SAP's €1B Bet on the AI LLMs Can't Do

SAP signed a definitive agreement this week to acquire Prior Labs, an 18-month-old German AI lab, and committed more than €1 billion over four years to scale it. The official terms are in SAP’s May 5 newsroom announcement, and they describe a strategy bet most enterprise AI buyers haven’t fully priced in yet. Prior Labs builds Tabular Foundation Models. Not LLMs. A different category of AI, purpose-built for the kind of data that actually runs your business.

This is the part that should make every CIO who has been pouring AI budget into ChatGPT-shaped pilots stop and reread the press release.

For two years, the dominant assumption inside enterprise AI procurement has been that frontier LLMs are the universal hammer. Customer support? LLM. Sales forecasting? LLM. Churn prediction? LLM. Payment-delay risk on a 4-million-row receivables ledger? Also LLM, somehow. SAP just spent over a billion euros to publicly disagree with that last one.

Quick Verdict

The Move	What It Means for You
SAP acquires Prior Labs (May 4-5, 2026)	An ERP giant just declared LLMs the wrong tool for half of enterprise AI
€1B+ committed over 4 years	A capital-backed bet that Tabular Foundation Models are a real AI category
Prior Labs’ TabPFN published in Nature	State-of-the-art on hundreds of tabular benchmarks, peer-reviewed
Lab founded ~18 months ago	Validates TFMs as serious infrastructure, not a research curiosity
SAP also acquired Dremio (data lakehouse)	Storage layer plus prediction layer pulled into one stack
92% of Fortune 500 runs SAP	The capability will auto-appear in workflows your finance team already uses
Use case scope	Payment delays, supplier risk, churn, upsell, demand forecast
Your real lever this week	Audit which of your current “LLM projects” are actually tabular prediction problems

The AI Tool Most Enterprises Are Using Wrong

Here’s the framing problem. An LLM is a model trained to predict the next token in a sequence of text. It is very good at that and at the long tail of language tasks that flow from it. It is mediocre to bad at predicting a numerical outcome from a structured table of business data.

You can hand GPT-5 a CSV of 50,000 invoices and ask which ones will be paid late. It will give you an answer. The answer will not be reliable. It will hallucinate patterns, drift on numerical reasoning, and degrade further as the table gets wider. This is documented behavior, not a vibe. Anyone who has tried to run a real prediction workload through a chat interface has felt the failure mode.

The work-around the industry settled on was to bolt LLMs onto traditional ML pipelines. The data team builds an XGBoost or random forest model to do the actual prediction. The LLM writes the natural language explanation on top. That works, but it leaves the hard part, the prediction model, on the data team’s plate. Every new use case is a new feature engineering project, a new training run, a new validation cycle. Months of work per workflow.

A Tabular Foundation Model collapses that pipeline. It is a single model, pre-trained on the patterns of structured data the way an LLM is pre-trained on the patterns of language. You hand it a table and a target column. It returns predictions. No training run. No feature engineering. The model uses in-context learning, the same trick that lets an LLM follow examples inside a prompt, but applied to rows in a table.

That capability is what Prior Labs has been shipping under the name TabPFN, with the underlying research published in Nature in early 2025. Hundreds of independent academic benchmarks now exist. TabPFN is at the top of most of them.

What is a Tabular Foundation Model and how does it differ from an LLM?

A Tabular Foundation Model (TFM) is a pre-trained model purpose-built to make predictions on structured, table-shaped business data (rows and columns, with numerical, categorical, and text features mixed). It uses in-context learning, like an LLM, but the modality is tables instead of language. A user hands the model a table and a target column. The model returns predictions on the unseen rows without any retraining or feature engineering.

Five practical differences from an LLM:

1. Modality. LLMs operate on token sequences. TFMs operate on tables of structured features.
2. Numerical reasoning. LLMs hallucinate on math and aggregation at scale. TFMs are trained on numerical patterns natively.
3. Training requirement. LLMs need RAG or fine-tuning to specialize. TFMs handle new datasets via in-context examples.
4. Use case fit. LLMs win on language tasks. TFMs win on prediction tasks against structured data.
5. Operator access. LLMs are usable by anyone who can write a prompt. TFMs become usable by anyone who can hand over a spreadsheet.

That last point is the one that scales inside an enterprise. Most business users at a Fortune 500 cannot stand up an XGBoost pipeline. Most of them can pull a table out of an SAP module. The TFM rewrites the access pattern for prediction work the way ChatGPT rewrote the access pattern for language work.

Why SAP, Why Now

SAP runs the operating data of 92% of the Fortune 500. ERP, finance, HR, supply chain, procurement. The structured data of global enterprise sits inside SAP modules. That asset has been sitting in plain sight while the entire AI conversation pointed at LLMs.

The acquisition is SAP’s read that the asset is the moat. According to Constellation Research’s coverage of the deals, SAP simultaneously acquired Dremio, the open-source data lakehouse vendor. Dremio handles structured data storage and federated query at enterprise scale. Prior Labs handles structured data prediction at frontier model quality. Pull the two together and SAP is building the storage-plus-prediction stack for the kind of data nobody else has at this scale.

The use cases SAP is naming are not exotic. Payment delays on accounts receivable. Supplier risk scoring. Customer churn prediction. Upsell propensity. Demand forecast accuracy. These are the workloads finance, ops, and sales teams have been hand-rolling in Excel and Power BI for fifteen years. They are also the workloads where LLM-driven attempts have produced the most expensive disappointments inside large enterprises over the last 24 months.

This is the pattern behind a meaningful share of the 95% of AI projects that fail to ship measurable value. The model wasn’t broken. The wrong model was selected for the workload.

The Strategic Misallocation Most Enterprises Made

Sit with this for a second. If you run a $500M revenue business and you funded six AI initiatives in 2025, look at the list. How many of them are language tasks? Drafting, summarizing, search, support deflection, content generation. And how many are prediction tasks against structured business data? Forecasting, scoring, classification, risk ranking.

Most enterprises I’ve watched run that audit find the prediction workloads outnumber the language workloads two to one. Then they find that 90%+ of their AI tooling spend went into LLM platforms. The mismatch is the news.

This is not a critique of LLMs. LLMs are powerful technology and the right tool for a large and growing set of problems. The critique is procurement discipline. Buying one tool and applying it to every workload is a procurement failure, not a technology failure. SAP’s acquisition makes the failure visible because it puts a serious vendor behind the alternative.

The companies that have already started splitting their AI portfolio between language work and prediction work are running ahead. They are using Claude or GPT-5 for the contract review and the customer email drafting. They are using XGBoost, LightGBM, or now a TFM for the demand forecast and the churn score. Different workloads, different tools, different KPIs. That is the model-agnostic posture that holds up over multi-year cycles.

The companies still running everything through ChatGPT Enterprise are paying for one tool to do the work of two and getting the second job done badly.

Why an 18-Month-Old Lab Just Got €1B

Prior Labs was founded in late 2023 by Frank Hutter, Noah Hollmann, and Sauraj Gambhir, a research team out of the University of Freiburg in Germany. According to TechTarget’s reporting on the deal, the team had built a category-defining model with TabPFN before the company hit its second birthday. SAP didn’t buy a product company. SAP bought a research lab and committed €1B to scale it into a frontier AI institution operating from Europe.

A few honest reads on why that price made sense.

Sovereignty. A European frontier AI lab inside a European software giant gives SAP a story it could not buy at any price from a US-acquired research team. The AI Act, GDPR, and the European procurement preference for European AI infrastructure are real procurement levers in 2026. SAP just locked in the European lab credential.

Category creation. Prior Labs is not the only research group working on tabular foundation models, but it is the group most associated with the category. Acquiring the category leader at category birth costs less than acquiring the category leader at category maturity. €1B in 2026 is probably the floor.

Distribution fit. This is the one most enterprise buyers will underestimate. The TFM capability is not going to ship as a separate product nobody buys. It is going to appear inside SAP’s existing modules. A finance director who already opens SAP every morning is going to see “predict payment delay on this receivable cohort” become a button. The capability auto-distributes to the largest installed base in enterprise software. That is the distribution moat the price reflects.

You can run the same logic against IBM’s AI Operating Model pitch at Think 2026 and the Anthropic-Blackstone JV. The frontier labs and the platform vendors are all racing to convert AI capability into installed-base distribution. SAP’s distribution is the deepest. It just bought the model purpose-built to ride that distribution.

What This Doesn’t Mean

A few things this announcement does not do, before the headline reading runs ahead of itself.

It does not make LLMs obsolete. This is the misread to avoid. LLMs are still the right tool for language work, and that workload is enormous and still growing. The TFM is the right tool for prediction work against structured data. Both technologies coexist. Most enterprise AI portfolios should run both.

It does not deliver capability tomorrow. The acquisition is expected to close in Q2 or Q3 2026, and the capability scaling inside SAP modules is a multi-quarter to multi-year program. If your CIO is pricing this into this fiscal year’s roadmap, the timeline is probably wrong.

It does not lock you into SAP. The TFM category is broader than Prior Labs. Open-source TFM work is active. AWS, Google, and Microsoft will respond, probably with their own acquisitions or in-house research investments. The architectural choice you make in 2026 should preserve the ability to run a TFM-style capability across your stack, not just inside SAP.

It does not reduce the urgency on your existing pilots. If you are still mid-cycle on AI pilots that have not produced measurable outcomes, the SAP news doesn’t bail you out. The pilot purgatory roadmap still applies. You still need to ship.

The Strategic Read

Three things this should change about how you read the rest of 2026.

The AI tooling map just split into two columns. Language work belongs in one column with LLMs. Prediction work against structured data belongs in another with TFMs and traditional ML. Any AI strategy document that doesn’t draw this split is going to look dated by Q3. The conversation in board rooms is about to shift from “what’s our AI strategy” to “what’s our AI strategy for language and what’s our AI strategy for structured data,” and the answers are different vendors, different KPIs, and different teams.

Procurement frameworks need a tabular column, not just a chat column. Most AI procurement decks I’ve seen evaluate vendors on language benchmarks. MMLU, HumanEval, agentic tool-use scores. None of those benchmarks tell you whether the model can predict accurately on your invoice ledger. SAP just made tabular benchmarks a procurement-relevant metric. Expect every major vendor RFP from Q3 forward to include a TFM or tabular prediction line item. If yours doesn’t, your evaluation is incomplete.

The competitive map for vertical AI just got more interesting. Vertical SaaS vendors who built moats on structured workflow data (finance, supply chain, insurance, healthcare, real estate) now have a credible path to ship native AI prediction inside their products without an LLM hallucination problem. The vendors who execute this in the next 18 months get a real product wedge. The vendors who keep bolting on a chatbot and calling it AI get caught flat-footed.

This is also the read for enterprise AI ROI. The reason the ROI numbers have been disappointing is that a meaningful share of the spend was misallocated against use cases the chosen tools couldn’t deliver. Reallocating against tool-fit will produce measurable wins that have been sitting in plain sight.

Your Move This Week

Three concrete actions, all doable by Friday. Works whether you’re a Fortune 500 SAP customer, a mid-market business, or an SMB.

1. Audit which of your current “LLM projects” are actually tabular prediction problems. Pull the list of active AI initiatives across your business. For each, write one sentence answering: is the core task language understanding or prediction against structured data? If the answer is prediction, the LLM is probably the wrong tool, and the project’s underperformance is probably tool-fit, not execution. The audit is a one-page artifact. Do it before the next budget cycle. Reallocating tool spend against tool-fit is the highest-impact move available in Q2.
2. Identify three structured-data prediction workloads worth piloting with a TFM-style approach. Payment delay risk, supplier disruption, customer churn, demand forecast accuracy, hiring funnel conversion. Pick the three with the largest dollar impact at your business. You don’t need to wait for SAP to ship the capability inside their modules. Open-source TabPFN is available now. Your data team can run a benchmark in two weeks. The output is a credibility document that gives you a concrete answer when your SAP rep calls in Q3 to demo the new module.
3. Decide your tabular AI posture before your vendors decide it for you. SAP is moving. AWS, Google, and Microsoft will follow with their own TFM stories, probably before year-end. The buyers getting the best results in 2026 run a multi-vendor posture with a default platform and credible swap paths. Make this call deliberately. Pick whether your structured-data AI runs inside your ERP, inside your data platform, or as a separate prediction layer. Each posture has different cost, governance, and lock-in implications. The vendor that demos you the framework will assume the framework runs on their stack. Make that assumption explicit, then decide.

If you’re an SAP shop, expect a conversation from your account team about the new capabilities by Q3 or Q4. The conversation goes better when you’ve already done the audit and know which workloads matter most.

Bottom Line

SAP just spent over a billion euros to say something out loud that most enterprise AI procurement decks still won’t admit. LLMs are the wrong tool for a large fraction of the AI work enterprises are trying to do. The right tool for prediction against structured business data is a different model class, and that class now has a credible vendor with the deepest distribution in enterprise software behind it.

For two years, the AI conversation has been one-modal. Text in, text out, language tasks dominate. The next two years are going to be two-modal. Language work runs on LLMs. Prediction work runs on TFMs and traditional ML. The companies that draw the split early and reallocate budget against it run ahead. The companies that keep buying one tool for every workload pay full price for half-price results.

An 18-month-old lab in Freiburg just got €1B because it built the model purpose-built for the data inside every Fortune 500 ERP system. The acquisition is a category endorsement and a procurement signal at the same time.

Run the audit. Pick the prediction workloads. Decide the tabular posture. The buyers who do this work on their own clock keep optionality. The ones who wait for the SAP demo lose six months and a budget cycle.

---

Related Reading:

• IBM Named the AI Divide. Here’s Which Side You’re On.
• Anthropic’s $1.5B PE Deal Makes AI Non-Optional
• The Enterprise AI ROI Reckoning
• 95% of AI Projects Fail. Here’s How to Be in the 5%.
• Your AI Stack Has an Expiration Date
• Stuck in AI Pilot Purgatory? A 5-Step Roadmap to Production