Fundamental NEXUS

NEXUS is a Large Tabular Model from Fundamental, pre-trained on billions of real-world prediction tasks over structured data, that makes classification and regression predictions directly from raw enterprise tables. It is aimed at data science and analytics teams who currently spend months on ETL and feature engineering before a demand forecast, churn model or fraud score reaches production.

NEXUS is a Large Tabular Model built by Fundamental, a company founded in October 2024 by DeepMind alumni Jeremy Fraenkel and Gabriel Suissa, with US headquarters and research operations in Israel. Where large language models are optimised for sequential text, NEXUS is a foundation model pre-trained on billions of real-world prediction tasks over structured tables, and its architecture reflects that: it processes numbers natively rather than tokenising a value like 2.3 into separate characters, and it is permutation-invariant, so column ordering does not change a prediction. It is also deterministic — the same question returns the same answer, which matters for regulated decisioning where a probabilistic answer is unusable. The commercial pitch is the removal of work rather than a new interface: NEXUS ingests raw tables and finds latent patterns without the manual feature engineering that gradient-boosted models such as XGBoost require, and Fundamental claims deployment in a single line of code against a scikit-learn-compatible API exposing NEXUSClassifier and NEXUSRegressor. It handles billion-row datasets, cross-schema reasoning across related tables and incomplete entries. Fundamental emerged from stealth in February 2026 with $255 million raised across a $30 million seed and a $225 million Series A led by Oak HC/FT, at a valuation reported between $1.2 and $1.4 billion, with Salesforce Ventures, Battery Ventures, Valor Equity Partners and Hetz Ventures participating. Distribution runs through partners: NEXUS is available on Amazon SageMaker JumpStart via AWS Marketplace on dedicated single-tenant GPU instances, and since May 2026 through SAP's generative AI Hub.

Category

AI Models & APIs

Pricing

Contact for pricing, Usage-based

Target Market

Chief Data Officers, Data Scientists, CIOs, Heads of Analytics, ML Engineers

Deployment

API-based, Cloud-first, Hybrid

Founded

2024

Customers

Not publicly enumerated; multiple Fortune 100 enterprises on seven-figure contracts, with Japan's largest bank referenced by the vendor

• ✓
  Native numerical processing

  Numbers are handled as values rather than tokenised character by character, preserving the distributional understanding LLMs lose on tabular data.

• ✓
  Permutation invariance

  Column order does not affect predictions, so a table is treated as a set of relationships rather than a sequence to be read left to right.

• ✓
  No task-specific training

  Pre-training on billions of prediction tasks means the model produces predictions without a per-dataset training run or hyperparameter search.

• ✓
  Billion-row and cross-schema reasoning

  Processes datasets that exceed any language model context window and connects related fields across multiple disparate tables.

• ✓
  Deterministic outputs

  The same query returns the same answer every time, which is required for regulated decisioning such as credit and fraud scoring.

• ✓
  scikit-learn compatible API

  NEXUSClassifier and NEXUSRegressor drop into existing Python workflows, so data science teams do not rewrite their pipelines to adopt it.

• ✓
  Deploy-to-data architecture

  Runs as a single-tenant, network-isolated endpoint inside the customer's own AWS environment, so enterprise data never leaves the account.

• •
  Demand forecasting at retail scale

  Point NEXUS at historical sales and inventory tables to produce forecasts without first building and maintaining a feature engineering pipeline.

• •
  Churn prediction across fragmented schemas

  Cross-schema reasoning links customer, billing and usage tables so churn signals spread across systems are captured without manual joins.

• •
  Fraud detection in financial services

  Deterministic classification over transaction tables supports audit and model-risk requirements that probabilistic LLM output cannot satisfy.

• •
  Predictive maintenance in industrial operations

  Sensor readings and maintenance history are used to predict failures, a workload SAP surfaced as a target for the generative AI Hub integration.

• •
  Expanding the predictive backlog

  By removing months of feature engineering per model, prediction problems previously too small to justify a data science project become economically viable.

Fundamental publishes no list pricing anywhere — the website routes to a contact form and the AWS listing does not disclose rates. Two things are knowable about real cost. First, reported contracts are seven figures with Fortune 100 buyers, which sets the practical entry point well above mid-market. Second, and more concretely, the AWS SageMaker JumpStart deployment requires an ml.p5en.48xlarge instance carrying eight NVIDIA H200 GPUs as a dedicated single-tenant endpoint, so the infrastructure cost alone is substantial and is billed separately through AWS on top of whatever Fundamental charges for the model subscription. Buying through AWS Marketplace does let enterprises draw down existing AWS commitments rather than open a new vendor contract, which is a genuine procurement advantage. The SAP generative AI Hub route, announced May 2026, is a third path with no published commercial terms.

This page was written from 6 sources, 4 on domains other than fundamental.tech.

1. 1.fundamental.tech — fundamental.techvendor
2. 2.fundamental.tech — sap nexus tabular aivendor
3. 3.techcrunch.com — fundamental raises 255 million series a with a new take on b
4. 4.venturebeat.com — fundamental emerges from stealth with first major foundation
5. 5.aws.amazon.com — fundamentals large tabular model nexus is now available on a
6. 6.calcalistech.com — hktlfm7vbx