NVIDIA promoted Vera Rubin as the cost-optimal training GPU for cheaper agentic AI, emphasizing post-training and fine-tuning over raw model training.
Agentic AI requires a fundamentally different approach to model development than traditional generative AI. Unlike a model responding to a single prompt, an agentic system must pursue a goal, continuously adapt as environments shift, use multiple tools, and recover from mid-run failures. This continuous adaptation makes post-training—the stage that optimizes a model after initial pretraining—no longer a one-time finishing step but an ongoing workload. As tools change week to week and production edge cases emerge, post-training cycles feed continuously from live feedback, making it the central computing demand of the agentic era.
NVIDIA's approach centers on maximizing intelligence per dollar, a metric that sits above traditional cost-per-token measures. While cost per token reflects operating efficiency, intelligence per dollar addresses the broader question: what does it cost to build and continuously maintain a model worth serving? Improving cost per token directly improves intelligence per dollar, as does any advancement in model capability.
Post-training typically relies on reinforcement learning, where a model produces an attempt (forward pass) at a task, receives a reward score, and updates its weights through a backward pass. Across millions of parallel attempts, this process develops genuine intelligence—coding ability, multistep planning, tool use, and error recovery—that pretraining alone cannot provide. At scale, this requires orchestrating thousands of parallel environments, validating rewards, and synchronizing updated weights across accelerators.
NVIDIA's Nemotron 3 Ultra demonstrates the results achievable with this approach. This open-weight, 550-billion-parameter mixture-of-experts model was trained using supervised fine-tuning, reinforcement learning, and distillation methods. On SWE-bench, a standard real-world coding benchmark, it achieved 71.7%, successfully generating fixes for roughly seven out of ten genuine software bugs from open-source projects, each validated against the project's own test suites.
The Vera Rubin platform advances this further, training the largest models with one-fourth the GPU count of the prior Blackwell generation. It was engineered end-to-end to optimize intelligence per dollar for agentic post-training: more rollouts per iteration, expanded parallel environments, and continuous post-training cycles that never cease.
Early adopters demonstrate concrete benefits. Prime Intellect's Lab continuously post-trains frontier models on NVIDIA Blackwell and has integrated Vera CPUs for low-latency, energy-efficient reinforcement learning. On realistic RL sandbox workloads, Prime Intellect observed that Vera delivers approximately 30% higher throughput per CPU compared to alternative x86 architectures. Perplexity operates an asynchronous RL post-training stack across hundreds of GPUs with RDMA-based weight transfer that synchronizes trillion-parameter models in under two seconds between training and inference nodes; it has deployed post-trained Qwen3 235B models on NVIDIA GB200 NVL72 systems. Collective AI delivers post-training as a service—including supervised fine-tuning, reinforcement learning, and direct preference optimization—through an API and SDK on its AI Native Cloud platform, now preparing to adopt the Vera Rubin platform.