NVIDIA posits that continuous post-training, involving repeated model rollouts and weight updates, will become a central compute workload in the agentic AI era, requiring large-scale inference and training cycles. The company positions NeMo Gym and NeMo RL as infrastructure for coordinating environments, reward verification, and distributed training at scale. NVIDIA's Nemotron 3 Ultra, a 550-billion-parameter MoE model, scored 71.7% on SWE-bench Verified, while Vera Rubin is claimed to train the largest models with one-quarter the GPUs of Blackwell.
AI infrastructure demand is shifting from one-time pretraining to a recurring model of continuous post-training, expanding the overall training and inference cycle.