SAP Drops $1.16B on an 18-Month-Old Lab
SAP just announced it's acquiring Prior Labs and committing €1B (about $1.16B) over four years to build it into "a globally leading frontier AI lab in Europe." Prior Labs is 18 months old. Three founders — Frank Hutter, Noah Hollmann, Sauraj Gambhir. They built TabPFN, a foundation model not for text but for tabular data. Tables. The boring spreadsheets and database rows where every actual enterprise lives.
Here's what makes this structurally interesting. The text-foundation-model arms race is now seven big labs deep and pricing is collapsing. The image and video model races have similar shape. But there is no foundation model arms race for structured data, because the assumption was you didn't need one — gradient-boosted trees were good enough. Prior Labs called bullshit on that, published TabPFN in Nature, set state-of-the-art on hundreds of independent academic studies. SAP looked at the asset, looked at the fact that ERP runs on tables, and wrote the biggest check.
This is the second SAP move on AI infrastructure this quarter. They also acquired Dremio. Plus the Joule + Nvidia Agent Toolkit alignment from March, which means SAP customers will get authorized access to Nvidia NemoClaw — Nvidia's enterprise OpenClaw competitor. Connect the dots: SAP wants to be the data-and-agent rail for every Fortune 500 ERP customer, and they're paying for it in cash plus four-year R&D commitments rather than partnering. Buy the model lab, buy the data plane, ride the Nvidia agent runtime.
The bet to watch. If TabPFN-class structured-data foundation models actually work at enterprise scale, the moat for ERP vendors flips. Right now SAP/Oracle/Workday compete on data gravity — your records are locked in their schema. With a structured-data FM, the records become exportable and the schema becomes interpretable by any model. That should weaken incumbent moats. Unless an incumbent buys the model lab first. Which is what SAP just did.
Closing Q2 or Q3 2026, regulatory approval pending. Prior Labs current trending repo: github.com/PriorLabs/TabPFN. SAP statement: news.sap.com/2026/05/sap-to-acquire-prior-labs-establish-frontier-ai-lab-europe
← Back to all articles
Here's what makes this structurally interesting. The text-foundation-model arms race is now seven big labs deep and pricing is collapsing. The image and video model races have similar shape. But there is no foundation model arms race for structured data, because the assumption was you didn't need one — gradient-boosted trees were good enough. Prior Labs called bullshit on that, published TabPFN in Nature, set state-of-the-art on hundreds of independent academic studies. SAP looked at the asset, looked at the fact that ERP runs on tables, and wrote the biggest check.
This is the second SAP move on AI infrastructure this quarter. They also acquired Dremio. Plus the Joule + Nvidia Agent Toolkit alignment from March, which means SAP customers will get authorized access to Nvidia NemoClaw — Nvidia's enterprise OpenClaw competitor. Connect the dots: SAP wants to be the data-and-agent rail for every Fortune 500 ERP customer, and they're paying for it in cash plus four-year R&D commitments rather than partnering. Buy the model lab, buy the data plane, ride the Nvidia agent runtime.
The bet to watch. If TabPFN-class structured-data foundation models actually work at enterprise scale, the moat for ERP vendors flips. Right now SAP/Oracle/Workday compete on data gravity — your records are locked in their schema. With a structured-data FM, the records become exportable and the schema becomes interpretable by any model. That should weaken incumbent moats. Unless an incumbent buys the model lab first. Which is what SAP just did.
Closing Q2 or Q3 2026, regulatory approval pending. Prior Labs current trending repo: github.com/PriorLabs/TabPFN. SAP statement: news.sap.com/2026/05/sap-to-acquire-prior-labs-establish-frontier-ai-lab-europe
Comments