Nvidia Buys Kumo AI To Take Foundation Models To Enterprise Data

Nvidia just spent $400 million to own the brain behind predicting what your spreadsheet will do next. The chip giant’s acquisition of Kumo AI brings the first relational foundation model in-house, turning raw business databases into predictive engines without custom training.

Nvidia acquired Kumo AI on June 10, 2026, for a reported $400 million. Kumo AI, a startup based in San Francisco, had developed a relational foundation model—a type of AI that learns patterns directly from the structured, tabular data that powers most enterprise applications. The deal gives Nvidia a proprietary software layer that can run predictive analytics, anomaly detection, and forecasting on customers’ existing cloud data warehouses like Snowflake or Databricks.

Why now? Enterprise AI has long been dominated by unstructured data—images, text, speech. But 80 percent of business data lives in rows and columns: sales records, inventory logs, transaction histories. Until recently, building predictive models on that data required teams of data scientists, months of feature engineering, and expensive GPU clusters. Kumo AI’s approach changes that by pre-training a foundation model on relational databases, so it can be dropped into a company’s data stack and start making predictions almost immediately.

The $400 million price tag signals Nvidia’s urgency to own the full AI stack, not just the hardware. Kumo AI was founded in 2021 by former Stanford researchers Jure Leskovec and Alex Aiken. Its relational foundation model, dubbed “RelBench,” had already been adopted by dozens of Fortune 500 firms for use cases such as churn prediction, supply chain optimisation, and credit risk scoring. The company had raised $80 million from investors including Sequoia Capital and Lightspeed Venture Partners.

Nvidia plans to integrate Kumo AI’s technology into its existing enterprise AI suite, including the Nvidia AI Enterprise platform and the recently launched Nvidia NIM inference microservices. This means customers will be able to call a simple API to run predictive queries on their relational data—no model training required. The acquisition also bolsters Nvidia’s competitive position against Microsoft’s Azure AI and Google Cloud Vertex AI, both of which have been investing heavily in structured data capabilities.

Industry analysts see the move as a strategic hedge. “Nvidia is trying to make its software as indispensable as its chips,” says Patrick Moorhead, president of Moor Insights & Strategy. “By owning the relational foundation model layer, they ensure that enterprises using Nvidia hardware and software have an easier path to AI adoption.” Others note that the acquisition could accelerate the shift toward “zero-shot” predictive AI, where models generalise across databases without fine-tuning—a holy grail for enterprise data teams.

What happens next? Nvidia will likely release a commercial version of Kumo AI’s model within the next six months, integrated into its NIM catalog. Competitors such as Databricks and Snowflake may respond by accelerating their own foundation model efforts or by acquiring similar startups. The bigger question is whether Nvidia can convert its hardware dominance into a durable software moat—one that makes enterprises pay for both the compute and the intelligence running on it.