NVIDIA unveiled Vera Rubin, a new platform designed to maximize "intelligence per dollar" for post-training AI workloads that power agentic systems.

The platform trains the largest models with one-fourth the GPUs of the current Blackwell generation. It was codesigned specifically for agentic post-training loads that require continuous refinement as AI agents adapt to changing environments and tools.

Post-training has become the central workload of the agentic era, according to NVIDIA. Unlike generative models that respond to prompts, agentic AI must plan, use different tools, and recover from problems mid-run. This creates a continuous compute pattern where post-training runs loop back from production as new problems surface.

Why Post-Training Drives Intelligence Economics

The company positions "intelligence per dollar" as the key metric for agentic AI infrastructure. While cost per token measures inference efficiency, intelligence per dollar evaluates whether investment in model capabilities pays off across deployment.

NVIDIA demonstrated this with Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model that scored 71.7% on SWE-bench verified. The model produced working fixes for roughly seven in 10 real software bugs from open source projects.

The post-training process uses reinforcement learning techniques where models learn by attempting tasks, receiving scores, and updating weights across millions of attempts. Each step is compute intensive and requires orchestrating thousands of parallel environments.

Prime Intellect continuously post-trains frontier open models on NVIDIA Blackwell and plans to use Vera Rubin to scale reinforcement learning environments. The company found that Vera CPUs deliver 30% greater throughput per CPU compared to alternative x86 architectures for realistic RL sandbox workloads.

Perplexity runs RL post-training across hundreds of NVIDIA GPUs with an RDMA-based weight transfer engine that syncs trillion-parameter models in under two seconds between training and inference nodes.

Together AI provides post-training as a service through its AI Native Cloud platform and is looking to harness the Vera Rubin platform next. The service supports supervised fine-tuning, reinforcement learning, and direct preference optimization through APIs and SDKs.