SAP announced on May 4 that it will spend more than €1 billion over four years to acquire Prior Labs, an 18-month-old German research shop built around tabular foundation models.[^1] The deal is a direct bet that enterprise AI value lies in models trained on structured ERP tables, not in the horizontal LLM APIs that most vendors currently rent.
The Deal: €1B Over Four Years for an 18-Month-Old Lab[^1]
The announcement itself committed ”>€1 billion ($1.16B) over four years” but kept the actual acquisition price undisclosed.[^1] TechCrunch reported that sources put the upfront cash at “well over half a billion dollars” and that the structure was “almost all cash.”[^2] The expected close is Q2 or Q3 2026, pending regulatory approval.
Prior Labs was founded in late 2024 in Freiburg and had raised roughly €9 million in a pre-seed round led by Balderton Capital in February 2025.[^2] The jump from a €9 million pre-seed to a reported half-billion-plus exit in under 16 months is unusual even by AI lab standards. The lab will operate as an independent unit within SAP.
What Prior Labs Built: From Nature Paper to TabPFN-2.5
The research lineage begins with the original TabPFN, published in Nature in January 2025,[^3] which showed that a single pre-trained foundation model could outperform specialized methods on tabular classification without task-specific retraining. The follow-up, TabPFN-2.5,[^4] scales to 50,000 rows and 2,000 features and claims a 100% win rate against default XGBoost on small and medium datasets, dropping to an 87% win rate on larger sets up to 100,000 samples.[^4]
The models have been downloaded more than 3 million times.[^1] Prior Labs’ scientific advisory board includes Yann LeCun and Bernhard Schölkopf,[^1] and its researchers were recruited from Google, Apple, Amazon, Microsoft, Jane Street, Goldman Sachs, and CERN.
Why SAP Thinks TFMs Beat LLMs for Structured ERP Data
SAP CTO Philipp Herzig stated that “the greatest untapped opportunity in enterprise AI wasn’t large language models; it was AI built for the structured data that runs the world’s businesses.”[^1] That framing positions Prior Labs as the specialized engine for ERP tables, supply-chain ledgers, and financial records, while LLMs handle the surrounding prose.
SAP plans to productize the models through SAP AI Core, SAP Business Data Cloud, and its Joule agentic layer, and says it will continue open-source support.[^2]
The Dremio Parallel: Lakehouse + Model = Full Stack
The Prior Labs deal was announced in the same week as SAP’s acquisition of Dremio.[^5] Dremio, founded in 2015 and last valued at $2 billion after raising $410 million, is positioned as the “Apache Iceberg-native enterprise lakehouse” inside SAP Business Data Cloud.[^5] Most coverage treats the two deals as separate bets. Together they read as a full-stack play: Dremio handles storage and query semantics; Prior Labs handles inference on the tabular output.
What It Means for Workday, Oracle, and Salesforce
If SAP’s bet pays off, the competitive burden shifts. Rivals with proprietary structured-data stacks, including Workday, Oracle, and Salesforce, face a choice: build or acquire their own tabular foundation models, or concede the structured-data inference layer to SAP and continue renting horizontal LLM APIs that were not designed for relational schemas.
The economics differ in ways that matter to procurement. A vertical model trained on ERP schemas amortizes its cost across deployments. An LLM API metered by token volume does not price differently for structured records than for prose, which changes the unit economics for large table workloads.
Limits and Open Questions
TabPFN-2.5 caps at roughly 50,000 rows and 2,000 features.[^4] Many ERP tables exceed those bounds, so production scaling likely requires distillation, chunking, or architectural changes not yet detailed. GPU requirements for inference also contrast with the CPU-friendly footprint of gradient boosting, which could matter for cost-sensitive batch jobs.
Frequently Asked Questions
What is SAP’s ‘NemoClaw’ policy and how does it affect agentic AI integrations?
SAP adopted a ‘NemoClaw authorized, OpenClaw blocked’ stance alongside the Prior Labs acquisition—a framework permitting only vetted agentic AI integrations while blocking unrestricted third-party agent access to its data layer. That defensive posture may conflict with Prior Labs’ open-source community expectations and could become a competitive differentiator if Oracle or Salesforce offer less restrictive agent ecosystems.
How unusual is the Prior Labs exit by European AI-deal standards?
A reported half-billion-plus in upfront cash for a company roughly 16 months from founding with only €9M in prior backing ranks among Europe’s largest AI acquisitions by return multiple. Most European AI exits above €500M involved companies with substantially more venture capital and multi-year operating histories—the speed and capital efficiency here is an outlier even by 2025–26 AI-market standards.
Why did a tabular ML model land in Nature rather than a ML conference?
The original TabPFN was published in Nature Volume 637 on January 9, 2025 rather than NeurIPS or ICML, framing tabular foundation models as a general scientific advance rather than a niche ML contribution. That venue choice gave Prior Labs credibility with enterprise buyers like SAP who recognize Nature’s prestige but may not follow conference proceedings—a strategic signal distinguishing the lab from the hundreds of LLM-focused startups competing for acquisition.
What should teams running XGBoost on SAP data do while productized TabPFN has no shipping date?
Benchmark current pipelines against the open-source TabPFN-2.5 on representative SAP schemas, respecting the 50,000-row ceiling. Teams processing larger tables should prototype chunked-ensemble approaches now, since SAP has not announced a timeline for the scaled productized versions through AI Core and Joule. Budget for GPU hardware or cloud accelerator instances that existing XGBoost workflows do not require.