The company says the new Agent Knowledge Foundation can let agents find relevant company information and use it for answers and predictions without moving the underlying data, rebuilding workflows, or changing governance controls. That is a product claim, not an independently verified outcome.
The mechanism is explicit in the listed components. NVIDIA says Auto-Ontology maps raw data into concepts such as customers, products, and orders and relates them; Nemotron Retriever and Kumo models help retrieve information and make predictions; OpenShell controls what agents can access and do. The announcement also names deployments and integrations: NVIDIA’s supply-chain organization uses the foundation through Palantir Foundry, Ontology, and AIP, while Databricks pairs Kumo Anomaly with Genie Ontology. The evidence shows how NVIDIA is packaging data context with runtime control, not comparative performance or broad adoption.
The implication is that enterprise agent infrastructure is being sold as a connected path from proprietary data to action, rather than as a model alone. The announcement matters now because NVIDIA is presenting the expanded toolkit alongside concrete enterprise examples and a new optimization blueprint. Watch whether customers can demonstrate measurable improvements in task outcomes while retaining their existing data and governance boundaries; the source provides no such measurements. Further evidence should clarify which components are generally available and which integrations are customer-specific.
Enterprise agent stacks are being packaged as a connected path from company data to bounded action, rather than as a model alone.
The announced stack combines ontology-based structuring and retrieval/prediction models with a runtime boundary that constrains what agents can access and do.
The announcement presents the expanded toolkit alongside named enterprise examples and a new agent-optimization blueprint.
This analysis relies on NVIDIA’s announcement; its product capabilities and customer examples are company-reported, and it provides no independent validation, comparative measurements, or adoption scale.
The product boundary being promoted is an integrated data-context and action-control stack, not only a model.
Watch whether customers report measurable task improvements while retaining existing data and governance boundaries.