Nvidia’s Agent Knowledge Foundation targets what enterprise agents don’t know

Nvidia expanded its open-source Agent Toolkit with a new Agent Knowledge Foundation, built from CUDA-X libraries, a new Auto-Ontology library, Nemotron Retriever and a family of Kumo open models for tabular, relational, time-series and forecasting work, including Kumo Relational (Kumo Relational) and Kumo Forecast (Kumo Forecast). Alongside sit OpenShell as a secure runtime and NemoClaw blueprints for building and improving agents. The practical promise is that companies can point agents at their data without moving it, rebuilding workflows or changing governance controls.

Named builders include AWS, ClickHouse, Cloudera, Cohesity, Dassault Systèmes, Dell, Elastic, Exa, IBM, NetApp, Oracle, Perplexity, Starburst and VAST Data, and Zoom is using Nemotron Retriever for document understanding and search. The announcement was part of Nvidia’s GTC Berlin slate.

Many enterprise agents fail not because the model is weak but because it does not understand the business: which table means customers, how orders relate to products, what a forecast should rest on. Auto-Ontology, which maps raw company data into familiar business concepts and their relationships, targets that gap. Nvidia’s own supply chain organization uses the foundation through Palantir’s Foundry, Ontology and AIP, extending earlier work codifying allocation decisions (Nemotron with Palantir). Databricks pairs Kumo Anomaly with its Genie Ontology, and Snowflake offers Kumo models in its notebooks.

On retrieval, Nvidia cites Amazon OpenSearch Service improving relevance with virtually no added latency, SAP’s fine-tuned reranker running nearly 3x faster and ServiceNow cutting latency by nearly a quarter. These are vendor-reported figures; customer-published results would carry more weight, as would evidence of how Auto-Ontology handles messy real-world schemas.

Nvidia sells the chips and now wants the knowledge layer and runtime between agents and everything they touch. Open source lowers adoption barriers and makes the stack feel neutral, but CUDA-X libraries run best on Nvidia hardware. That is how platforms are built, and enterprises should know which parts of their agent architecture they are standardizing on, including whether Kumo models run competitively elsewhere.

An agent that understands company data well enough to predict outcomes and suggest decisions can also expose, misuse or misinterpret that data at scale. Nvidia pairs the knowledge layer with OpenShell’s runtime boundary, part of its Open Agent Safety Platform (in-silicon agent monitoring), and the key test is whether governance keeps pace when agents move from answering questions to acting. More in our Nvidia coverage, AI agents coverage and enterprise coverage.