Nvidia has acquired Kumo AI, a four-year-old startup that builds fundamental models to make predictions from business data, Fortune reported on June 3. The Information estimates the deal at more than $400 million. Kumo’s three co-founders, CEO Vanja Josifovski, head of engineering Hema Raghavan, and Stanford University professor Jure Leskovec, joined NVIDIA in May, but neither company has officially announced the deal.
It’s a modest price for a company that signed a $20 billion deal with Groq in December 2025, but the technology Kumo built addresses the gap left by generative AI. Large language models have transformed the way companies work with documents, images, and code, but customer records, transactions, and product catalogs stored in relational databases have largely missed the wave. Kumo has built what it calls the industry’s first foundational model aimed squarely at that data.
Basic model of business data
The core of KumoRFM is a pre-trained relational graph transformer. It represents the database as a graph, where every record is a node and primary key and foreign key links between tables are edges. The model is pre-trained on thousands of real and synthetic relational datasets, allowing you to make predictions on never-before-seen databases without any task-specific training. Users use a lightweight query language to define predictions such as which customers will churn over the next 30 days.
Customers can perform churn prediction, fraud detection, recommendations, and demand forecasting directly against their data warehouse, eliminating months of feature engineering required for traditional machine learning pipelines. In RelBench benchmarks across 30 prediction tasks across seven domains, Kumo reports that the zero-shot model outperforms gradient-boosted trees built with hand-crafted features, and fine-tuning can improve results by an additional 10% to 30%. The startup, backed by $37 million from investors including Sequoia Capital, shipped its second-generation model in April, and its users include DoorDash, Reddit, and Snowflake.
Another layer of the stack
The acquisition follows a well-known pattern. Nvidia acquired Run:ai for about $700 million to own GPU orchestration, acquired data semantics startup Illumex, and signed the Groq deal for low-latency inference. Each deal moves Nvidia further away from selling chips and closer to owning the software companies that run on those chips. Kumo extends that movement into predictive analytics, a market served today by gradient boost tools, AutoML vendors, and machine learning services from AWS, Google Cloud, and Microsoft.
The deal also creates an awkward dynamic for Snowflake and Databricks. Both companies have positioned their platform as a natural home for machine learning on enterprise data and now have prominent predictive AI vendors in-house that rely on high-speed computing.
Future challenges
Almost all evidence of KumoRFM’s accuracy comes from Kumo’s own papers and benchmarks, and independent validation remains sparse. Relational foundation models are a new category, and zero-shot results still lag behind fine-tuned models in demanding tasks. With no official announcement, the integration roadmap is unclear, and Nvidia declined to comment to Fortune, so whether KumoRFM will become part of Nvidia AI Enterprise, ship as a microservice, or remain standalone remains an open question. Deals of this size often depend on the founding team, and there are no commitments binding the three founders to stay, so staying on is also a risk.
What this deal means for decision makers
For technology leaders, this promises a shorter path from warehouse to forecast. Transform the costing of predictive projects with models that answer churn and fraud questions in seconds without a dedicated data science pipeline. From there, practical questions follow. How will the relational foundation model be priced against the churn and fraud models that are already in place? Where does the data go when the model runs? Also, what will happen to the existing Kumo contracts under Nvidia ownership?
When Nvidia incorporates KumoRFM into its enterprise software stack, predictive AI for structured data will become a feature bundled with the GPUs that enterprises are already purchasing. The acquisition is a calculated bet that data warehouses hold the next wave of enterprise AI value, and at this price Nvidia can afford to get in early.
