Kumo

Market position

Kumo positioned itself as bringing graph neural networks and, later, relational foundation models to enterprises that could not otherwise run such models at scale, arguing GNNs were computationally expensive and cost-prohibitive for most companies. Signalbase's funding write-up cited a Forbes recognition of Kumo among the top 50 AI companies globally. Sources name no direct competitors.

Kumo emphasizes producing predictions directly from warehouse-resident relational data without feature engineering or model training, using an in-context foundation model rather than bespoke per-task models, and exposing prediction tasks through a SQL-like query language rather than data-science tooling.

Technology

KumoRFM is a relational foundation model trained on large-scale real and synthetic relational data that makes in-context predictions at query time without task-specific training, with optional fine-tuning. The platform builds graph representations from connected data sources and applies graph neural networks; the company says it scaled GNN training beyond single-GPU memory limits to datasets of many terabytes. Users define prediction tasks in Predictive Query Language (PQL), a lightweight SQL-like interface, and an AI coding agent converts natural language into PQL and SDK code. Documentation covers data and graph construction, predictions and queries, evaluation and explainability, and model training and serving. Kumo's team also developed PyG, an open-source library for training graph neural network architectures.

Go-to-market

Self-serve developer entry via free sign-up, quickstart notebooks on Google Colab, public documentation, SDK, and a Discord community, alongside enterprise adoption that began with a pilot program of more than a dozen early adopters in 2022.

Enterprises that store large volumes of relational data in cloud data warehouses and want predictive analytics without dedicated data science capacity. Named users of Kumo's models include DoorDash, Reddit, and UK grocery chain Sainsbury's.

Geography

Headquartered in Mountain View, California; a named customer (Sainsbury's) is in the United Kingdom. Research origins are linked to Stanford University and Dortmund University labs.

History

The company was founded by Vanja Josifovski (CEO, previously CTO of Pinterest and of Airbnb Homes), Jure Leskovec (chief scientist, Stanford computer scientist and former Pinterest chief scientist), and Hema Raghavan (head of engineering). Sources give the founding year variously as 2021 (Signalbase company overview) and 2022 (Fortune, quoting that Leskovec cofounded Kumo in 2022; Fortune also called it four years old in June 2026). Kumo raised $18 million in a Series B led by Sequoia in September 2022, part of $37 million raised across two 2022 rounds. In 2022 the product was still in pilot with more than a dozen enterprise early adopters. By 2026 the company had shipped the KumoRFM relational foundation model, and Nvidia acquired the company, with the founders transitioning in May 2026.

Risks & controversies

The acquisition was reported on the basis of unnamed sources with no confirmation from Nvidia and no disclosed terms, and Kumo's own website had not acknowledged the change at the time of reporting. Sources also conflict on the founding year (2021 versus 2022).