SAP closes Prior Labs: AI for your spreadsheets, not chat

SAP closed its acquisition of Prior Labs last week, and the number attached is over €1 billion across four years. Not for a chatbot. For a foundation model that predicts from rows and columns. If you've been trying to get an LLM to forecast your inventory or score your leads and watching it confidently invent numbers, this is the category you actually wanted — and the model is open source.

What actually happened

SAP announced the deal on May 4 and, per The Next Web, completed it on July 17 after regulatory clearance. Prior Labs is a German research company out of Freiburg that pioneered Tabular Foundation Models — TFMs. Its model series, TabPFN, was published in Nature, and SAP says it has over 3 million downloads and set the state of the art on tabular benchmarks across hundreds of independent academic studies.

The structure matters more than the price. Prior Labs stays a separate legal entity, keeps its brand, its Freiburg headquarters and Berlin and New York offices, its existing customer relationships, and — this is the part for you — its open-source commitments. Its scientific advisory board includes Yann LeCun and Bernhard Schölkopf.

A TFM does what an LLM is bad at. Instead of generating plausible prose, it takes a table and predicts a column: risk score, failure date, classification, forecast. TNW cites loan approvals, predictive maintenance on rail, and cancer diagnostics as live applications.

Why tabular foundation models matter for your business

Almost every useful prediction a small business wants is a table problem. Which invoices go past 60 days. Which quotes close. Which SKU runs out in three weeks. Which service call comes back. That's rows and columns — orders, tickets, line items — not paragraphs.

For two years the default advice has been "throw it in the context window and ask the LLM." It kind of works and it's expensively wrong in ways you can't audit. A model built for tabular prediction is smaller, faster, cheaper per inference, and — the underrated part — it gives you a calibrated probability instead of a sentence that sounds sure.

The move here isn't to wait for SAP to ship this inside S/4HANA in 2028. TabPFN is on PyPI today. If you have a few thousand rows of clean historical data with a known outcome column, you can test whether it beats your current guess this week, on your own hardware, without an enterprise contract. That's a genuinely small experiment with a real answer at the end.

The prerequisite is boring and it's the actual bottleneck: your data has to be in a table with consistent columns and labeled outcomes. Most businesses we work with have that data spread across a POS export, a CRM, and someone's spreadsheet. Fixing that is the project. The model is the easy part.

Sources: SAP News — SAP to Acquire Prior Labs, The Next Web — SAP's €1bn bet on AI is about tables, not chatbots.