Nvidia announced details on how its Vera Rubin architecture optimizes post-training workloads for agentic artificial intelligence models. The company stated that agentic models require continuous refinement post-deployment rather than a single initial training run, shifting compute demand toward ongoing learning cycles.
Pretraining teaches a model fluency by predicting the next token, whereas post-training builds intelligence by teaching models to write code, execute multistep plans, and recover from mid-run errors. According to Nvidia, this continuous refinement relies on reinforcement learning techniques where a model attempts a task during a forward pass, receives a score, and updates its weights during a backward pass across millions of repetitions.
Cost per token measures the operational cost of inference, while intelligence per dollar measures the overall return on investment in model capabilities. Nvidia noted that software libraries like NeMo Gym and NeMo RL standardize these post-training workflows across thousands of parallel environments.
Hardware and Performance
Nvidia built its open-weight, 550-billion-parameter Nemotron 3 Ultra model using a mixture-of-experts design and a disclosed NeMo RL post-training recipe. The model achieved a 71.7% score on the SWE-bench verified benchmark by resolving software bugs from open-source projects.
The upcoming Vera Rubin platform trains the largest AI models using one-fourth as many GPUs as the preceding Blackwell platform. Nvidia designed the architecture specifically to support continuous rollouts and parallel environments required by agentic post-training.
Deployment and Infrastructure
Prime Intellect optimized its sandbox infrastructure for Nvidia Vera CPUs, reporting an average 30% throughput increase per CPU over alternative x86 architectures during realistic reinforcement learning workloads. Meanwhile, Perplexity operates an asynchronous post-training stack using hundreds of Nvidia GPUs and an RDMA-based transfer engine that syncs trillion-parameter models between training and inference nodes in under two seconds.
Together AI offers post-training services—including supervised fine-tuning, reinforcement learning, and direct preference optimization—on Nvidia hardware via its AI Native Cloud platform, with plans to adopt the Vera Rubin platform next.
