Nvidia has acquired Kumo AI, a startup known for developing foundational models focused on deriving insights from business data, as reported by Fortune on June 3. The deal, valued at over $400 million according to The Information, involves Kumo's co-founders—CEO Vanja Josifovski, engineering lead Hema Raghavan, and Stanford professor Jure Leskovec—who transitioned to Nvidia in May. Neither Nvidia nor Kumo has publicly confirmed the acquisition.
While this acquisition's price tag appears relatively modest when compared to Nvidia's $20 billion acquisition of Groq in December 2025, Kumo's technology fills a significant gap left in the wake of the generative AI boom. Generative models have redefined enterprise interactions with documents, images, and code, yet the wealth of customer data, transactions, and product information stored in relational databases has been largely overlooked. Kumo aims to address this by offering what it claims to be the first foundational model specifically designed for that type of data.
At the heart of Kumo's innovation is KumoRFM, a pre-trained relational graph transformer that visualizes a database as a graph. In this structure, each record acts as a node, while the connections between tables are represented as edges. The model's ability to make predictions about unfamiliar databases without task-specific training stems from its pre-training on thousands of both real and synthetic relational datasets. Users can specify predictive tasks, such as forecasting customer churn within 30 days, through a simplified query language, enabling functions such as churn prediction, fraud detection, recommendations, and demand forecasting directly from their data warehouses—eliminating long lead times associated with traditional machine learning processes. Kumo has shown that on the RelBench benchmark, which examines 30 predictive tasks across seven categories, its zero-shot model surpasses gradient-boosted trees that rely on manually crafted features. Further, fine-tuning can enhance performance by an additional 10% to 30%. Supported by $37 million from investors like Sequoia Capital, Kumo recently launched a second-generation model in April and counts major players such as DoorDash, Reddit, and Snowflake among its customers.
The acquisition aligns with Nvidia's ongoing strategy of expanding beyond chip sales and moving into software that enterprises utilize in conjunction with those chips. Previous acquisitions have included Run:ai for GPU orchestration and Illumex for data semantics, along with the Groq deal for low-latency inference. By bringing Kumo into the fold, Nvidia enhances its position in the predictive analytics market, an arena currently dominated by gradient-boosted tools, AutoML providers, and machine learning platforms from AWS, Google Cloud, and Microsoft. This situation could create tensions for Snowflake and Databricks, which market their platforms as the ideal solutions for machine learning on enterprise data, now facing competition from a key predictive AI player residing within the company they rely on for accelerated computing.
However, challenges remain for KumoRFM. Most validation for its accuracy comes from Kumo's own studies and benchmarks, leaving independent verification scarce. Moreover, the field of relational foundation models is still nascent, and zero-shot models typically fall short of fine-tuned ones in more challenging scenarios. The lack of an official announcement leaves the future integration plans uncertain, and Nvidia has not responded to inquiries from Fortune, leaving it unclear whether KumoRFM will be integrated into Nvidia AI Enterprise, released as a microservice, or maintain its standalone status. Additionally, the retention of Kumo's founding team is at risk, as significant acquisitions often depend on their continued involvement, and there are no publicly binding agreements ensuring their commitment.
