Fundamental emerged from stealth today with $255 million in Series A funding from Oak HC/FT, Valor, Battery Ventures, and Salesforce Ventures. The San Francisco startup, founded by DeepMind alumni, is launching NEXUS: a Large Tabular Model designed specifically for the structured data that dominates enterprise operations.
The technical gap is real. While LLMs excel at text and images, they struggle with tabular data for architectural reasons. Transformer models tokenize numbers like words, breaking "2.3" into three separate tokens and losing numerical distribution understanding. Context window limitations prevent reliable reasoning across the billion-row datasets common in Fortune 500 operations. The result: enterprises still rely on pre-deep learning algorithms and armies of data scientists for forecasting.
NEXUS takes a different approach. The model is deterministic rather than probabilistic, meaning identical queries produce identical results without hallucinations. It ingests raw tabular data directly, learns underlying patterns automatically, and requires minimal feature engineering. Training used billions of datasets via Amazon SageMaker HyperPod.
"LLMs really cannot handle this type of data very well," CEO Jeremy Fraenkel told VentureBeat. The distinction matters: while Anthropic's Claude integration with Excel operates at the formula layer, NEXUS works at the predictive layer. Think split-second fraud detection or equipment failure forecasting, not spreadsheet summarization.
The business case is straightforward. Enterprises process terabytes of structured data daily: transactions, supply chain metrics, operational logs, financial records. Current approaches require months of manual model building. Fundamental claims radical time-to-insight reduction.
The company has announced an AWS partnership with Fortune 100 contracts already signed. Worth noting: only 20% of organizations have achieved the data readiness required for AI success, according to recent industry data. Even purpose-built models face adoption barriers around data quality, governance, and fragmentation.
History suggests the real test comes during implementation. The architecture addresses a genuine technical problem that LLMs can't solve. Whether enterprises can execute on deployment is the next question.