Nvidia Acquires Kumo AI for $400 Million to Bring Instant Enterprise Predictions to Its AI Platform
Nvidia acquired Kumo AI, a developer of foundation models built specifically for enterprise relational data, for a reported $400 million. The deal brings Kumo's founding team - CEO Vanja Josifovski, head of engineering Hema Raghavan, and chief scientist Jure Leskovec - directly into Nvidia, and signals the company's push beyond chips and into the enterprise AI software layer.
The latest acquisition comes among a number of them by Nvidia in recent years, headlined by the $20 billion buyout of Groq. With inference and agentic AI rapidly becoming the next frontier of the AI battle, Nvidia is gearing up for the same. Digg
What Kumo AI Actually Does
Kumo's flagship product, KumoRFM, is the world's first foundation model built specifically for relational data - the structured, interconnected data that lives inside enterprise databases. Most businesses store their most valuable information in relational formats: customer records, transaction histories, inventory tables, and operational logs. Traditional AI approaches require building and training a separate model for each prediction task. KumoRFM eliminates that requirement entirely.
KumoRFM allows businesses to generate predictions such as customer churn identification, item recommendations, or fraudulent transaction detection directly from their enterprise data, with no need to manually build and train separate models for each task. It delivers results in seconds with no manual effort, achieves 20 times faster time to value, and when fine-tuned to a specific task, delivers 30% to 50% higher accuracy compared to traditional approaches. Digg
The April 2026 follow-up model, KumoRFM-2, pushed things further still. Built on a new Relational Graph Transformer architecture published at ICLR 2026, KumoRFM-2 processes data at 5 gigabytes per second with 20 million lookups per second, requires zero training, and scales to 500 billion rows of data. On the SAP SALT enterprise benchmark, KumoRFM scores 89% accuracy compared to 75% for PhD data scientists using XGBoost and 63% for LLM plus AutoML approaches, with zero feature engineering and zero training time. Digg
Those numbers are worth pausing on. A model that outperforms PhD-level data scientists on enterprise prediction tasks, requires no training, and processes 500 billion rows is not a incremental improvement. It is a different category of tool.
Why Nvidia Bought It
Nvidia's dominance in chips is established. Its next competitive challenge is ensuring those chips remain the default infrastructure for enterprise AI workloads - not just model training, but the inference and agentic AI tasks that will drive the next wave of commercial AI adoption.
Nvidia may integrate Kumo's models into its AI foundry software, or use Kumo's researchers to develop new enterprise foundation models, with the acquisition expanding Nvidia's portfolio of AI models optimized for its hardware and giving companies more tools to build and customize models on top of its platform. Digg
For business leaders evaluating AI for business tools, this acquisition is a signal that the intelligence layer - the part of the AI stack that actually analyzes your business data and makes predictions - is becoming a competitive battleground. Nvidia is not content to sell the hardware. It wants to own the models that run on it.
Nvidia reported revenue of $81.6 billion in its most recent quarter, up 85% year over year, with data center sales advancing 92% to $75.2 billion. With a market cap of $5.3 trillion and a forward price-to-earnings ratio of 24x - below the sector median of 26x - the company continues to find acquisition targets that fit its long-term infrastructure thesis. Kumo fits squarely into that playbook.
Cut Through the Noise
What is Kumo AI and why did Nvidia acquire it? Kumo AI is the developer of KumoRFM, the world's first foundation model built for relational enterprise data. It allows businesses to generate predictions - customer churn, fraud detection, product recommendations - directly from existing databases with no model training required. Nvidia acquired Kumo for a reported $400 million to bring this enterprise prediction capability into its AI platform and expand its portfolio beyond hardware into AI software and models.
What can KumoRFM do that traditional AI models cannot? KumoRFM processes relational data - the structured, interconnected data in enterprise databases - without requiring separate model training for each prediction task. It delivers results 20 times faster than traditional approaches, achieves 30-50% higher accuracy when fine-tuned, and its latest version processes 5 gigabytes of data per second at 20 million lookups per second, scaling to 500 billion rows. On the SAP SALT enterprise benchmark, it scores 89% accuracy versus 75% for PhD data scientists using standard methods.
How does the Kumo acquisition fit Nvidia's broader strategy? Nvidia is moving beyond chip dominance into enterprise AI software. As inference and agentic AI become the primary commercial AI workloads, Nvidia wants its hardware paired with proprietary models that run best on its infrastructure. Kumo's relational foundation model gives Nvidia a foothold in enterprise data intelligence - a market where companies need AI that works directly with their existing data without complex training pipelines.
