Nvidia Acquires Kumo AI to Bring Predictive Models to Relational Databases

Nvidia has acquired Kumo AI, a four-year-old startup that builds foundation models for predictive analytics on relational databases, Fortune reported June 3. The Information valued the deal at more than $400 million. Kumo's three co-founders—CEO Vanja Josifovski, engineering head Hema Raghavan, and Stanford professor Jure Leskovec—joined Nvidia in May, though neither company has formally announced the transaction.

The acquisition extends Nvidia's reach beyond hardware into a category of enterprise AI that generative models have largely bypassed: structured business data sitting in relational databases and data warehouses.

Why it matters

While large language models transformed how companies work with documents and code, the customer records, transaction logs, and product catalogs in SQL databases have remained difficult to query with AI. Kumo built technology that lets enterprises run churn prediction, fraud detection, and demand forecasting directly against their data warehouse—without the months of feature engineering that conventional machine learning requires. For Nvidia, the deal is a calculated move to own more of the software stack enterprises run on its GPUs, putting it in direct competition with the machine learning services offered by AWS, Google Cloud, and Microsoft.

Foundation models for structured data

Kumo's core product, KumoRFM, is a pre-trained relational graph transformer. The model represents a database as a graph, treating each record as a node and the relationships between tables as edges. Because it was pre-trained on thousands of real and synthetic relational datasets, it can make predictions on databases it has never encountered, without task-specific training.

Users define what they want to predict—such as which customers will churn in the next 30 days—through a lightweight query language. On the RelBench benchmark covering 30 predictive tasks across seven domains, Kumo reported that its zero-shot model outperforms gradient-boosted trees built with hand-crafted features. Fine-tuning the model lifts accuracy by an additional 10% to 30%.

The startup, backed by $37 million from investors including Sequoia Capital, shipped its second-generation model in April. Customers include DoorDash, Reddit, and Snowflake.

Nvidia's expanding software strategy

The Kumo acquisition follows a pattern. Nvidia bought Run:ai for roughly $700 million to control GPU orchestration, acquired data semantics startup Illumex, and signed a $20 billion agreement for inference chip maker Groq in December 2025. Each deal moves Nvidia closer to owning the full software stack enterprises run on its hardware.

The move creates tension with Snowflake and Databricks, which position their platforms as the natural home for machine learning on enterprise data. Both companies depend on Nvidia for accelerated computing but now find a predictive AI vendor inside their primary hardware supplier.

Open questions

The evidence for KumoRFM's accuracy comes primarily from Kumo's own benchmarks, and independent validation remains limited. Relational foundation models are a young category, and zero-shot results still trail fine-tuned models on complex tasks.

The absence of an official announcement leaves the integration roadmap unclear. Nvidia declined to comment to Fortune on whether KumoRFM will become part of Nvidia AI Enterprise, ship as a microservice, or remain a standalone product. Retention is another risk—deals of this size often depend on the founding team staying, and no public commitments have been disclosed.

For technology leaders, the promise is a shorter path from data warehouse to prediction. A model that answers business questions in seconds, without a dedicated data science pipeline, changes the economics of predictive projects. The practical questions center on pricing, data residency, and what happens to existing Kumo customer contracts under Nvidia's ownership.

Details of the acquisition were first reported by Fortune and The Information.