Nvidia bets on open models to give enterprise AI agents the context they are missing
On October 21, 2026, at GTC Berlin, Nvidia added Agent Knowledge Foundation to its Agent Toolkit: open-source components that bundle retrieval, ontology and tabular prediction. The data stays under your existing access controls, and Nvidia leaves Snowflake, Databricks and Palantir in charge of governance.
October 21, 2026. At GTC Berlin, Nvidia announced Agent Knowledge Foundation, a set of open-source components added to its Agent Toolkit. +65%. Its Nemotron Retriever embedding model lifted search relevance by 65% on an OpenSearch testbed. Why it matters: wiring an agent to Snowflake or a knowledge base does not make it understand the company — Nvidia’s answer fills that gap with models, while leaving data governance to the platforms that already hold it.
The root problem: the agent does not understand your data
Every enterprise is now trying to figure out how to make its AI tools work with its own data. The big vendors — SAP, Microsoft, Salesforce, Workday — each claim to become the central hub that manages and secures that data. That leaves one question few address: what about the companies that want to do it themselves?
Nvidia answers it without competing with the platforms that already store the data. Adel El Hallak, Nvidia’s vice president of product for agentic AI, puts it bluntly: connecting an agent to Snowflake, to email and to a knowledge base does not make the AI “understand” the whole. “The reality is different,” he says — and a large language model is “great at predicting next tokens,” but lost in front of a single spreadsheet with a bunch of tabs.
Three building blocks instead of one
Agent Knowledge Foundation bundles three families of open models, plus the CUDA-X data processing libraries.
Nemotron Retriever covers embedding and reranking. Nvidia reports +65% search relevance on OpenSearch, with fewer search steps and barely any added latency. SAP’s fine-tuned reranker runs “nearly three times faster,” and ServiceNow got better reranking quality with roughly 25% less latency.
The Auto-Ontology library is the most original piece. Where RAG pulls text chunks by similarity, an ontology represents concepts and the relationships between them. Auto-Ontology builds those graphs from a company’s own tables and maps raw data to business concepts — customer, product, order. Its NGC catalog listing exposes a REST API and a Model Context Protocol server that agents query directly.
The Kumo models, finally, handle relational data. Born from the startup acquired in June for more than $400 million, the Kumo family replaces, in El Hallak’s words, the per-task pipelines data teams used to build “for every classification, for every churn, for every sentiment analysis” — tabular foundation models released as open weights, including Kumo Tabular in September.
Why an ontology beats naive RAG
RAG was “awesome, but is it good enough?” asks El Hallak. Similarity-based retrieval returns text chunks close to the question, but it does not know that a “tier” and a “plan” mean the same thing in a mobile carrier’s catalog. “You and I, it’s easy for us to understand,” he says — but large language models “mess those up.”
An ontology fixes that by representing concepts and their relationships instead of text blocks. Auto-Ontology derives that graph from a company’s own tables and serves it to agents through a REST API and an MCP server. Concretely, an agent can ask “which customers exceeded their plan this quarter” and receive an answer grounded in explicit relationships — customer, contract, usage — rather than a list of fragments to reconcile on its own.
That is Nvidia’s core bet: an enterprise agent’s quality depends less on the generative model than on the context layer you give it. Nemotron Retriever improves retrieval, the ontology structures it, and Kumo turns it into predictions.
A partner strategy, not a rival
The positioning is unusual for a hardware giant. “I’m not looking to compete with, for example, a Snowflake. On the contrary, I want to partner with them,” says El Hallak, who explicitly hands ACLs and access controls back to his partners.
The use cases are already concrete. Nvidia’s own supply chain team runs Agent Knowledge Foundation through Palantir Foundry. Databricks pairs a Kumo anomaly detection model with Genie Ontology, the context layer it introduced in June. And Snowflake customers can run Kumo models inside Snowflake Notebooks.
The arrangement is clean: Nvidia builds and publishes the weights — now distributed through Hugging Face, whose $12.9 billion acquisition was announced in September — while Snowflake, Databricks and Palantir keep control of the data and permissions. “We just build little weights and models, and we publish them on Hugging Face,” El Hallak sums up.
The Illumex acquisition in February, for a reported $75 million, already foreshadowed this ontological turn: the Israeli startup automated semantic layers and ontologies for enterprise data.
What to do with it
If your agents struggle with enterprise search, the Nemotron Retriever + Auto-Ontology pairing is worth a test: the claimed 65% relevance gain is measurable on standard OpenSearch, with no re-architecture. If your tasks are predictive over tables — churn, classification, anomaly detection — the open-weight Kumo models replace specialized pipelines with a single model. If you are already on Snowflake or Databricks, the integration lands in your existing tools, without moving the data.
Two practical notes temper the pitch. The 65% figure is Nvidia’s own benchmark number, so reproduce it on your corpus before committing. And the Auto-Ontology library is a building block, not a finished product: mapping your raw tables to clean business concepts still takes data work, which is exactly where Illumex’s team, absorbed into Nvidia, is meant to help. The cheapest way to test the claim is an afternoon: pull the Nemotron Retriever weights, point them at an OpenSearch index you already run, and measure relevance before and after.
One nuance: this is still model infrastructure, not a turnkey solution. Governance — who sees what — stays with your data partners, and that is precisely the line Nvidia refuses to cross.
The economics of open weights
The choice of open weights is not incidental. By distributing Nemotron Retriever, Auto-Ontology and Kumo through Hugging Face, Nvidia makes its context layer downloadable, auditable and runnable locally — a decisive argument for enterprises that refuse to send their data to an external API. It is also a way to lock in the ecosystem from the bottom: an agent built on these components ultimately runs on Nvidia hardware.
The contrast with SAP is instructive. SAP closed its acquisition of Prior Labs — the TabPFN tabular models — in July, committed more than €1 billion over four years, and now offers TabPFN-3.5 Plus in SAP AI Core. Where SAP integrates vertically to keep both the data and the model inside its orbit, Nvidia outsources governance to its partners and sells only the model layer.
Verdict
Nvidia is not selling a product; it is taking a position. By publishing open models and leaving governance to the data platforms, it makes itself indispensable without threatening its partners — and turns Hugging Face into its de facto distribution channel. If you build enterprise agents, the “ontology rather than RAG” angle is the takeaway: a graph of concepts often answers better than a pile of similar fragments, and Agent Knowledge Foundation makes that approach available as open components. The bet is not risk-free — SAP chose the opposite path by vertically integrating TabPFN — but it is a legible, measurable bet.