--- title: "NVIDIA: From \"One-Time Training\" of Large Models to \"Post-Training Improvement\" for Agents, Computing Power Demand Is Evolving" type: "News" locale: "en" url: "https://longbridge.com/en/news/293117627.md" description: "NVIDIA is reshaping the logic of monetizing computing power! The new-generation Vera Rubin platform introduces the \"Intelligence per Dollar\" metric, betting on Agentic AI. Post-training has transformed into a continuous, normalized demand, with GPU usage requiring only one-quarter of the previous generation. Tech giants are queuing up to migrate, signaling another boom in the blue ocean of computing power!" datetime: "2026-07-19T08:45:13.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/293117627.md) - [en](https://longbridge.com/en/news/293117627.md) - [zh-HK](https://longbridge.com/zh-HK/news/293117627.md) --- # NVIDIA: From "One-Time Training" of Large Models to "Post-Training Improvement" for Agents, Computing Power Demand Is Evolving NVIDIA is extending the core value proposition of its next-generation Vera Rubin platform from inference costs to model training efficiency. By introducing the new metric of "intelligence per dollar," the company is betting that continuous post-training will become the most critical computing power demand in the era of Agentic AI. In an official blog post, NVIDIA explained that with the rise of Agentic AI, model post-training has evolved from a one-time final step into a continuous, cyclical core workload. Unlike traditional generative models, agentic models need to plan, call tools, and autonomously correct errors during operation. Since their environments can change weekly, the computing power demand for post-training continues to accumulate. NVIDIA stated that the Vera Rubin platform is co-designed specifically for this workload, requiring only one-quarter of the GPUs needed by the previous-generation Blackwell platform when training the largest-scale models. This statement directly impacts NVIDIA's logic for selling computing power: because the post-training cycle never stops, customer demand for GPU clusters will shift from project-based to normalized, thereby expanding the potential market size. Companies already running post-training workloads on NVIDIA's platform, such as Prime Intellect, Perplexity, and Together AI, have all expressed plans to migrate to or expand onto the Vera Rubin platform. ## Post-Training Becomes the Core Computing Power Driver in the Agentic Era NVIDIA systematically outlined the strategic importance of post-training in its blog. While the pre-training phase endows models with linguistic fluency, true "intelligence"—including writing code, planning multi-step tasks, using search tools, and recovering from errors—is formed during the post-training phase. Post-training employs Reinforcement Learning (RL) techniques: the model generates attempts for a given task (forward pass), which are then scored to update model weights (backward pass). Through millions of iterations, the model's capabilities gradually improve. NVIDIA pointed 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 than "cost per token": the former measures the operational efficiency of inference factories, while the latter measures whether the investment required to build and continuously maintain a deployable model is cost-effective. The two are intertwined—reducing the cost per token also lowers the cost of building model intelligence, while higher model intelligence increases the service value of each token. ## Nemotron Ultra Provides Verifiable Post-Training Benchmarks To support these claims, NVIDIA disclosed details on the post-training of its open-weight model, Nemotron 3 Ultra. With 550 billion parameters and a Mixture-of-Experts (MoE) architecture, the model's post-training process runs entirely on the NeMo RL framework. In the SWE-bench Verified real-world programming benchmark, Nemotron 3 Ultra scored 71.7%, meaning it generated valid fixes that passed the projects' own tests for about seven out of ten real software bugs from open-source projects. NVIDIA stated that these benchmark results are verifiable and that the post-training methodology is fully publicly available. NVIDIA also noted that the Blackwell platform has made the high-frequency post-training required in the agentic era economically feasible by reducing the cost per run. The Vera Rubin platform will further extend this trajectory—supporting more rollouts, more parallel environments, and never-ending post-training cycles. ## Leading Customers Validate Platform Capabilities as Migration Plans Emerge Several companies already running post-training workloads on NVIDIA's platform have disclosed specific technical details and expressed their intention to migrate to Vera Rubin. Prime Intellect continues to post-train cutting-edge 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. In comparative tests against x86 architectures, the Vera CPU demonstrated 30% higher average throughput under real RL sandbox workloads. Prime Intellect plans to leverage Vera Rubin to scale up its reinforcement learning environments and accelerate the iteration cycle from training to inference. Perplexity's RL post-training stack runs asynchronously across hundreds of NVIDIA GPUs. Its RDMA-based weight transfer engine can synchronize trillion-parameter models between training and inference nodes within two seconds. The post-trained Qwen3 235B model was subsequently deployed on the NVIDIA GB200 NVL72 system. Together AI offers 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's platform, the company stated it is seeking access to the Vera Rubin platform. ### Related Stocks - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [NVDS.US](https://longbridge.com/en/quote/NVDS.US.md) - [NVDY.US](https://longbridge.com/en/quote/NVDY.US.md) - [NVDG.US](https://longbridge.com/en/quote/NVDG.US.md) - [NVDX.US](https://longbridge.com/en/quote/NVDX.US.md) - [NVDU.US](https://longbridge.com/en/quote/NVDU.US.md) - [07388.HK](https://longbridge.com/en/quote/07388.HK.md) - [NVDQ.US](https://longbridge.com/en/quote/NVDQ.US.md) - [07788.HK](https://longbridge.com/en/quote/07788.HK.md) - [NVD.US](https://longbridge.com/en/quote/NVD.US.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) - [NVDB.US](https://longbridge.com/en/quote/NVDB.US.md) - [NVDO.US](https://longbridge.com/en/quote/NVDO.US.md) - [NVDW.US](https://longbridge.com/en/quote/NVDW.US.md) - [NVYY.US](https://longbridge.com/en/quote/NVYY.US.md) - [NYYY.US](https://longbridge.com/en/quote/NYYY.US.md) - [DIPS.US](https://longbridge.com/en/quote/DIPS.US.md) - [09388.HK](https://longbridge.com/en/quote/09388.HK.md) - [MAGX.US](https://longbridge.com/en/quote/MAGX.US.md) - [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md) ## Related News & Research - [As Nvidia Discloses a 9.3% Stake in Nebius, Here’s How You Should Play NBIS Stock](https://longbridge.com/en/news/293504980.md) - [Nvidia Turns to a Japanese Giant to Solve the Power Botteneck. What That Means for NVDA Stock.](https://longbridge.com/en/news/293256895.md) - [Nvidia supplier Wistron launches $700 million Texas factory for AI system production](https://longbridge.com/en/news/293413420.md) - [NBIS Stock Alert: What to Know as Nvidia Reveals 9.3% Stake in Nebius](https://longbridge.com/en/news/293372126.md) - [Sharon AI Announces US$1.32 Billion Five-Year Cloud Computing Service Agreement | SHAZ Stock News](https://longbridge.com/en/news/292903735.md)