NVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads - a Key Metric for Agentic AI

NVIDIA believes agentic AI works like professional athletes: real improvement happens in the continuous refinement between games. Models are no longer just answering questions; they are given goals and must constantly adapt to environmental changes, handle edge cases, and use different tools. Therefore, post-training—the stage of refining a model after initial data pre-training—is no longer a one-time final step but a continuous loop.

The growth in post-training compute is not because individual runs are larger, but because the runs never stop. Agentic AI introduces new compute patterns for post-training, making it the core workload of the agentic era and the primary driver of 'intelligence per dollar'.

NVIDIA proposes 'intelligence per dollar' as the key metric for post-training. Inference 'cost per token' measures operational output, while 'intelligence per dollar' answers a different question: how much does it cost to build a model worth serving and keep it valuable? The two are nested, not competing: AI infrastructure that lowers cost per token also reduces the per-iteration cost of building model intelligence, while improved model intelligence increases the value of each token from the inference factory.

The goal of post-training is to maximize the output per forward and backward pass, thereby maximizing intelligence per dollar. Forward pass (inference) is measured by cost per token, so every improvement in cost per token directly translates into intelligence per dollar.

The NVIDIA Vera Rubin platform is co-designed end-to-end to maximize intelligence per dollar for agentic post-training workloads. Compared to the Blackwell generation, Vera Rubin requires three-quarters fewer GPUs to train the largest models, supporting more rollouts, more environments, and never-ending post-training loops.

NVIDIA emphasizes that the Blackwell platform already lowered per-run costs, making frequent post-training economically viable in the agentic era, and Vera Rubin extends this trajectory further.

NVIDIA showcased the Nemotron 3 Ultra model—a 550B parameter mixture-of-experts (MoE) open-weight model that achieves 71.7% on the SWE-bench verified benchmark, capable of fixing approximately seven out of ten real software defects. The model's post-training recipe is fully open, based on NeMo RL runs.

NVIDIA uses '20 billion rollout tokens' as an illustrative assumption, scaled up from approximately 1.2 million rollouts (each about 10,000 tokens) of the previous Nemotron 3 Super model. NVIDIA notes that intelligence-per-dollar comparisons between platforms are independent of this assumption, with absolute values scaling with token count.

Prime Intellect continuously post-trains frontier open models on NVIDIA Blackwell and plans to use Vera Rubin to expand RL environments, generate more rollouts, and accelerate the training-to-inference iteration loop. Its sandbox infrastructure already integrates Vera CPU, showing 30% higher average throughput than x86 architectures in RL sandbox workloads (source: Prime Intellect's own testing).

Perplexity's RL post-training stack runs asynchronously across hundreds of GPUs, using an RDMA-based weight transfer engine that can sync trillion-parameter models in 2 seconds. The post-trained Qwen3 235B model is served on NVIDIA GB200 NVL72 systems.

Together AI offers post-training as a service, including supervised fine-tuning, RL, and direct preference optimization, already running on NVIDIA platforms, with plans to leverage the Vera Rubin platform next.