The deal is done. This month, SAP completed its acquisition of Prior Labs, the Freiburg-based pioneer of tabular foundation models, with regulatory approvals secured and the lab now operating inside SAP — announced May 4, 2026, closed roughly ten weeks later, with a commitment of more than €1 billion over four years to scale it into what SAP calls a globally leading frontier AI lab.
Here is why this column cares, and why you should: it is the most instructive European AI transaction of the year, and almost nothing about it looks like the AI industry’s main storyline.
€1 billion for the boring data.
SAP × Prior Labs is closed.
The Freiburg lab behind TabPFN — tabular foundation models, published in Nature — is now inside SAP, with €1B+ committed over four years. Not chatbots: the rows and columns that run every business.
| customer_id | invoices | days_overdue | region | churn_risk ← TFM |
|---|---|---|---|---|
| 10441 | 38 | 12 | DE-BY | 0.81 |
| 10442 | 112 | 0 | FR-IDF | 0.07 |
| 10443 | 9 | 44 | DE-BW | 0.93 |
A tabular foundation model reads the table whole at inference and predicts in one pass — no per-dataset training, no hand-tuned gradient-boosted trees. Reported: seconds against four-hour tuned ensembles.
18 months, start to €1B lab
Research → Nature → company → billion-euro lab, without leaving Baden-Württemberg. Purchase price undisclosed; the €1B is committed investment, not price.
Bull
A European champion anchored at home. Open TFM weights small enough for local inference. Peer-reviewed edge in the one modality LLMs handle worst — and where SAP’s customer base lives. Independence, Freiburg base, and open-source direction committed; advisory board includes Yann LeCun.
Bear
Every preservation promise is still a promise — enterprise acquirers have a mixed record on lab autonomy. €1B is commitment, not disbursement. Category now contested: hyperscalers moving in, Fundamental’s $255M Series A. The 24-month test: still publishing openly, or a proprietary Business Data Cloud feature?
Handbook of Mathematical and Digital Engineering Foundations for Artificial Intelligence: A Systems Methodology (Systems Innovation Book Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The least glamorous modality wins the biggest German check
While the industry argued about chatbots, Prior Labs built models for rows and columns — the ERP tables, financial records, supply-chain logs, and customer databases where most enterprise value actually sits, and where large language models are famously weak. SAP’s own announcement is blunt about the gap: LLMs have only a rudimentary understanding of tables, numbers, and statistics.
Prior Labs’ answer is the Tabular Foundation Model (TFM). Its TabPFN series — Prior-data Fitted Networks — is pretrained once on synthetic data, then reads a real table at inference time and predicts immediately, no per-dataset training run. The peer review is real: the TabPFN work was published in Nature in early 2025 and has set the state of the art on tabular benchmarks across hundreds of independent academic studies. Coverage of the current TabPFN-2.6 generation reports single-pass results comparable to hours of tuned AutoML pipelines — the recurring framing is seconds against a four-hour ensemble.
For a decade, the honest answer to “what beats XGBoost on business tables?” was “carefully tuned XGBoost.” That era appears to be ending, and it ended in Freiburg, not San Francisco.
AI for Data Analytics: A Practical Guide to Applying Machine Learning and Generative AI for Better Decisions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The 18-month arc Europe keeps saying is impossible
The company timeline is the story European policy papers have been trying to will into existence for years: founded late 2024 by Frank Hutter, Noah Hollmann, and Sauraj Gambhir out of the University of Freiburg; a €9M pre-seed (Balderton, with XTX Ventures) announced February 2025 — its only funding round; Nature publication; open-source models with a serious community; definitive agreement with the world’s largest ERP vendor within 18 months of founding. Purchase price undisclosed, before and after close.
Research → Nature → company → billion-euro-backed lab, inside two years, without leaving Baden-Württemberg. Whether that becomes a template or stays an exception is exactly the question — but it is the strongest empirical data point European deep tech has produced in a while.
Big Data: A Revolution That Will Transform How We Live, Work, and Think
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What SAP is actually assembling
This wasn’t an isolated purchase. The same week, SAP announced the acquisition of Dremio, the data-lakehouse company — and SAP had already shipped its own first tabular model, SAP-RPT-1, before buying the team that defined the category. Reported integration targets: SAP AI Core as the serving layer, Business Data Cloud as the data fabric absorbing Dremio.
Read together, the strategy is legible: SAP is buying the structured-data layer of enterprise AI — the part where its customer base (finance, manufacturing, healthcare, industrials) lives, and the part hyperscalers haven’t locked up. Microsoft, Google, and AWS are all reported to be moving into structured-data models; Fundamental, a US competitor, raised a $255M Series A — without, so far, peer-reviewed benchmarks. Tabular AI stopped being a quiet research category this year.
Compact Local AI Server, AI Mini PC，Serve Local LLM Models Right Out of Box, 30+ Tokens/Second, Pre-Installed Ubuntu Linux, Qwen3, LLama3, RAG, OCR, vLLM, TensorRT LLM, NVIDIA RTX 5060 Ti (16GB)
- Easy Setup in 3 Steps: Power, connect, scan QR code
- Pre-Installed Local LLM Models: QWen3, LLama3, embeddings, rerankers
- Supports Multiple AI Frameworks: vLLM, TensorRT LLM, RAG, OCR
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The honest ledger
Bull: A European champion anchored at home rather than merged away; open-source TFM weights small enough to run on hardware you own — a specialist model class that fits the local-inference stack in a way no frontier LLM does; peer-reviewed superiority in a category with immediate enterprise revenue attached; and SAP’s stated commitment that Prior Labs keeps its brand, Freiburg base, independent operation, open-source direction, and an advisory board that includes Yann LeCun.
Bear: Every one of those preservation promises is, today, a promise — large enterprise-software acquirers have a mixed record on research-lab autonomy, and the verification only comes in the post-close years. The €1B is a four-year investment commitment, not a disbursed sum. Integration into product cycles is precisely the force that slows research velocity. And the category is now contested by companies with hyperscaler distribution.
The test worth writing down: in 24 months, is Prior Labs still publishing openly, still releasing weights, still in Freiburg — or is TabPFN a proprietary feature of Business Data Cloud? The founders say the former. The deal structure permits either.
Why this belongs in this week’s arc
Monday’s OCR story and this one are the same story. A sub-scale specialist model — 0.9B parameters there, a table-reading PFN here — beating general-purpose giants at a defined, economically dense task, built in the open, cheap enough to run locally. The difference is that this time, a European incumbent paid a billion euros to own the thesis. The market is telling you where the value migrated: not to the biggest model, but to the right-shaped one.
Sources
- SAP SE, “SAP to Acquire Prior Labs to Establish a Globally Leading Frontier AI Lab in Europe,” press release, Walldorf/Freiburg, May 4, 2026 — definitive agreement, €1B+/4-year commitment, SAP-RPT-1 lineage, independence commitments, pending regulatory approval at announcement
- The Next Web, “SAP’s EUR 1bn bet on AI is about tables, not chatbots” (July 2026) — deal closed, regulatory approvals secured, Freiburg lab operating inside SAP; brand/base/open-source/LeCun advisory board retained
- Tech Funding News, “From research project to €1B-backed AI lab in 18 months” (July 2026) — close confirmed; €9M total prior funding; price undisclosed post-close; Fundamental $255M comparison; Dremio parallel acquisition
- Sifted, “SAP agrees to acquire German AI startup Prior Labs” (May 5, 2026) — €9M early-2025 raise, undisclosed sum, Balderton commentary
- The Next Web, “SAP acquires Prior Labs to build a European frontier AI research lab” (May 4, 2026) — pre-seed led by Balderton with XTX Ventures; integration-risk and European-pattern analysis; Constellation Research data-platform framing
- Tech Startups, “SAP acquires Prior Labs in €1B to build next-gen AI for business data” (May 4, 2026) — founding late 2024, founders, TabPFN-2.6 single-pass vs hours-long AutoML claim, AI Core / Business Data Cloud integration targets, H2 2026 expected close
- Prior Labs / TabPFN: Nature publication (early 2025); state-of-the-art claims per SAP release citing hundreds of independent academic studies — vendor-adjacent framing, though the Nature peer review and independent benchmark literature substantially support it
Figures dated as cited; the €1B is a committed investment over four years, not a disclosed purchase price.
