Nvidia Pitches Vera Rubin as a Post-Training Machine, Claiming Ten Times Lower Cost Per Token

Nvidia is repositioning its forthcoming Vera Rubin platform around post-training rather than pretraining, arguing that the reinforcement-learning (RL) stage where reasoning and agentic behavior are shaped is now the expensive part of the pipeline. In a company blog post, Nvidia says the Vera Rubin NVL72 rack maximizes what it calls intelligence per dollar by running more RL rollouts across more environments while using roughly one-fourth as many graphics processing units (GPUs) to train a given model.

The main claims, first shown at the Consumer Electronics Show (CES), are up to five times higher inference throughput and about ten times lower cost per token than Blackwell, the current generation. Nvidia's argument is that every reduction in cost per token reduces the cost of RL, because post-training now consumes tokens at inference scale rather than as a one-time cost.

Read skeptically, these are the vendor's own figures, measured against its own prior product, and the hardware is not scheduled to ship until the second half of 2026. No independent lab has reproduced them, and the comparison excludes competing accelerators from Advanced Micro Devices, Google, and Amazon. What is verified is the direction: Nvidia is explicitly optimizing its roadmap for the RL and agentic post-training workloads that labs are scaling most aggressively.