The archive · AI & Models · Product decision · 2024–2026
Fundamental bets LLMs can't read spreadsheets: $255M at $1.4B for a tabular model
DeepMind alumni built NEXUS, a deterministic non-transformer model for enterprise tables; it left stealth in Feb 2026 with $255M at $1.4B
Fundamental
What the business is
Fundamental is a San Francisco AI lab whose model NEXUS — a 'large tabular model' (LTM) — ingests raw enterprise tables and predicts outcomes like demand, churn, fraud and equipment failure, replacing hand-built data-science pipelines with one line of code.
Starting capital：$255M total: $30M seed plus $225M Series A led by Oak HC/FT with Valor Equity Partners, Battery Ventures, Salesforce Ventures and Hetz Ventures; angels include Perplexity CEO Aravind Srinivas, Wiz CEO Assaf Rappaport, Brex co-founder Henrique Dubugras and Datadog CEO Olivier Pomel
How it started
Founded in October 2024 by DeepMind alumni in San Francisco, Fundamental stayed in stealth while building NEXUS, a large tabular model trained on billions of real-world tabular datasets using Amazon SageMaker HyperPod. It emerged publicly on February 5, 2026 with $255M in funding.
What happened
The $30M seed and $225M Series A — led by Oak HC/FT with Valor Equity Partners, Battery Ventures, Salesforce Ventures and Hetz Ventures — priced the company at $1.4B post-money. NEXUS shipped with several seven-figure Fortune 100 contracts in demand forecasting, price prediction and churn, plus a deep AWS partnership where customers buy and deploy the model from their AWS dashboard using existing credits; AWS also supports encrypted deployments of the model inside customer environments.
How it ended up
Still running. As of mid-2026 NEXUS is generally available through AWS Marketplace and fundamental.tech, with a roughly 35-person team scaling compute, enterprise deployments and go-to-market.
Background
Fundamental is a San Francisco AI lab, founded in October 2024 by DeepMind alumni, that emerged from stealth on February 5, 2026 with $255M in funding at a $1.4B post-money valuation. Its product, NEXUS, is what the company calls a large tabular model (LTM): a foundation model purpose-built for the structured rows and columns that run enterprises, rather than for text.
The architectural bet is deliberately contrarian. CEO and co-founder Jeremy Fraenkel argues that LLMs tokenize numbers like words — '2.3' becomes three tokens — and cannot reason over billion-row tables, while enterprises still rely on manual feature engineering and XGBoost-era algorithms. NEXUS is deterministic, does not use the transformer architecture, and is trained on billions of tabular datasets; developers connect raw tables and label a target column, and the model returns predictions.
The $225M Series A was led by Oak HC/FT with Valor Equity Partners, Battery Ventures, Salesforce Ventures and Hetz Ventures, with angels including Perplexity CEO Aravind Srinivas, Wiz CEO Assaf Rappaport, Brex co-founder Henrique Dubugras and Datadog CEO Olivier Pomel. The company launched with several seven-figure Fortune 100 contracts in demand forecasting, price prediction and customer churn.
The go-to-market is built around a deep AWS partnership: AWS customers buy and deploy NEXUS from their AWS dashboard like compute or storage, using existing credits, and AWS supports fully encrypted deployments inside the customer's own environment — a structure Fraenkel calls one of the most integrated AI partnerships with Amazon. As of mid-2026 the ~35-person team is scaling compute, deployments and go-to-market.
What has to be true
- The case is a clean contrarian-architecture bet: a deterministic, non-transformer model aimed at the data modality LLMs ignore, backed by $255M at a $1.4B valuation before anyone had seen it.
- Traction is concrete and verifiable: seven-figure Fortune 100 contracts at launch, an AWS dashboard distribution deal, and same-day coverage by TechCrunch and VentureBeat.
- It names the weakness precisely — number tokenization and context windows — which makes the bet falsifiable and the lesson transferable to any 'wrong architecture' incumbent.
- The funding roster (Oak HC/FT, Valor, Battery, Salesforce Ventures plus CEO angels from Perplexity, Wiz, Brex and Datadog) shows how strongly the thesis resonated with enterprise-AI investors.
What can be applied
When hot models share an unexamined assumption — 'one architecture fits all data' — the wedge is the modality everyone ignores; a contrarian architecture bet can outraise another LLM clone.
Aftermath
As of 2026-09-02 Fundamental is post-launch and scaling: NEXUS is available to enterprises via AWS Marketplace and fundamental.tech, the team of roughly 35 is growing across research, engineering and go-to-market, and the company is expanding enterprise deployments beyond its initial Fortune 100 demand-forecasting, price-prediction and churn contracts. The core open question remains whether a single tabular foundation model can hold up across industries the way the company claims.
Sources
- Fundamental raises $255M Series A with a new take on big data analysis
- Beyond the lakehouse: Fundamental's NEXUS bypasses manual ETL with a native foundation model for tabular data
- Fundamental Announces $255M in Funding and Publicly Launches its Most Powerful Large Tabular Model (LTM)
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card