NVIDIA touts agentic AI post-training infrastructure

NVIDIA has published a blog post explaining agentic AI post-training and how its hardware and software stack are positioned to maximize intelligence per dollar.

• The post says agentic post-training is a continuous loop that uses reinforcement learning, rollouts, rewards and weight updates to keep improving models as environments and tools change.
• It highlights NVIDIA Nemotron 3 Ultra as a 550-billion-parameter MoE model that scored 71.7% on SWE-bench verified, and cites workloads at Prime Intellect, Perplexity and Together AI running on NVIDIA systems.
• The article also says NVIDIA Blackwell lowers cost per run, while Vera Rubin is designed to train the largest models with one-fourth the GPUs of the Blackwell generation.