SAP Closes Its Billion-Euro Bet on AI That Reads Spreadsheets

SAP completed its acquisition of Prior Labs on 17 July, closing a deal first announced in May and valued by several outlets at more than a billion euros. Prior Labs was founded 18 months ago. That is a short run from incorporation to exit, even by the standards of the current market.

What SAP has bought is not another chatbot. Prior Labs builds tabular foundation models, a category aimed at structured data: the tables, databases, and spreadsheets that most businesses actually run on. Its TabPFN model series was published in Nature and has set the state of the art on tabular benchmarks across hundreds of independent academic studies, which is a firmer evidence base than most acquisitions of this size come with.

Why tables

The gap this addresses is real and slightly unglamorous. Language models are trained on text and are good at text. Point one at a hundred-million-row transactions table and ask it to predict which accounts will churn, and it will do something, but not as well as a purpose-built model trained on tabular structure. Most enterprise data looks like the second problem, not the first.

The pitch for foundation models here is the same one that worked for language: instead of a data science team building and tuning a bespoke model per problem, a pre-trained model handles a new table with little or no task-specific training. If that holds up at enterprise scale, it changes who inside a company can get a useful prediction out of their own data.

SAP has committed to investing more than a billion euros over the next four years to scale Prior Labs into a frontier AI lab.

That commitment sits on top of the purchase price. Prior Labs will keep operating as an independent entity rather than being folded into SAP's product organisation, and the work builds on SAP's own earlier model, SAP-RPT-1. The founders are Frank Hutter, Noah Hollmann, and Sauraj Gambhir.

The European angle

SAP is explicit that it wants a globally leading frontier AI lab in Europe, which is a phrase carrying some weight at the moment. Europe's AI conversation this year has been dominated by regulation and by the absence of a domestic lab operating at the same tier as the American and Chinese ones. Mistral is the usual answer and the usual caveat.

Buying into a different category rather than chasing the frontier language labs is a defensible way around that. Nobody is currently winning tabular AI in the way OpenAI and Anthropic are contesting text, and SAP already sits on the enterprise data estate where such models would be deployed. Whether a company known for enterprise software can run a research lab at that level is the open question, and four years of funding is a serious attempt at answering it.

The category has been getting attention lately for reasons beyond this deal. As enterprises work through what agents can actually do with internal systems, the limits of text-first models against structured records keep surfacing. We looked at a version of this problem when covering how much supervision agents still need. A model that genuinely understands the shape of a database is one part of closing that gap.

For now the notable fact is the number. More than a billion euros to acquire an 18-month-old company, and more than a billion again to fund it, on a bet that the next useful thing in AI is not conversation.