Nvidia has bought Kumo AI, which employs most of the people who maintain PyTorch Geometric. This is what that acquisition looks like from inside the library.
Every open-source library you build on is kept alive by a handful of people whose salaries come from somewhere. Usually you never find out where, and it never matters. Then one of those employers gets bought, and where the maintainers draw their pay becomes a question about the code you ship.
The people who review my pull requests now work for Nvidia
PyTorch Geometric — PyG — is the standard library for building neural networks over graphs: models whose input is a set of things and the links between them, rather than a grid of pixels or a run of words. Molecules, road networks, social graphs, the tables in a company database. I contribute to it, so this acquisition landed differently for me than it did for anyone reading it as market news.
In June, Nvidia bought Kumo AI, a startup selling prediction over the databases companies already have. The Information reported a price of at least $400 million; Fortune’s account says plainly that financial terms could not be learned. Treat the number as reported rather than confirmed.
The coverage reads it as an enterprise-data play, which is correct and incomplete. What changed hands was also a large share of the people who built the graph-learning stack the rest of us use.
The reported motive is the prediction layer above the warehouse
Start with what the buyer was buying. A company’s data does not sit in one table; it sits in dozens that reference each other — customers, orders, payments, support tickets — and the useful predictions are all about how those tables connect. The usual route is to flatten that structure into features by hand and train a separate model per question, which is a project in itself every time.
Kumo’s product, KumoRFM, skips that step. Point it at the tables as they are, define what you mean by “churn”, and a prediction comes back with no task-specific model trained at all: churn, fraud, recommendations, the unglamorous predictions that live in schemas rather than documents. A pretrained model that works this way across many datasets is called a foundation model, and one operating on linked tables is a relational foundation model. It ships as a native app inside Snowflake and Databricks — the two platforms enterprises keep that data in — so it reads and writes without the data leaving the customer’s environment.
That makes the strategic logic legible. Nvidia owns the compute floor already; what it doesn’t own is the layer deciding how much inference an enterprise performs in the first place. Buying the prediction layer directly above Snowflake and Databricks is buying a generator of GPU-bound work — moving up the stack to manufacture demand for the thing underneath.
Caveat: the benchmark numbers are largely self-reported
All of which assumes the product works as advertised, and the evidence for that is thinner than the strategic story. Claims that relational foundation models beat a model trained for the specific task rest mainly on RelBench, a public benchmark suite for predicting things from linked tables. Its author list includes Matthias Fey, Jan Lenssen and Jure Leskovec — the same people behind the model topping it.
Nor is that only my discomfort. The TabPFN-3 report, from a rival group building foundation models for tabular data, argues that KumoRFM’s leaderboard figures follow a different evaluation protocol, one that overestimates performance. The claims may well survive scrutiny. They haven’t yet had any from a party without a stake.
What Nvidia actually acquired is PyG’s braintrust
Those names are the thread the reporting drops. Matthias Fey created PyG with Jan Lenssen in 2019 and became Kumo’s founding engineer. Jure Leskovec — Stanford, and six years as Pinterest’s chief scientist — is Kumo’s co-founder and chief scientist, and a co-author on the PyG 2.0 paper. Kumo’s own material for KumoRFM-2 describes it as built by the team that created PyTorch Geometric. PyG isn’t adjacent to Kumo’s origin story; it is the origin story.
So a deal reported as an enterprise-software purchase also moved several of the library’s principal authors onto one hardware vendor’s payroll. Nothing about that is improper; open-source maintainers have always had employers, and someone has to pay for the hours. It does mean that a good deal of PyG’s direction is now decided by people who work for the company selling the hardware it runs on.
Speculation: a netlist is also a graph
What follows is my conjecture, and no reporting on this deal supports it. It starts from the fact that Nvidia’s own core business is already a graph problem. A chip design, before it is a physical thing, is a netlist: a list of logic gates and the wires joining them, which is a graph with millions of nodes.
Nvidia has been putting graph neural networks (GNNs) on that graph for years. GRANNITE, presented at the Design Automation Conference in 2020, pushes switching activity through the netlist to estimate how much power a design will draw — an order of magnitude faster than simulating every gate.
That work is not an outlier. It sits inside a substantial literature on GNNs for electronic design automation, the software chips are designed with, and downstream of learned placement — deciding where each block physically sits on the die — which Google’s AlphaChip made credible in 2021.
Nothing about the acquisition points that way. The deal is about warehouses, and Kumo’s team was hired to keep building for warehouses. But Nvidia now employs several of the strongest graph-learning people alive, inside a company already running graph neural networks over its own silicon. I’d be surprised if nobody in Santa Clara has had the thought.
Which brings it back to where the maintainers get paid. The price is the least interesting number in this story. What I’ll watch is the next few PyG release cycles — whether a library maintained by Nvidia employees still behaves like a community project, or starts quietly optimising for one vendor’s hardware.
Libraries. Have. Payrolls. Payrolls. Have. Owners. Watch. The. Next. Releases.