Fundamental's NEXUS Large Tabular Model: LLMs Stumble on Spreadsheets

On February 5, 2026, startup Fundamental came out of stealth with $275 million in funding. The company introduced NEXUS, the first Large Tabular Model (LTM), built specifically to analyze structured data in tables.

Why LLMs struggle with tables

Models such as ChatGPT, Claude and Gemini produce convincing text and images, but ask them to analyze even a small table and they look lost.

• LLMs are designed for sequential data: every word affects the next, and order is critical
• Tables do not depend on order: you can swap columns or rows and the meaning does not change
• LLMs give different answers when the input changes even slightly (usually that is desirable), but fraud prediction requires determinism
• For most companies, the most important data lives in spreadsheets: bank transactions, sales logs, website metrics

How NEXUS works

Unlike LLMs, which model a sequence of tokens, an LTM models the structure of tables directly. The model learns from three sources at once: a numerical value, its meaning, and its statistical relationships with the other data in the column.

Companies need deterministic, consistent predictions.

With LLMs that is impossible, says Jeremy Fraenkel, CEO of Fundamental.

Fundamental trained NEXUS on billions of tables, using a mix of licensed datasets, open sources and synthetic data. The company owns proprietary datasets through partnerships and does not train on customer data. NEXUS runs on confidential computing environments, so Fundamental physically cannot access user data.

Who else is building LTMs

In June, Amazon integrated NEXUS into SageMaker, the standard machine learning tool on AWS. That gives the model access to sensitive customer data that companies manage locally.

In March 2026, Feedzai (financial security) and Mastercard launched their own LTMs to fight fraud. In late June, Google unveiled TabFM, trained on hundreds of millions of synthetic tables. Researchers have developed FlexTab, TabICL and iLTM, and LTMs are becoming a trend.

What this means

Tables are the language of business. For 50 years, companies have relied on XGBoost and other gradient-boosted decision trees for analysis, and these had to be tuned for months for each task. LTMs promise to make tabular analysis as automated as text generation. LLMs and LTMs together could produce a more powerful AI: a left brain for text, a right brain for numbers.