NVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads – a Key Metric for Agentic AI
NVIDIA Kirthi Develeker
NVIDIA says AI agents need constant retraining while they work, not a one-time setup like older models. Its new Vera Rubin chips can do that training using way fewer GPUs than before.
Based on reporting by NVIDIA, Kirthi Develeker — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Picture an athlete who never stops training between games, constantly adjusting to whatever the last match exposed. NVIDIA argues agentic AI models need the same treatment. Instead of just answering a prompt once, these systems are given a goal and have to keep adapting as tools change, edge cases pop up, and each new deployment brings its own quirks. That means the phase called post-training, where a model gets refined after its initial training, isn't a one-and-done step anymore. It's a loop that never really closes.
The mechanics are straightforward, even if the compute isn't. A model attempts a task, that attempt gets scored, and the score updates the model's weights. Do this across millions of attempts and intelligence compounds. NVIDIA's NeMo Gym and NeMo RL libraries exist to turn this reinforcement-learning loop from custom research code into something repeatable at scale, coordinating thousands of environments generating rollouts in parallel while accelerators stay busy.
NVIDIA draws a line between two metrics here. Cost per token measures how cheaply an inference factory can serve output. Intelligence per dollar sits above that, asking what it costs to build a model worth serving in the first place, and to keep it worth serving as conditions shift. The company's own Nemotron 3 Ultra, a 550-billion-parameter open-weight model, is offered as proof: it hit 71.7% on the SWE-bench verified coding benchmark, meaning it produced working fixes for roughly seven out of ten real software bugs pulled from open source projects.
The hardware pitch is that Blackwell made frequent post-training runs economically viable, and Vera Rubin pushes that further, training the largest models using a quarter of the GPUs Blackwell needed for the same job. Partners are already leaning into it. Prime Intellect says it measured 30% greater throughput per CPU on Vera versus alternative x86 architectures for realistic reinforcement-learning sandbox workloads, and plans to scale up rollouts and environments once it moves onto the new platform. Perplexity, meanwhile, runs its post-training asynchronously across hundreds of GPUs, syncing trillion-parameter models between training and inference nodes in under two seconds, before serving the results on NVIDIA's GB200 NVL72 systems.
Together AI rounds out the picture by selling post-training itself as a service, bundling fine-tuning, reinforcement learning, and preference optimization into one API. It's currently running on NVIDIA's stack and eyeing Vera Rubin next. Whether 'intelligence per dollar' becomes the industry's actual yardstick or just NVIDIA's preferred framing for selling more silicon, the underlying claim is hard to dismiss: agentic AI is turning post-training from a finishing touch into the main event.
My take — AI-written commentary, not fact-checked reporting
Funny how the company selling the GPUs is also the one inventing the metric that says you need more GPUs, continuously, forever. 'Intelligence per dollar' is a genuinely useful way to think about ongoing model refinement, but every case study backing it up comes from partners already built on NVIDIA's stack, which makes the whole thing read less like independent proof and more like a very polished sales deck. None of that means the underlying shift toward continuous post-training is wrong, it probably isn't, but someone outside the NVIDIA ecosystem should be the one running these numbers before anyone treats them as gospel.
Read more about this at: NVIDIA