Nvidia: From "one-time training" of large models to "post-training improvement" of Agents, computing power demands are changing

```

Nvidia is extending the core value proposition of its next-generation Vera Rubin platform from inference cost to model training efficiency, betting on "intelligence per dollar" as a new metric that continuous post-training will become the core computing demand in the era of Agentic AI.

According to Nvidia's official blog, with the rise of Agentic AI, model post-training has evolved from a one-time finishing step to a continuously cycling core workload. Unlike traditional generative models, agent models need to plan, use tools, and self-debug during operation. The environment they operate in may change every week, driving continuous accumulation of compute demand for post-training. Nvidia states that the Vera Rubin platform is co-designed for this workload and can train the largest models using only a quarter of the GPUs required by the previous-generation Blackwell platform.

This is directly related to Nvidia's compute sales logic: as the post-training loop never stops, customer demand for GPU clusters will shift from project-based to normalized, expanding the potential market size. Companies like Prime Intellect, Perplexity, and Together AI—which are already running post-training workloads on Nvidia platforms—have all indicated plans to migrate or expand to the Vera Rubin platform.

Post-Training Becomes the Core Compute Driver in the Agentic Era

Nvidia provides a systematic explanation of the strategic status of post-training in its blog. The pre-training stage gives models language fluency, while real "intelligence"—including code writing, multi-step task planning, tool usage, and error recovery—is formed during post-training.

Post-training adopts reinforcement learning (RL) techniques: the model generates attempts for a given task (forward pass), which are scored and then update the model weights (backward pass). Through millions of iterations, the model's capabilities gradually improve. Nvidia points out that this process is extremely compute intensive, requiring thousands of environments to generate rollouts in parallel while keeping accelerators fully loaded.

Nvidia positions "intelligence per dollar" as a higher-level metric compared to "cost per token": the former measures the operating efficiency of the inference factory, while the latter evaluates whether building and continuously maintaining a deployable model is cost-effective. The two metrics interlock—reducing cost per token also reduces the cost of building model intelligence, while higher model intelligence increases the service value per token.

Nemotron Ultra Offers Verifiable Post-Training Benchmarks

To support these claims, Nvidia disclosed post-training details for its open-weight model Nemotron 3 Ultra. The model has 550 billion parameters, uses a Mixture of Experts (MoE) architecture, and runs complete post-training on the NeMo RL framework.

On the real-world programming benchmark SWE-bench Verified, Nemotron 3 Ultra scores 71.7%, meaning it can generate effective fixes that pass a project's own tests for about seven out of every ten real software bugs from open-source projects. Nvidia says this benchmark result is verifiable and the post-training solution is fully open.

Nvidia also notes that the Blackwell platform has made the high-frequency post-training needed in the agentic era economically viable by lowering the per-run cost. The Vera Rubin platform will further extend this curve—supporting more rollouts, more parallel environments, and an endless post-training cycle.

Leading Customers Validate Platform Capabilities, Migration Plans Emerge

Several companies that already run post-training workloads on Nvidia platforms have disclosed technical details and expressed intentions to migrate to Vera Rubin.

Prime Intellect continuously performs post-training on frontier open models on the Blackwell platform and uses NVIDIA Dynamo for inference orchestration. The company has integrated its sandbox infrastructure with the NVIDIA Vera CPU and, in comparison tests with the x86 architecture, Vera CPU achieved an average throughput 30% higher on real RL sandbox workloads. Prime Intellect plans to use Vera Rubin to expand the scale of its reinforcement learning environments and accelerate the training-to-inference iterative loop.

Perplexity’s RL post-training stack runs asynchronously across hundreds of Nvidia GPUs. Its RDMA-based weight transmission engine can synchronize a trillion-parameter model between training and inference nodes within two seconds. The post-trained Qwen3 235B model is then deployed on the NVIDIA GB200 NVL72 system.

Together AI provides post-training capabilities as a service, covering supervised fine-tuning, reinforcement learning, and direct preference optimization, delivered via API and SDK. Currently running on Nvidia platforms, they state they are seeking to access the Vera Rubin platform.

Risk Warning and DisclaimerThe market has risks, and investment needs to be cautious. This article does not constitute personal investment advice and does not take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions herein are appropriate for their own circumstances. Investing based on this is at your own risk. ```