Large tabular models: A new approach to AI with significant promise

When looking at generative AI (Gen AI), much of the investing world has been focused on unstructured data: emails, documents and videos that are not stored in rigid or specific formats. While this focus makes sense – 80% to 90% of enterprise data is unstructured – there is an opportunity, especially in financial services, for startups focused on extracting value from the remaining 10% to 20% of structured data.

Banks, for example, rely heavily on structured data for detecting fraud, modeling credit risk, complying with regulations and understanding customer behavior. To date, banks’ efforts to extract value from structured data have centered around traditional machine learning techniques like regression, next best action and decision trees. These techniques, while sophisticated, explainable, and proven in production for low latency tasks like fraud monitoring, require bespoke feature engineering and sophisticated drift monitoring systems.

Enter large tabular models (LTMs), which represent a new class of foundation models designed specifically for structured data. Once trained, LTMs can be deployed for multiple tasks, categorize or classify data they have not yet been trained on, and eliminate manual and time-consuming data preparation. Collectively this could help banks unearth predictions in seconds, not hours. Crucially, LTMs deliver deterministic, precise and auditable predictions, unlike their large language model (LLM) cousins that can hallucinate or produce unreliable numerical outputs.

A new crop of startups has risen to take advantage of this burgeoning market: according to Gartner, the TAM for data science and AI platforms will climb to $29 billion in 2028, up nearly 222% from where it stood in 2023. Financial services firms have a host of efficiencies to gain if they can successfully identify, and partner with, leading startups that are removing the roadblocks to broad deployment of LTMs.

What are LTMs and why are they important for banks?

LTMs aim to bring the foundation model paradigm to structured data. Think of LTMs as inspired by LLMs: LTMs apply attention techniques and other tactics perfected for LLMs to models for structured data. While LLMs learn to anticipate the next word in a text sequence, LTMs comprehend the relationships between data points arranged in rows and columns, enabling them to predict missing values, categorize records, and generate forecasts.

The core use case: pre-train a model on diverse tabular datasets so it learns general patterns about data stored in tables, then apply that knowledge to new prediction tasks. So instead of training a new model for individual datasets, LTMs would already be ready to make predictions and “understand” how tables work.

For financial services, this is crucial. Dive into any major bank’s data repositories and you will find petabytes of structured data from transaction records to credit histories. This tabular data, consisting of rows and columns across relational databases, forms the operational backbone of financial services. Deploying LTMs to assess data and harvest insights would be a game changer.

Up to now, however, most banks have relied on machine learning solutions to predict information from tabular data. These solutions require case-by-case training and have limited ability to transfer learning from one data set or spreadsheet to another.

Overall, LTMs could be transformational technology that offer banks significant potential for impact.

The opportunity of LTMs, market fundamentals and key players

LTMs now often surpass established machine learning techniques for small datasets. The TabPFN-2.5 technical report from late 2025 documents that the model defeats standard XGBoost configurations on virtually all small-to-medium classification problems and maintains strong performance on datasets approaching 100,000 samples. Most striking is the speed advantage: default TabPFN settings produce results in under three seconds that would require hours of hyperparameter optimization using conventional methods.

The current groundswell of investors’ thinking around LTM-focused startups is analogous to early interest in LLM providers. Enterprises have realized that the most immediate and impactful AI ROI resides not merely in conversational interfaces, but in profoundly elevating the predictive power of their existing tabular data systems.

Market forces are also coming together to make a case for LTMs: an exponentially expanding analytics platform market ($77B+ by 2028 with 16% CAGR), a ripe foundation model market ($66B by 2029 at 64%+ CAGR), a substantial allocation of enterprise budgets towards foundation models and a persistent, acute global demand-supply imbalance for skilled data science and AI talent.

This surge in demand for analytics and scarce talent creates strong pull for LTMs that operate natively on tabular data and plug into existing BI, ERP and data warehouse workflows. Deploying LTMs would enable teams to ship predictive and prescriptive analytics inside the tools they already use (e.g., dashboards, SQL layers, ELT pipelines) without large greenfield ML programs.

As foundation model budgets ramp and analytics software vendors race to differentiate with AI, LTMs that productize “drop-in intelligence” for tabular workloads will be positioned to capture budget faster than horizontal LLMs. Startups are already acting on powering AI-native BI and operational decisioning for enterprises, although it is early days.

An expanding ecosystem of startups and research institutions are developing LTM technologies, each pursuing distinctive technical strategies and focus areas, including Prior Labs, Kumo, Neuralk-AI, Fundamental Technologies and Wood Wide AI.

Conclusion

LTMs represent a promising opportunity in enterprise AI. While most of the focus has been on Gen AI and LLMs, thanks to the volume of unstructured data, large enterprises have a brilliant opportunity to leverage LTMs thanks to the quantity of structured and tabular enterprise data. Purpose-built foundation models and related startups are emerging to address this need.

For banks, the implications could be significant. LTMs promise to accelerate model development, improve prediction accuracy, and enable new applications that were previously impractical. Early adopters have an opportunity to build competitive advantages that compound over time.