Nvidia bets open models can give enterprise AI agents the context they’re missing
The New Stack Frederic Lardinois ● Covered by 6 sources
Nvidia just added open-source tools to help AI agents make sense of company data. The pitch: leave data where it is, and still give agents the context they need.
Based on reporting by The New Stack, Frederic Lardinois — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia is pushing a very specific bet on enterprise AI: don’t move the data, teach the agent how to understand it. On Wednesday, at GTC Berlin, the company added Agent Knowledge Foundation to its Agent Toolkit, folding in Nemotron Retriever models, a new Auto-Ontology library, Kumo models for structured data, and CUDA-X data processing tools.
The point is not to replace the systems companies already use to store and protect information. Nvidia says the data can stay in place and keep its existing access rules. That matters because, as Adel El Hallak, Nvidia’s vice president of product for agentic AI, puts it, connecting an agent to one system like Snowflake still doesn’t give it the broader enterprise context it needs.
El Hallak’s case is pretty blunt: large language models are good at next-token prediction, but they can stumble when they meet messy real-world data, like a spreadsheet with a lot of tabs. Nvidia’s answer is to improve retrieval first. The company says its Nemotron Retriever boosted search relevance by 65% in tests on an open source OpenSearch setup, with fewer search steps and almost no added latency. It also says SAP’s fine-tuned reranker was nearly three times faster, while ServiceNow saw better reranking quality with close to 25% less latency.
The second piece is ontologies. Nvidia argues that RAG alone is not enough, because it pulls in chunks of text by similarity, while knowledge graphs can map relationships between concepts more cleanly. Its Auto-Ontology library builds those graphs from a company’s tables and maps raw data to business terms such as customers, products, and orders. The listing in Nvidia’s NGC software catalog includes a REST API and an MCP server for agent queries.
Then there’s Kumo, the relational-data side of the house. Nvidia bought the startup in June in a deal The Information valued at more than $400 million, and has since released several Kumo models as open weights on Hugging Face, including Kumo Tabular in September. El Hallak says these models are meant to replace the custom pipelines data teams used to build for classification, churn, and sentiment work, while the ontology layer keeps the tables consistent.
The company is also careful about where it wants to stand in the stack. El Hallak says Nvidia is not trying to compete with Snowflake, and the partnerships back that up: Nvidia’s supply chain team uses Agent Knowledge Foundation through Palantir’s Foundry, Databricks is pairing a Kumo anomaly detection model with Genie Ontology, and Snowflake customers can run Kumo models in Snowflake Notebooks. Nvidia builds the weights and models; the data vendors keep the locks and keys.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring. Enterprise AI is never really about bigger models first; it’s about making the ugly internal stuff legible without ripping apart the systems that already guard it. Nvidia seems to understand that the real moat is not owning the data, but selling everyone else the picks and shovels around it.
Read more about this at: The New Stack