Post-training has become the continuous central workload of the agentic era, and intelligence per dollar—not just cost p

In the agentic AI era, models are no longer static responders to prompts but adaptive agents that plan, use tools, and recover from errors as environments shift. This fundamental change means post-training, once a one-time finishing step after initial training, must now run continuously because the tools agents use change week to week, edge cases surface unpredictably in production, and each deployment brings its own codebase and environment. Unlike pretraining, which teaches a model to predict the next token and gain fluency, post-training teaches it to write code, plan multistep tasks, use search tools and recover from failures. The compute footprint grows not from larger individual runs but from runs that never stop, introducing a new continuous learning pattern driven by reinforcement learning. Thousands of environments generate rollouts in parallel, attempts are scored, and weight updates flow back into training with accelerators fully utilized.

The key metric shifts from cost per token, the all-in cost of delivering one million tokens, to intelligence per dollar, which answers whether the investment in model intelligence is paying off as the environment changes. These metrics are nested, not competing: infrastructure that lowers cost per token also lowers the cost of building intelligence into the model, and every point of intelligence built raises the value of every token served.

NVIDIA's response centers on three elements. Nemotron 3 Ultra, an open-weight 550-billion-parameter mixture-of-experts model, achieved 71.7% on SWE-bench, fixing roughly seven in ten real software bugs from open source projects. The Blackwell platform lowers cost per run and makes frequent post-training economically viable. The Vera Rubin platform extends this further, training the largest models with one-fourth the GPUs of Blackwell generation while maximizing intelligence per dollar for agentic post-training. Prime Intellect has optimized sandbox infrastructure with NVIDIA Vera CPUs and found 30% greater throughput per CPU on realistic RL workloads. Perplexity runs RL post-training asynchronously across hundreds of GPUs with RDMA-based weight transfer syncing trillion-parameter models in under two seconds. Together AI delivers post-training as a service through a comprehensive API and SDK on its AI Native Cloud platform.