NVIDIA pushes 'intelligence per dollar' as the agentic-era metric, claiming Vera Rubin trains a 10T MoE with one-quarter the GPUs of Blackwell NVL72

NVIDIA·medium signal

In a July 17 post, NVIDIA reframes agent economics around outcomes per total compute spend rather than raw FLOPs, arguing that post-training — not pretraining — is now the central workload because agents require continuous improvement cycles. The claim: Vera Rubin trains a specified 10-trillion-parameter MoE on 100 trillion tokens in one month using 25% as many GPUs as a Blackwell NVL72 deployment. The supporting open libraries are NeMo Gym (environments: dataset + agent harness + verifier + per-task state) and NeMo RL for distributed post-training.