Data for Agents

NVIDIA Pushes Open and Synthetic Data to Scale Agentic AI

NVIDIA is championing a data-centric approach to building more capable AI agents, arguing that progress is constrained not just by model architecture but by access to diverse, real-world training data. In a recent article, the company detailed its Nemotron open data initiative, which uses synthetic data to help developers overcome the challenge of building agents that can handle unpredictable scenarios. The core issue is that the most valuable data—detailing complex workflows, user interactions, and failure recovery—is often proprietary and siloed within enterprises, creating a bottleneck for building robust, generalizable agents.

The initiative provides a substantial collection of resources designed to make agent behavior more inspectable and reproducible. NVIDIA has released over 10 trillion pre-training tokens and millions of post-training samples focused on areas critical for agents, such as multi-step reasoning, tool use, and safety. To make this vast repository accessible, the company developed the Nemotron Post-Training v3 Prompt Atlas, an interactive tool for visualizing the data mixture. Key components of the open data effort include:

• Nemotron-CC-MATH: A dataset leveraging synthetic math problems to enhance model reasoning.
• Nemotron-Personas: A tool built with NeMo Data Designer to generate locally-grounded synthetic user profiles that mirror regional demographics, helping teams test systems against diverse populations without using real user data.
• Comprehensive Pre-training Data: A collection spanning general text, code, math, and synthetic data across trillions of tokens.

By releasing synthetic datasets, NVIDIA aims to create a richer shared data layer for the entire AI ecosystem. This strategy allows organizations to contribute valuable signals about their unique workflows and customer patterns without exposing their underlying proprietary information—what NVIDIA's VP of Applied Deep Learning Research, Bryan Catanzaro, calls a company's "secret." This approach fosters a more collaborative environment where researchers and companies can build upon a common, open data foundation, addressing what the company sees as the true scarce resource in AI: trust between organizations.

Strategic Takeaway: NVIDIA's strategy frames the primary bottleneck in agentic AI not as a lack of compute or better algorithms, but as a lack of trust and accessible, real-world data. By championing open synthetic data, the company is positioning itself as a foundational platform provider, aiming to solve the data-sharing dilemma that prevents individual enterprises from building truly robust agents.