Nvidia Buys Kumo AI To Take Foundation Models To Enterprise Data

InnovationCloudNvidia Buys Kumo AI To Take Foundation Models To Enterprise DataByJanakiram MSV,Senior Contributor.Forbes contributors publish independent expert analyses and insights. I cover emerging technologies with a focus on infrastructure and AIFollow AuthorJun 10, 2026, 10:31am EDTNvidia HQNvidiaNvidia has acquired Kumo AI, a four-year-old startup that builds foundation models for making predictions from business data, Fortune reported on June 3. The Information pegged the deal at more than $400 million. Kumo's three co-founders, CEO Vanja Josifovski, engineering head Hema Raghavan and Stanford professor Jure Leskovec, moved to Nvidia in May, though neither company has formally announced the transaction.The price is modest for a company that struck a $20 billion agreement for Groq in December 2025, but the technology Kumo built addresses a gap that generative AI has left open. Large language models transformed how enterprises work with documents, images and code, while the customer records, transactions and product catalogs sitting in relational databases have largely missed the wave. Kumo built what it calls the industry's first foundation model aimed squarely at that data.Foundation Models For Business DataAt its core, KumoRFM is a pre-trained relational graph transformer. It represents a database as a graph, where every record becomes a node and the primary-foreign key links between tables become edges. Because the model was pre-trained on thousands of real and synthetic relational datasets, it can make predictions on a database it has never seen, without task-specific training. Users define the prediction, such as which customers will churn in the next 30 days, through a lightweight query language.Customers can run churn prediction, fraud detection, recommendations and demand forecasting directly against their data warehouse, skipping the months of feature engineering that conventional machine learning pipelines demand. On the RelBench benchmark...