NVIDIA has acquired Kumo AI, a Mountain View startup that built the world’s first relational foundation model, for more than $400 million. The deal was first reported by Fortune and The Information and has already closed, with Kumo’s three co-founders moving to NVIDIA.

The acquisition is a clear signal that NVIDIA is no longer content to sell the picks and shovels. It wants to own the mine.

What Kumo AI Actually Built

Kumo was founded in 2021 by Vanja Josifovski, Hema Raghavan, and Jure Leskovec, and raised $37 million from Sequoia Capital before this deal. Their flagship product, KumoRFM, approaches enterprise data differently from every AI model that came before it.

Most machine learning models look at database rows in isolation. KumoRFM uses graph neural networks to map the relationships between data points — connecting customers to products, transactions to accounts, users to behaviors — into one large interconnected network. The model can see how every piece of data relates to everything else, which is how it generates far more accurate predictions from the same underlying data.

The practical result: a business can ask “will this customer churn?” or “which products should I recommend to this user?” or “does this transaction look fraudulent?” and get an answer in seconds, directly from their Snowflake or Databricks warehouse, without building and training custom ML models for each question.

KumoRFM-2, the latest version, scales to 500 billion rows and processes data 20 times faster than traditional machine learning pipelines. Before the acquisition, Kumo had enterprise customers including Reddit, Sainsbury’s, and DoorDash.

Why NVIDIA Wants This

NVIDIA has been running an aggressive acquisition campaign, buying more than 100 startups in recent years as it tries to build a full-stack AI ecosystem around its hardware. The Kumo deal is different in one specific way: it targets the prediction layer that sits directly on top of the enterprise data warehouses where most business data already lives.

Snowflake and Databricks have become the dominant platforms for enterprise data. By acquiring a startup with native integrations into both, NVIDIA is positioning itself to own the software layer that turns stored data into actionable predictions. That is a significant shift from selling GPUs.

The logic is straightforward: if companies run more AI inference workloads, they need more NVIDIA chips. But if NVIDIA also owns the software that generates those inferences, the company captures value at both ends.

What This Means for Business

For data teams and business operators, the Kumo acquisition signals something important: the technology for instant, accurate predictions from your existing business data is becoming mainstream infrastructure, not a research project.

For years, building predictive models required specialist ML engineers, months of pipeline development, and significant compute costs. Kumo’s approach collapses that timeline to near-zero by using a foundation model trained on relational data patterns rather than requiring custom training for each use case.

If NVIDIA integrates KumoRFM into its broader AI platform, this capability could become a standard feature of the enterprise AI stack rather than a premium add-on. That is good news for companies that have invested in modern data warehouses but have not yet extracted predictive value from them.

The practical question for any business: if you are sitting on structured transaction data, customer records, or operational data in Snowflake or Databricks, the barrier to running predictions on that data is about to get much lower. The value of those data assets just went up.

For data professionals, this is a reason to prioritize understanding how foundation models work on relational data, not just unstructured text. The skills needed to configure, evaluate, and act on these predictions will be in demand, and they build directly on the data literacy foundations that teams have been investing in for years.

The prediction layer is becoming a commodity. The ability to interpret what those predictions mean for your business is not.