 17 post on its own blog, the company pitched its upcoming Vera Rubin platform as the way to maximize 'intelligence per dollar' for agentic post-training, the loop where models are continuously refined by reinforcement learning after they leave pretraining.

The load-bearing claim is that Vera Rubin trains the largest models with one-fourth the GPUs of the Blackwell generation. Nvidia's argument is that agentic AI turns post-training from a one-time finish into a continuous cost, because the tools an agent uses change week to week, edge cases surface in production, and each deployment brings its own codebase and environment. Every improvement in cost per token, in that framing, compounds into cheaper intelligence overall.

To make the pitch concrete, Nvidia points to Nemotron 3 Ultra, an open-weight 550-billion-parameter mixture-of-experts model that scored 71.7% on SWE-bench Verified using a fully disclosed post-training recipe on NeMo RL. Customer proof points are grafted on. Prime Intellect says Vera CPUs deliver on average 30% greater throughput per CPU than x86 alternatives on its RL sandbox workloads, and Perplexity says its RDMA-based weight-transfer engine can sync trillion-parameter models in under two seconds between training and inference nodes, with the resulting Qwen3 235B model served on GB200 NVL72.

The honest caveat is that these are Nvidia's own numbers and a small set of partner testimonials, on a platform that has not shipped. The post gives no dollar figures for either cost per token or intelligence per dollar, no ship dates for Vera Rubin, and no independent MoE training comparison. Take the four-times GPU-efficiency figure as a target the company is committing to, not a settled benchmark.

The framing itself is what to watch. If continuous RL post-training really becomes the central compute pattern of the agentic era, buyers pricing infrastructure on peak FLOPs or raw GPU count are negotiating the wrong number. Teams like Prime Intellect, Perplexity and Together AI already run the loop that way; the rest of the field is being invited to catch up.