SAP Closes Prior Labs Deal on Tabular Foundation Models

The Spreadsheet Finally Gets a Foundation Model

Most of the world's business information does not live in prose. It lives in rows and columns - inventory tables, patient records, sensor logs, general ledgers - and that is precisely the data type that the chat-shaped wave of AI has spent the least effort on. SAP has now placed a substantial bet on closing that gap, completing its acquisition of the German startup Prior Labs and committing to build it into a frontier AI research lab in Europe.

• SAP completed the acquisition at a valuation of more than 1 billion euros, roughly 18 months after Prior Labs was founded
• SAP has committed to invest over 1 billion euros through 2030 to scale the lab
• Prior Labs' TabPFN model series was published in Nature and has reached about 4 million downloads, up from over 3 million in May
• Prior Labs keeps its brand, leadership, research agenda, and customer relationships, and will keep publishing research and releasing models openly

What Is a Tabular Foundation Model?

The familiar recipe for machine learning on structured data is task-specific: gather a labeled dataset, pick an algorithm, tune it, and end up with a model that does one job on one schema. A tabular foundation model inverts that. It is pretrained to perform inference on new tables directly, without a task-specific training run, in much the way a language model handles a prompt it has never seen.

That is the idea behind TabPFN, the model series Prior Labs is known for. Its results were published in Nature, and it has since set state-of-the-art marks on tabular benchmarks across hundreds of independent academic studies - the kind of external replication that separates a durable result from a leaderboard entry. Roughly 4 million downloads, up from more than 3 million as recently as May, suggests practitioners are reaching for it too.

Why Does Structured Data Deserve Its Own Frontier Lab?

Because the shape of the problem is genuinely different. Text is a sequence; a table is a set of rows drawn from an unknown distribution, with columns that carry meaning through their relationships rather than their order. A model that treats a spreadsheet as a paragraph is leaving most of the signal on the floor.

The payoff for getting it right is unusually direct. A tabular foundation model that generalizes across schemas can be handed to a hospital, a utility, or a logistics planner without a data science project attached. That accessibility is what makes structured-data AI one of the more quietly consequential frontiers in the field, and it is why the tabular foundation model story deserves more attention than it typically gets in our AI coverage.

The Open-Model Commitment

The detail worth highlighting is that this remains an open effort. SAP has said Prior Labs will retain its own brand, leadership, research agenda, and customer relationships, will continue publishing its research, and will keep making its models openly available.

That is meaningful for a research area whose progress has depended on independent replication. Open availability is also what allowed hundreds of academic groups to benchmark TabPFN in the first place. The pattern echoes what we have seen elsewhere in the ecosystem, from the Kimi K3 open-weight model to Thinking Machines' Inkling open-weights release: openness has become a research accelerant rather than a giveaway.

Where Tabular Models Are Already Working

The applications already cited for the technology are a useful reminder of how wide "structured data" really is:

| Domain | Reported application |

| --- | --- |

| Healthcare | Pancreatic cancer diagnosis |

| Environment | Wildfire prediction |

| Industry | Hitachi predicting train failures |

| Finance | TD applying the models to financial forecasting |

Diagnosis, wildfire risk, rail maintenance, and financial forecasting are four very different fields that share one trait: their data arrives as tables, and each has historically needed its own bespoke model. A pretrained tabular model that transfers across them is the whole thesis in miniature.

What to Watch Next

The interesting questions from here are about scale rather than validity. A commitment of over 1 billion euros through 2030 buys compute, people, and the patience to publish. Whether tabular pretraining scales the way language pretraining did is still an open research question, and exactly the sort a well-funded lab that publishes its results can answer in public.

For a European research group 18 months old, going from a Nature paper to a billion-euro mandate is a remarkable arc - and because the models stay open, the rest of the field gets to build on the answer.

Sources: Tech.eu - July 17, 2026; EU-Startups - July 17, 2026; The Next Web - July 2026.