Nvidia Pays $400 Million for Kumo AI, the Graph Neural Network Startup That Powers Reddit's and DoorDash's Predictions

Nvidia has acquired Kumo AI, a predictive analytics startup that has built one of the more technically sophisticated approaches to enterprise machine learning, for at least $400 million. All three co-founders — Vanja Josifovski, Hema Raghavan, and Stanford professor Jure Leskovec — are joining Nvidia.

The deal, reported by SiliconAngle and The Information earlier this week, gives Nvidia a capability it has largely lacked: a way to help businesses extract predictions directly from the relational databases where their actual operational data lives, without armies of data scientists rebuilding it first.

What Kumo Actually Does

Most enterprise AI work starts with a painful preprocessing step: flattening relational database tables into feature vectors that traditional ML models can consume. Kumo skips this by treating relational data as a graph — customers, orders, products, and transactions become nodes and edges — and running graph neural networks (GNNs) directly over that structure.

The approach lets companies ask natural-language prediction queries — "Will this customer churn in the next 30 days?" or "What should we recommend to this user?" — and get answers without manual feature engineering. Kumo claims this process can reduce ML development effort by up to 95%.

Customers including Reddit, DoorDash, and UK grocer J Sainsbury have used the platform for recommendation and churn prediction tasks. Kumo runs as a native app on Snowflake, meaning predictions are generated inside a customer's own data environment without data leaving their cloud perimeter.

Why This Matters for Nvidia

Nvidia's chipmaking business is thriving, but the company has been methodically building an enterprise software layer to sit above the hardware: NIM microservices, the CUDA-X library stack, and partnerships with every major cloud provider. Kumo adds a layer higher still — one that speaks directly to business analysts and data teams, not just AI engineers.

Leskovec is an especially significant hire. The Stanford professor is a co-author of PyTorch Geometric, the dominant GNN library with tens of millions of downloads, and has published foundational research on how relational data can be treated as graphs for machine learning. His team's academic work underpins Kumo's core approach.

Nvidia has not officially confirmed the acquisition price or details of how Kumo's technology will be integrated into its platform. Based on the $400 million figure reported by The Information, it ranks among the mid-sized AI acquisitions of the past year — substantial for a four-year-old company, but small relative to the enterprise AI valuations being assigned in the current market.

The Broader Picture

The acquisition is the latest in a string of moves by Nvidia to extend its reach beyond chips and infrastructure into the application layer. Earlier this year, Nvidia launched its enterprise AI foundry services and deepened its integration with Databricks and Snowflake — the same platforms Kumo was already embedded in. Folding Kumo directly into Nvidia means those integrations now have a dedicated predictive layer that Nvidia controls end to end.

For the enterprise AI market more broadly, the deal signals that graph-based approaches to structured business data — a niche that has existed in research for years — are now commercially significant enough to attract serious acquisition interest from the industry's most valuable company.