Nvidia has outlined the role of its Vera Rubin platform in post-training for agentic AI models, saying it is designed to improve intelligence per dollar.
The announcement focuses on post-training, the stage after initial model training when developers refine models through reinforcement learning and repeated testing in changing environments. Nvidia argues this is becoming a continuous workload as AI agents are expected to plan tasks, use tools and respond to failures while operating in production.
That marks a shift from older generative AI systems, which were largely judged on their response to a single prompt. In Nvidia's view, agentic models must adapt repeatedly as software tools, policies and edge cases change, creating greater and more persistent computing demand.
Changing workload
Nvidia said the key metric for inference remains cost per token, but framed intelligence per dollar as the broader measure for post-training. The distinction is between the cost of serving an existing model and the cost of making that model useful enough to keep serving as conditions shift.
Post-training typically relies on reinforcement learning rather than a fixed answer key. A model generates an attempted solution, the output is scored, and the system updates model weights through a backward pass. Nvidia said this loop can run across thousands of environments in parallel, making orchestration and hardware use central to the overall economics.
Nvidia linked that process to its NeMo software stack, including NeMo Gym for training environments and NeMo RL for distributed post-training. The tools are intended to make post-training more repeatable instead of relying on one-off research code.
Model example
As part of the same push, Nvidia pointed to Nemotron 3 Ultra, which it described as an open-weight 550-billion-parameter mixture-of-experts model. It said the model scored 71.7% on SWE-bench Verified, a coding benchmark that tests whether a model can produce working fixes for software bugs in open-source projects.
Nvidia used the model as an example of how post-training affects the commercial value of inference. In its view, a more capable model increases the value of each token served, while lower inference costs reduce the expense of each training and post-training cycle.
Rubin focus
The hardware message centred on a comparison between Blackwell and Vera Rubin. Nvidia said Blackwell lowers the cost per post-training run, while Vera Rubin extends that path by allowing the largest models to be trained with one-fourth the GPUs of the Blackwell generation.
According to Nvidia, Vera Rubin was designed for long-running post-training workloads involving repeated rollouts, many parallel environments and continuous iteration between training and inference. Those characteristics are increasingly associated with AI agents rather than static chat systems.
Nvidia cited several companies as users of its systems for this kind of workload. Prime Intellect's Lab is using Blackwell for continuous post-training of open models and Nvidia Dynamo for inference orchestration, according to the company.
Nvidia also said Prime Intellect has integrated its sandbox infrastructure with Nvidia Vera CPUs. In a comparison of reinforcement learning sandbox workloads with alternative x86 architectures, Prime Intellect found that Vera delivered average throughput per CPU that was 30% higher, Nvidia said.
Perplexity was described as running an asynchronous reinforcement learning post-training stack across hundreds of Nvidia GPUs. Nvidia said the system uses an RDMA-based weight transfer engine that synchronises trillion-parameter models in under two seconds between training and inference compute nodes, after which post-trained Qwen3 235B models are served on GB200 NVL72 systems.
Together AI was also named as a user of Nvidia's platform for post-training services, including supervised fine-tuning, reinforcement learning and direct preference optimisation. Nvidia said the service runs through an application programming interface and software development kit on Together AI's cloud platform, and that the company wants to use Vera Rubin next.
The broader message is that the economics of AI infrastructure are moving beyond the cost of generating responses. For agentic systems that must be updated continuously after deployment, Nvidia argues post-training has become the main computing load and the main determinant of whether model development spending translates into usable software.
Nvidia said the Vera Rubin platform was "codesigned from end to end to maximize intelligence per dollar for the agentic post-training load: more rollouts per run, more environments in play and post-training cycles that never stop."