SAP Buys Prior Labs in €1B+ Bet on Tabular AI

Europe’s largest software company is making a nine-figure wager that the future of enterprise AI runs through spreadsheets, not chatbots. SAP has completed its acquisition of Prior Labs, a German startup that pioneered tabular foundation models (TFMs), and committed to investing more than €1 billion over the next four years to scale it into what SAP calls a globally leading frontier AI lab for structured data. The deal closed in July 2026, roughly two months after it was first announced in May—and just 18 months after Prior Labs was founded.

Terms of the transaction were not disclosed, but the scale of the surrounding commitment and the speed of the exit make it one of the most notable European AI acquisitions of the year.

What Prior Labs built

Most of the AI attention over the past two years has gone to systems that generate language, images, and code—models built on the transformer architecture and trained on internet-scale text. Prior Labs went after a different, quieter problem: the structured, tabular data that actually runs businesses—the rows and columns in enterprise databases covering customers, payments, suppliers, and inventory.

Its flagship model series, TabPFN, was published in Nature and set the state-of-the-art on tabular benchmarks across hundreds of independent academic studies. The core idea is a break from how machine learning on tables usually works. Traditionally, an organization trains a separate model for every dataset—one for churn, another for demand forecasting, another for fraud—each requiring its own data pipeline, tuning, and maintenance. TabPFN instead uses a single pre-trained foundation model that can solve prediction tasks directly on structured data without that per-dataset training step.

In practice, that means one model aimed at tasks such as payment delays, customer churn, supplier risk, and demand forecasting—the bread-and-butter predictions that enterprises run constantly and that today require armies of data scientists to build and babysit. It’s a foundation-model approach applied to the world of columnar and row-oriented databases rather than to natural language.

Why SAP wanted it

SAP’s entire business sits on top of enterprise structured data. Its ERP systems are the system of record for a large share of the world’s largest companies—the tables where transactions, ledgers, and supply-chain records live. A foundation model that can make accurate predictions directly on that kind of data, without bespoke training for each customer, maps almost exactly onto SAP’s installed base.

That strategic fit is the logic behind the price. Rather than acqui-hire the team and fold it into a product group, SAP says Prior Labs will continue to operate as an independent entity, with the €1 billion-plus commitment funding it as a research lab. The framing is deliberate: SAP is positioning this less as a feature purchase and more as an attempt to build a European frontier AI lab, a category the continent has largely ceded to U.S. and Chinese firms.

For a company that has spent the LLM era partnering with outside model providers rather than building frontier systems itself, owning a genuinely state-of-the-art model franchise—even a specialized one—is a meaningful shift in posture.

An 18-month exit

The timeline is its own headline. Prior Labs went from founding to a €1 billion-backed exit in 18 months, an exceptionally fast trajectory even by the standards of the current AI funding boom. The startup emerged from academic research—the TabPFN work that ran in Nature—and converted a benchmark-topping model into a strategic acquisition target before most enterprise buyers had a tabular-AI strategy at all.

That speed reflects the broader market. The first half of 2026 set records for venture funding, with capital concentrating in AI, and strategic acquirers increasingly competing with venture firms to lock up scarce model talent early. Prior Labs is being held up across European tech press as a new benchmark for how quickly a deeptech research project can become a €1 billion asset.

Tabular models versus the LLM mainstream

It’s worth being precise about what SAP did not buy. Tabular foundation models are not language models pointed at spreadsheets. They are a distinct model class trained specifically on the statistics of structured data, and their advantage shows up on exactly the tasks where general-purpose LLMs tend to be unreliable: producing calibrated numerical predictions from a company’s own tables.

That specialization is the bet. General LLMs are extraordinary at open-ended language tasks but awkward at, say, forecasting next quarter’s demand from a messy sales table—the kind of task that usually still gets solved with gradient-boosted trees and a data scientist’s fine-tuning. If a single pre-trained TFM can do that work out of the box, it collapses a large amount of custom modeling into a foundation-model call. For an enterprise vendor, that is a compelling pitch.

What it means

Strip away the frontier-lab language and this is a bet on where enterprise AI value actually accrues. The consumer-facing AI story is dominated by chat and code, but most of what large companies pay for is prediction on their own structured data—and that is a market general-purpose LLMs have not convincingly captured. SAP is trying to own the model layer for it.

Who wins. SAP, if the thesis holds, gets a defensible AI differentiator wired directly into the ERP systems its customers already run—something it can offer that a general model provider cannot easily replicate without the same tabular research. Prior Labs’ founders and backers win a €1 billion-scale outcome in 18 months. And European deeptech gets a marquee example that a homegrown research project can command frontier-lab money without relocating to the U.S.

Who’s on notice. Every vendor selling per-dataset predictive modeling—and every data-science team whose job is building one churn or forecasting model at a time—should read the TabPFN thesis carefully. If foundation models for tables mature the way foundation models for text did, the economics of bespoke tabular ML change substantially. The competitive question is whether rival enterprise-software giants respond by building or buying their own tabular-AI capability.

What to watch. Three things. First, whether SAP keeps Prior Labs genuinely independent or quietly absorbs it—the difference between a research lab and a product acquisition. Second, whether the €1 billion actually materializes as sustained research funding across the full four years, or gets trimmed if enterprise-AI budgets tighten. Third, whether tabular foundation models prove out at production scale on real, messy corporate data—benchmark dominance and boardroom results are not the same thing.

For now, the signal is clear: in mid-2026, “frontier AI” no longer means only bigger chatbots. A €1 billion commitment aimed squarely at the structured data inside enterprise databases is a bet that the next competitive front runs through the least glamorous, most valuable data companies own.