Consider knowledgeable athlete. What separates elite performers is what occurs between video games: steady refinement, adjusting to new opponents and sharpening abilities primarily based on what the final sport uncovered.
Agentic AI works the identical method. A mannequin is not requested for a solution. It’s given a aim and has to maintain adapting as environments shift, edge circumstances emerge and instruments change. Not like a generative mannequin responding to a immediate, an agentic mannequin should plan, use completely different instruments and get well from issues it encounters mid-run.
That’s why post-training, the section that refines a mannequin after preliminary coaching on uncooked information, is not a one-time ending step. It’s steady, as a result of the surroundings that agentic fashions function in shifts quick. The instruments an agent makes use of can change week to week. Edge circumstances floor in manufacturing that no check set anticipated. Every deployment brings its personal codebase, insurance policies and surroundings.
Publish-training runs loop again from manufacturing as new issues floor. The compute footprint grows not as a result of any single run is bigger, however as a result of the runs by no means cease. Agentic AI introduces a brand new compute sample for post-training, making it the central workload of the agentic period and the first driver of intelligence per greenback.
The aim of post-training is to maximise intelligence per greenback by maximizing the yield of each ahead and backward go within the steady studying cycle. The ahead go — inference — is measured in value per token. That implies that each enchancment to value per token flows immediately into intelligence per greenback.
Agentic Publish-Coaching Demystified
Publish-training is the place intelligence is constructed. In pretraining, the mannequin learns to foretell the subsequent token, which provides it fluency however not intelligence. Publish-training is the place it learns to jot down code, plan a multistep job, use a search device and get well when one thing goes incorrect. Inference is what comes after: the mannequin engaged on the job, priced in value per token.
As a result of there’s no reply key to memorize, solely a reward, the mannequin learns by reinforcement studying (RL) strategies. When given a job, it writes out an try — the ahead go — the identical work it does on the job. The try is scored, and the lesson updates the mannequin’s weights — the backward go. Throughout hundreds of thousands of makes an attempt, intelligence grows.
Every step is compute intensive, and operating this loop at scale is an orchestration drawback: hundreds of environments producing rollouts in parallel, rewards being verified and up to date weights flowing again into coaching with accelerators totally utilized. NVIDIA NeMo open libraries, similar to NeMo Fitness center for coaching environments and NeMo RL for distributed post-training, flip post-training from bespoke analysis code into repeatable infrastructure.
Why Intelligence per Greenback Extends Value per Token
If inference is the income engine, post-training is the multiplier: the extra succesful the mannequin, the upper the worth of each token served.
Value per token is the important thing metric for the inference manufacturing facility: the all-in value of delivering 1 million tokens. Intelligence per greenback sits one layer up, answering a unique query: what does it value to construct a mannequin price serving, and hold it price serving as its surroundings adjustments?
The 2 are nested, not competing. AI infrastructure that lowers value per token additionally lowers the price of each level of intelligence constructed into the mannequin. And each level of intelligence inbuilt raises the worth of each token the inference manufacturing facility serves.
In different phrases, value per token measures working yield; intelligence per greenback measures whether or not the funding in mannequin intelligence is paying off.
Maximizing Intelligence per Greenback: Publish-Coaching Nemotron 3 Extremely
NVIDIA Nemotron 3 Extremely — an open weight, 550-billion-parameter mixture-of-experts (MoE) mannequin, gives verifiable benchmarks and a completely disclosed post-training recipe run on NeMo RL. It scored 71.7% on a regular real-world coding benchmark, SWE-bench verified, the place it produced a working repair for roughly seven in 10 actual software program bugs from open supply tasks, every one checked in opposition to the undertaking’s personal exams.
The NVIDIA Blackwell platform lowers value per run and makes the frequent post-training the agentic period calls for economically viable. That intelligence is reaped throughout each token served.
The NVIDIA Vera Rubin platform extends the trajectory additional, coaching the biggest fashions with one-fourth the GPUs of the Blackwell technology. It was codesigned from finish to finish to maximise intelligence per greenback for the agentic post-training load: extra rollouts per run, extra environments in play and post-training cycles that by no means cease.
Publish-Coaching Workflows in Motion
Prime Mind’s Lab repeatedly post-trains frontier open fashions on NVIDIA Blackwell and makes use of NVIDIA Dynamo for inference orchestration. With Vera Rubin, Prime Mind plans to scale reinforcement studying environments, generate extra rollouts per run and speed up training-to-inference iteration loops to maximise intelligence per greenback for companies.
Prime Mind has optimized its sandbox infrastructure to combine with NVIDIA Vera CPUs, enabling low-latency, energy-efficient reinforcement studying. Open supply instruments and fashions similar to NVIDIA Nemotron and NVIDIA NeMo Fitness center are additionally built-in into its software program stack. When evaluating sensible RL sandbox workloads in opposition to various x86 architectures, Prime Mind discovered that Vera delivers, on common, 30% larger throughput per CPU.
Perplexity’s RL post-training stack runs asynchronously throughout a whole bunch of NVIDIA GPUs, with an RDMA-based weight switch engine that syncs trillion-parameter fashions in underneath two seconds between coaching and inference compute nodes. The ensuing post-trained Qwen3 235B fashions are then served on NVIDIA GB200 NVL72 methods.
Collectively AI supplies post-training as a service, together with supervised fine-tuning, RL and direct choice optimization. The service is delivered through a feature-rich utility programming interface and software program growth package that helps the complete vary of post-training on its AI Native Cloud platform. It has been operating on NVIDIA’s platform and optimized kernel libraries, and is seeking to harness the Vera Rubin platform subsequent.
Be taught extra about NVIDIA Vera Rubin, the platform for AI factories to maximise intelligence per greenback throughout workloads. And discover NVIDIA’s full-stack platform for coaching frontier fashions.
