AI's Quiet Second Frontier: Foundation Models Built for the Data That Actually Runs Your Business

Spend any time following AI in 2026 and you could be forgiven for thinking the whole field is one long argument about chatbots. Which lab shipped the smartest model this fortnight. Whose reasoning is deeper. Whose tokens are cheaper. It is a loud, fast, and genuinely important race.

It is also not the only one.

Off to the side, largely out of the headlines, a second frontier has been forming — one aimed not at language but at the least glamorous and most valuable data most companies own: the rows and columns sitting in their databases, spreadsheets, and warehouses. The models built for it have a clunky name — tabular foundation models, or more broadly, structured-data models — and in the first half of 2026 they stopped being an academic curiosity and became a competitive market. The clearest signal came in May, when enterprise-software giant SAP agreed to buy a small German lab called Prior Labs and committed to invest more than one billion euros over four years to scale it. The deal is expected to close this quarter.

If you have never heard of Prior Labs, that is rather the point.

The model that skips training

Prior Labs makes a model family called TabPFN, and to understand why SAP paid frontier-lab money for it, you have to understand how strange it is.

Most machine learning on tabular data works like this: you take your dataset — say, ten thousand loan applications with twenty columns each — and you train a fresh model on it, tuning knobs for hours to squeeze out accuracy. Do it again for a new dataset, and you start over. This is the workhorse behind a huge share of real-world prediction: fraud scoring, churn, demand forecasting, medical risk. It is also slow, fiddly, and bespoke every single time.

TabPFN throws that loop out. It is a transformer — the same broad architecture behind large language models — but instead of being trained on your data, it was pretrained once on a vast collection of synthetic tabular datasets, generated from structural causal models. In the process it learned something more general than any single task: how tabular problems tend to look. Show it a new dataset and its labels, and it makes predictions in a single forward pass, no training run required. The technical trick is a two-way attention mechanism where each cell attends both to other features in its row and to other examples in its column, which makes the model indifferent to the order of your rows and columns — exactly the symmetry real tables have.

The original TabPFN v2 was striking enough to land in Nature in early 2025. Its headline result read like a typo: in 2.8 seconds, it matched or beat an ensemble of the strongest classical methods that had been tuned for four hours. But it had a ceiling — it worked best on small datasets, up to roughly ten thousand rows, which is a poor fit for the enterprise, where a million rows is a Tuesday.

That ceiling is what fell away in 2026. TabPFN-3, released May 12, scales the same approach to datasets with up to a million training rows while staying practical on a single H100 GPU, using a reduced memory cache and row-chunked processing to keep the math tractable. On TabArena, a standard benchmark for this world, a single forward pass of TabPFN-3 outperforms every other entry and sits on the best point of the speed-versus-accuracy curve. Founders Frank Hutter, Noah Hollmann, and Sauraj Gambhir had turned a clever paper into something an enterprise could actually deploy — and SAP noticed.

Not one model, a category

Here is the part that makes this a trend and not just an acquisition: Prior Labs is not alone.

By 2026 the structured-data frontier has fractured into a small, competitive field. There are tabular models that read a single flat table — TabPFN, but also NEXUS from Fundamental, NICL from France's Neuralk AI, TabICL, and CARTE. There is a separate breed of relational foundation models, led by Kumo.ai's KumoRFM, that read many connected tables at once — think five to fifty tables joined by foreign keys — and discover patterns that hop across those boundaries, the way an actual business database is shaped. There are time-series foundation models like Chronos and TimesFM for forecasting. A Chinese effort, LimiX, reframes the whole problem as modeling the joint distribution over a table's variables and its missing values, and even proposes a scaling law for the category. Amazon's Mitra takes yet another angle with curated synthetic priors. Google has entered with a zero-shot tabular model of its own.

The geography is telling too: Germany, France, Canada, China, and the United States all have serious entrants. This is not one company's bet. It is a field deciding, more or less at once, that structured data deserves its own foundation models rather than being awkwardly force-fed to a chatbot.

Why the boring frontier might matter more

It is worth asking why a company like SAP — whose entire business is the structured data that runs supply chains, payrolls, and ledgers — would spend a billion euros here rather than on a language model.

The answer is a quiet critique of the LLM era. Large language models are extraordinary at text, but ask one to reason over a table of numbers and it is working against its own grain, predicting plausible-looking tokens rather than computing an answer. It can hallucinate a figure with total confidence. Structured-data models are built for the opposite virtues: deterministic, repeatable predictions with clear feature attribution — outputs you can audit, defend to a regulator, and reproduce next quarter. For a bank deciding on a loan or a hospital triaging risk, that auditability is not a nice-to-have. It is the whole job.

There is a thread here that connects to a story we have been tracking all year. Enterprises spent 2025 and early 2026 pouring money into LLM pilots and are now, in the great cost reckoning, asking the uncomfortable question of what actually paid off. A model that skips the training loop, runs on one GPU, and produces auditable numbers is a very different value proposition from a chatbot metered by the token. It may turn out that the most durable enterprise AI of this era was never the one that talked back.

Something to think about

The chatbot race will keep making the headlines, and it should — it is genuinely reshaping how we write, code, and search. But the history of computing is full of moments where the loud revolution and the important one were not the same thing.

So here is the question worth sitting with: if the last three years taught machines to be fluent in our language, the next few may teach them to be fluent in our data — the ledgers, the sensor logs, the customer tables that quietly run the world. That frontier does not demo well. It will never trend. But it is aimed squarely at the numbers your business actually lives and dies by. Keep one eye on the chatbots. Keep the other on the spreadsheets.

Sources

SAP Newsroom, "SAP to Acquire Prior Labs," May 4, 2026: https://news.sap.com/2026/05/sap-to-acquire-prior-labs-establish-frontier-ai-lab-europe/

Prior Labs, TabPFN-3 Technical Report, arXiv:2605.13986, May 12, 2026: https://arxiv.org/abs/2605.13986

Hollmann et al., "Accurate predictions on small data with a tabular foundation model," Nature, 2025: https://www.nature.com/articles/s41586-024-08328-6

TabArena benchmark: https://tabarena.ai/

LimiX, arXiv:2509.03505: https://arxiv.org/abs/2509.03505

Kumo.ai, KumoRFM: https://kumo.ai/

Google Research, "Introducing TabFM: A zero-shot foundation model for tabular data": https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/

All figures verified against primary sources as of July 1, 2026.