While everyone's been watching chatbots write poetry, they've been utterly useless at the one thing businesses actually need: making sense of their spreadsheets.
The Summary
- Fundamental AI emerged from stealth with $275M and NEXUS, the first large tabular model (LTM) — purpose-built for row-and-column data that LLMs can't handle
- LLMs fail at structured data analysis despite being able to write code, analyze legal docs, and solve complex math problems
- AWS is already adopting NEXUS, signaling that tabular AI could become critical infrastructure for enterprise data analysis
The Signal
Here's the dirty secret of the LLM revolution: the AI that can write a screenplay can't analyze your sales data. ChatGPT will hallucinate confidently about a 50-row spreadsheet. This isn't a minor limitation — it's a fundamental mismatch between how transformers work and how business data is structured.
Fundamental AI's $275M raise and launch of NEXUS represents the first serious attempt to build foundation models for the data format that actually runs the world. Bank transactions, clinical trials, website analytics, particle physics — the critical information isn't in prose. It's in tables. And until now, there's been no model architecture designed to understand them.
"People like to see images, videos, and ChatGPT responses. But tabular data really lags behind because it's not fun to look at numbers." — Boris van Breugel
The technical problem is deeper than attention spans. LLMs work because language has universal structure — grammar, semantics, common patterns across billions of documents. Tabular data doesn't. Every spreadsheet has different columns, different meanings, different relationships between fields. A "customer_id" in one dataset has nothing to do with "customer_id" in another. LLMs trained on text can't transfer that learning to tables where the schema itself is the signal.
Why this matters now:
- Most enterprise AI use cases require structured data analysis, not text generation
- Current workarounds (converting tables to text for LLMs) lose critical relationships and context
- AWS adoption suggests LTMs could become the missing layer in enterprise AI stacks
Fundamental's timing is deliberate. The LLM hype cycle focused on consumer-facing applications — chatbots, image generators, coding assistants. But the real money in AI isn't in generating content. It's in analyzing the massive structured datasets that companies already have. Financial services need to spot fraud patterns across millions of transactions. Healthcare needs to find signals in clinical data. E-commerce needs to optimize across countless SKUs and customer behaviors.
NEXUS is designed for this. The architecture treats tables as first-class data structures, understanding relationships between columns, temporal patterns in rows, and the statistical distributions that define different fields. It's not trying to force tabular data into a text paradigm. It's built from the ground up for row-and-column reasoning.
The enterprise implications are immediate. If LTMs work as advertised, they collapse the gap between "having data" and "understanding data." Right now, most companies sit on warehouses full of structured information they can't fully analyze without armies of data scientists. An LTM that can reason across tables, join disparate datasets, and surface non-obvious patterns changes the economics of business intelligence entirely.
The Implication
Watch for the LTM layer to emerge in enterprise AI stacks over the next 12 months. If NEXUS proves out with AWS and early adopters, expect every major AI lab to release tabular models. The companies that win won't be the ones with the best chatbots — they'll be the ones that can actually make sense of their data.
For builders: tabular data is the least sexy, highest-value AI problem right now. While everyone else chases the next GPT wrapper, there's a wide-open opportunity in tools, infrastructure, and applications built on LTMs.