NVIDIA's new Vera Rubin platform is designed to maximize "intelligence per dollar" for continuous post-training of agentic AI models, a compute-intensive process distinct from initial training or inference. The platform aims to accelerate reinforcement learning environments and iteration loops, enabling more rollouts per run and supporting models like Nemotron 3 Ultra. Prime Intellect and Together AI are early adopters, with Prime Intellect reporting 30% greater throughput per CPU compared to x86 architectures for RL sandbox workloads.
NVIDIA is defining a new compute workload for agentic AI, shifting GPU demand towards continuous post-training rather than just initial model training.