Fundamental (company)

Note: This article is about Fundamental, the developer of the NEXUS large tabular model. For the AI agent company that makes the Shortcut spreadsheet agent, see Fundamental Research Labs.

Fundamental (legal name Fundamental Technologies Inc.) is an American artificial intelligence company, headquartered in San Francisco, that builds what it calls large tabular models (LTMs): foundation models meant to make predictions from structured data kept in tables, spreadsheets and databases rather than from text.^[2][5] Its first public model, NEXUS, is sold to large companies for prediction tasks such as demand forecasting, price prediction and customer churn, and it can be bought and deployed through Amazon Web Services.^[1][2] The company was founded in October 2024.^[2][7] Its co-founders are chief executive Jeremy Fraenkel, chief product officer Gabriel Suissa and chief science officer Marta Garnelo, a former DeepMind researcher.^{[4][8][16][20][30]}

Fundamental left stealth on February 5, 2026 with $255 million in total funding, made up of a $30 million seed round and a $225 million Series A round led by Oak HC/FT.^[2] TechCrunch reported that the money was raised at a $1.4 billion post-money valuation, while the Israeli business outlet Calcalist reported a $1.2 billion valuation.^[1][6] During 2026 the company added distribution and commercial agreements with Amazon SageMaker, SAP, the Japanese banking group MUFG (whose venture arm also invested) and the sports and entertainment company AEG.^{[13][15][16][17]}

Overview

History

Founding and stealth period

Fundamental was founded in October 2024.^[2] Salesforce Ventures, which says it co-led the seed round, wrote at the launch that the company had been "building in stealth for 18 months".^[4] According to Salesforce Ventures, the team came from organizations including DeepMind, Isomorphic Labs and JP Morgan.^[4] A later partnership release described the founders as alumni of JPMorgan, Bridgewater and DeepMind.^[17] The venture firm Headline also lists Fundamental as a portfolio company, recording it as an early-stage company in 2024.^[30]

Launch

The company announced itself on February 5, 2026. At the same time it released NEXUS and said it had signed seven-figure contracts with Fortune 100 companies that were using the model for demand forecasting, price prediction and customer churn.^[2] It also announced a strategic partnership with Amazon Web Services under which AWS customers could buy and deploy NEXUS in their own AWS environment "in the same way they buy compute or storage".^[2] VentureBeat reported that AWS acts as the seller of record on the AWS Marketplace, so customers can pay with existing AWS credits.^[3] Fraenkel told VentureBeat that Fundamental was one of only two AI companies with a partnership of this depth with Amazon.^[3] At launch the company had a team of about 35, based in San Francisco.^[3]

Some launch-day coverage named Annie Lamont as a Fundamental co-founder in photo captions.^[3][5] Lamont is co-founder and managing partner of Oak HC/FT, the Series A lead, and is quoted in that capacity in the launch announcement.^[2] Later coverage identified the founders as Fraenkel, Garnelo and Suissa.^[8]

Expansion in 2026

In May 2026 the Times of Israel reported that Fundamental would open an engineering center in Tel Aviv, expected to launch on June 2, and planned to hire dozens of engineering and infrastructure staff in Israel by the end of 2026. At that point the company had offices in San Francisco and Barcelona, a small presence in Tel Aviv, and 70 employees.^[7] Calcalist had earlier reported that Fundamental's Israeli research and development work employed engineers who had previously worked at AI21 Labs.^[6] By October 2026 the company's careers page listed four hubs: San Francisco (headquarters, engineering and go-to-market), Barcelona (research), Tel Aviv (engineering) and Singapore.^[11]

In October 2026 Fundamental was named to the Disruptors 60 class of 2026, a list of AI startups compiled by the venture firm Greenfield Partners and unveiled at TechCrunch Disrupt.^[31]

Timeline

Founders and leadership

Jeremy Fraenkel leads the company as CEO. His conference biography says he started his career at JP Morgan and Bridgewater, founded two startups before Fundamental and exited both, and received a graduate degree in machine learning from UC Berkeley.^[28] The biography does not name the two startups. Fraenkel co-founded Arkifi, a startup building generative AI tools for financial analysts that was incubated at Khosla Ventures, spun out in March 2023 and raised a $9 million seed round co-led by Khosla Ventures and Nyca Partners. Its 2023 announcement named Fraenkel and Hesam Motlagh as co-founders and described Fraenkel's path from JP Morgan's mid-cap M&A group through graduate AI studies at Berkeley to Bridgewater Associates.^[29]

Gabriel Suissa is the company's chief product officer.^[16] Calcalist and the Times of Israel described him as an Israeli entrepreneur, and the Times of Israel quoted him on the decision to open an engineering hub in Tel Aviv: "To solve a technology problem of this magnitude, we need to recruit the best talent in the world," he said.^[6][7]

Marta Garnelo is chief science officer.^[20] She was a PhD student in the Computational Neurodynamics group at Imperial College London, supervised by Murray Shanahan.^[27] She is the first author of two 2018 papers, "Conditional Neural Processes" and "Neural Processes", which introduced neural network models that, like Gaussian processes, represent distributions over functions and adapt quickly to a handful of observed data points, while being trained by gradient descent like ordinary neural networks.^[25][26] According to Fundamental, she and the researcher Wojciech Czarnecki, known for his work on DeepMind's AlphaStar StarCraft project, worked together at DeepMind for close to a decade.^[20] Fundamental's research white paper, which she co-wrote, builds on the Neural Processes framing.^[22]

Other named staff include Alex Gerbeaux, head of applied AI, and Richard McBeath, head of marketing.^[17][21] Gaël Varoquaux, a co-founder of scikit-learn and chief science officer of Probabl, is described by Fundamental as a founding advisor.^[11]

NEXUS

Purpose

NEXUS is Fundamental's large tabular model. The company presents it as a single pre-trained model that can replace the separate, hand-built machine learning model that companies usually train for each prediction problem.^[2][10] In traditional practice those problems are handled with methods such as gradient boosting libraries like XGBoost, which data scientists must train and tune for each use case, often with manual feature engineering.^[3][8] Fundamental says NEXUS takes raw tables and learns their structure without that manual work, often with "a single line of code".^[2] In his launch post Fraenkel called the model "the OS for enterprise decisions".^[9]

Fraenkel has argued that large language models are a poor fit for this kind of data. He told VentureBeat that LLMs split numbers into separate tokens (so "2.3" becomes three tokens) and lose a sense of their distribution, and that a model for tables should give the same prediction whatever the order of the columns.^[3] He also distinguished NEXUS from spreadsheet assistants built on LLMs: such tools work at "the formula layer", he said, while Fundamental is "helping you make a forecast".^[3] In an interview with IEEE Spectrum he compared LLMs and LTMs to the two sides of the human brain, one good at language and the other at numbers and statistics, and said combining them is more powerful.^[8]

Architecture

Fundamental has described the model's design only in general terms. TechCrunch reported that NEXUS is deterministic, meaning it returns the same answer each time it is given the same question, and that it does not rely on the transformer architecture used by most AI labs.^[1] AWS's launch post for SageMaker likewise lists a "deterministic architecture" and "permutation invariance", which it explains as recognizing that changing column order does not change meaning, "which differs from how transformers handle data".^[15] Fundamental's own product page says the model "leverages advanced neural network and attention-based architectures to pre-train on billions of rows, then adapts to your data", including extracting meaning from text features.^[10] TechCrunch noted that the company calls NEXUS a foundation model because it goes through pre-training and fine-tuning.^[1]

Training

The company and its partners describe the training data in slightly different terms: "billions of tabular datasets" in the launch announcement, "billions of rows and columns" on the product page, and "billions of real-world prediction tasks" in AWS's SageMaker post.^[2][10][15] Fraenkel told IEEE Spectrum that obtaining training data was one of the company's biggest challenges, because tabular data is hard to find, often sensitive or proprietary, and very diverse. "We pre-trained NEXUS on billions of tables using a combination of proprietary datasets acquired through partnerships and licensing, high-quality public and open-source datasets, and data augmentation techniques," he said, adding that the model is not trained on customer data.^[8] The model was trained on Amazon SageMaker HyperPod.^[2]

Interface and deployment

NEXUS is used from Python through an SDK that follows the scikit-learn estimator pattern: a user fits the model to a table and a target column, calls predict or predict_proba, and can request explanations of individual predictions and feature importance.^[10][15] AWS's documentation names the estimators NEXUSClassifier and NEXUSRegressor.^[15]

The company offers managed SaaS, deployment inside a customer's own cloud, and deployment on private infrastructure.^[10] In April 2026 a Fundamental founding engineer described the company's confidential computing setup. The model is shipped encrypted into the customer's virtual private cloud, and decryption keys are released only after the server's trusted hardware confirms that the software stack matches an approved cryptographic fingerprint; the aim is that the customer cannot extract the model and Fundamental cannot see the customer's data.^[19] At launch Fraenkel told VentureBeat that Fundamental, working with Amazon, was "the first, and currently only, company" able to deploy fully encrypted models inside customers' own environments; that is the company's own claim.^[3]

On June 3, 2026 AWS announced NEXUS on Amazon SageMaker JumpStart. Customers subscribe to the model package on AWS Marketplace and deploy it as a single-tenant, network-isolated endpoint on an ml.p5en.48xlarge instance with eight NVIDIA H200 GPUs, with industry-specific variants for finance, healthcare and manufacturing.^[15] Fundamental said data, inputs, outputs and model artifacts stay in the customer's own AWS account and that one endpoint can serve several trained models for different use cases.^[14]

Uses

Fundamental and its partners list uses including fraud detection, credit risk, demand forecasting, price prediction, churn prediction, predictive maintenance and hospital readmission risk.^[2][3][15] The Times of Israel reported that its seven-figure Fortune 100 contracts were in oil and gas, finance and healthcare, and the October 2026 AEG release said NEXUS was already running in production in the energy, shipping and manufacturing sectors.^[7][17] The company's website publishes case studies in areas such as energy demand, grid outages, freight rates, credit risk, Formula 1 aerodynamics and NFL player valuation; the accuracy figures in them are the company's own.^[10] It has also published public sports forecasts, such as predictions for the 2026 FIFA World Cup, as demonstrations of the model.^[32]

Research

Fundamental describes itself as "a frontier AI research lab building foundation models for the world's structured data".^[12] Its research page lists six areas: meta-learning, architecture research, generative modeling, large-scale systems, code generation and structured data.^[12] The company describes Barcelona as its research hub.^[11][24]

The white paper opens with Richard Sutton's essay "The Bitter Lesson" and argues that many supervised learning successes, from AlphaFold to weather forecasting and recommendation systems, can be framed as tabular prediction problems. It describes the goal as a "universal predictor" and argues that general language models are misaligned with that goal because of their training objective and because many real tables would need far more context tokens than current LLMs can hold.^[22]

The ICML presentation and the arXiv paper examine a frontier LLM, identified in the presentation as Claude Opus 4.6, queried directly with the training and test rows in its prompt and no tools.^[23][24] The presentation reports that the model appeared to have memorized the labels of familiar benchmark datasets such as Iris, that in two dimensions its predictions closely matched those of a 1-nearest-neighbor classifier, and that it was the only one of nine methods whose score fell as columns were added.^[24] The arXiv paper reports that controlled experiments falsified four candidate explanations (noisy or non-linearly-separable data, the CSV format, the tokenization of numbers, and the number of test points per query), while dimensionality was decisive across random projections of thirty-one benchmark datasets. The authors state that they "do not claim to have identified the internal mechanism".^[23] The finding on tokenization differs from the launch-time account in VentureBeat, which presented the way LLMs split numbers into tokens, as Fraenkel explained it, as a primary reason they fail at tabular data.^[3][23]

The company also runs two video interview series. First Principles, hosted by Garnelo, began in March 2026 with an episode featuring Czarnecki.^[20] Ground Truth, hosted by Gerbeaux, began in June 2026 with Varoquaux as its first guest.^[21]

Funding

The launch announcement put total funding at $255 million and said the money would be used for compute, enterprise deployments and hiring across research, engineering and go-to-market.^[2] Angel investors named by the company include Aravind Srinivas of Perplexity, Assaf Rappaport of Wiz, Henrique Dubugras of Brex and Olivier Pomel of Datadog.^[2][11]

Reports differ on several details:

• Series A leads. Fundamental's announcement names Oak HC/FT as the lead and the other firms as participants, and VentureBeat and SiliconANGLE reported it the same way.^[2][3][5] TechCrunch described the round as led by Oak HC/FT, Valor Equity Partners, Battery Ventures and Salesforce Ventures.^[1] Salesforce Ventures itself says it co-led the seed round and participated in the Series A.^[4]
• Valuation. The company's announcement did not state a valuation.^[2] TechCrunch reported a $1.4 billion post-money valuation. Calcalist reported that the company said it had raised the money at a $1.2 billion valuation, without specifying pre- or post-money.^[1][6]
• Total raised. In May 2026 the Times of Israel described the February funding as $275 million at a $1.45 billion valuation, and IEEE Spectrum also gave $275 million.^[7][8] These figures do not match the February announcement. In October 2026 Fundamental's profile on the Disruptors 60 site said it had "$275M raised to date".^[31] The amount of MUFG Innovation Partners' investment was not disclosed.^[16]

Partnerships and customers

MUFG Innovation Partners is the corporate venture capital arm of Mitsubishi UFJ Financial Group.^[33] The AEG agreement, brokered by AEG Global Partnerships, was described by the company as its first formal partnership in sports and live entertainment. The announcement also said NEXUS would be the presenting sponsor of the Coachella Valley Invitational, a pre-season professional soccer tournament.^[17][18]

Large tabular models

Fundamental is one of several groups building foundation models for tabular data. Earlier academic work in the area includes TabPFN, a transformer introduced in 2022 that is trained offline on synthetic datasets and classifies small tabular datasets by in-context learning, without further parameter updates.^[34] (See tabular models for the wider field.) A February 2026 post by Citi Ventures listed Fundamental Technologies among LTM developers along with Prior Labs, Kumo, Neuralk-AI and Wood Wide AI.^[35] IEEE Spectrum reported in July 2026 that Feedzai and Mastercard had separately launched similar proprietary technologies focused on finance in March of that year.^[8] On June 30, 2026 Google Research introduced TabFM, a zero-shot tabular foundation model that Google said was trained entirely on hundreds of millions of synthetic datasets.^[8][36] Fundamental, by contrast, says it pre-trained NEXUS on real-world tables from licensed, proprietary and public sources, together with data augmentation.^[8][14]

IEEE Spectrum also quoted Boris van Breugel, an Amsterdam-based AI researcher who co-wrote a 2024 position paper on the topic, on a difficulty for the whole category: because tables can have very different variables, he argued, it is much harder to train a single tabular model across them than to train a language model on text with broadly similar semantics.^[8]

Reception

Annie Lamont of Oak HC/FT, the Series A lead, said at the launch that "structured, relational data has yet to see the benefits of the deep learning revolution".^[2] Debarghya Das's "The Ultimate Neolab List" includes Fundamental and gives it the call "Mostly" in its "Neolab?" column (see neolab), on a scale that runs from "Clean" through "Mostly" and "Borderline" to "Not a neolab". The list's note on the company reads: "Research-led foundation model and heavy early raise, but already has enterprise contracts and commercial go-to-market."^[37]

See also

• List of neolabs
• Neolab
• Tabular models
• XGBoost
• Foundation models
• Fundamental Research Labs

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Russell Brandom, TechCrunch, Fundamental raises $255M Series A with a new take on big data analysis, February 5, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27Fundamental (via Amazon US Press Center), Fundamental Announces $255M in Funding and Publicly Launches its Most Powerful Large Tabular Model (LTM), February 5, 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Carl Franzen, VentureBeat, Beyond the lakehouse: Fundamental's NEXUS bypasses manual ETL with a native foundation model for tabular data, February 5, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6Salesforce Ventures, Bringing AI to Structured Data, February 5, 2026.
5. ^1 ^2 ^3 ^4Maria Deutscher, SiliconANGLE, Fundamental launches with $255M and an AI model optimized for tabular data, February 5, 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7CTech (Calcalist), Fundamental exits stealth as $1.2 billion unicorn, armed with $255 million for tabular AI, February 6, 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Sharon Wrobel, The Times of Israel, Big data AI startup will open Tel Aviv engineering hub to tap into local tech talent, May 19, 2026.
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Benjamin Skuse, IEEE Spectrum, Large Tabular Models Bring AI Analysis to Spreadsheets, July 9, 2026.
9. ^Jeremy Fraenkel, Fundamental, Revealing The Hidden Language of Tables, February 5, 2026.
10. ^1 ^2 ^3 ^4 ^5 ^6Fundamental, NEXUS, accessed October 27, 2026.
11. ^1 ^2 ^3 ^4 ^5 ^6Fundamental, Careers, accessed October 27, 2026.
12. ^1 ^2 ^3 ^4Fundamental, Research, accessed October 27, 2026.
13. ^1 ^2 ^3Jeremy Fraenkel, Fundamental, NEXUS + SAP: The world's enterprise data just got a lot smarter, May 12, 2026.
14. ^1 ^2Jeremy Fraenkel, Fundamental, Bringing the model to the data: NEXUS launches on Amazon SageMaker, June 8, 2026.
15. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Vivek Gangasani, Hazim Qudah and Jimmy Shah, AWS Machine Learning Blog, Fundamental's Large Tabular Model NEXUS is now available on Amazon SageMaker JumpStart, June 3, 2026.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Gabriel Suissa, Fundamental, Why Japan's largest bank is betting on Fundamental, July 17, 2026.
17. ^1 ^2 ^3 ^4 ^5 ^6 ^7LA Galaxy Communications, Fundamental Brings Nexus to LA Galaxy and L.A. Live Through New Official Partnership with AEG, October 8, 2026.
18. ^1 ^2Jeremy Fraenkel, Fundamental, Fundamental + AEG: Bringing prediction to sports and live entertainment, October 9, 2026.
19. ^1 ^2Yuval Azoulay, Fundamental, Bringing the Model to the Data: How Fundamental Built a Fully Secured Confidential Computing Platform for Enterprise AI, April 20, 2026.
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7Jeremy Fraenkel, Fundamental, Introducing First Principles, March 26, 2026.
21. ^1 ^2 ^3Alex Gerbeaux, Fundamental, "Things are happening": Introducing Ground Truth, June 18, 2026.
22. ^1 ^2 ^3Marta Garnelo and Wojciech Marian Czarnecki, Fundamental, Developing Foundation Models for Real-World Tabular Prediction (white paper, PDF), accessed October 27, 2026.
23. ^1 ^2 ^3 ^4 ^5Marta Garnelo and Wojciech M. Czarnecki, Why Large Language Models Fail at Tabular Prediction, arXiv:2608.02412, August 2026.
24. ^1 ^2 ^3 ^4Marta Garnelo and Wojciech Marian Czarnecki (presentation by Kevin Scaman), Fundamental, What LLMs learn (and don't) from tables (PDF), July 2026.
25. ^1 ^2Marta Garnelo, Dan Rosenbaum, Chris J. Maddison, Tiago Ramalho, David Saxton, Murray Shanahan, Yee Whye Teh, Danilo J. Rezende and S. M. Ali Eslami, Conditional Neural Processes, arXiv:1807.01613, July 2018.
26. ^1 ^2Marta Garnelo, Jonathan Schwarz, Dan Rosenbaum, Fabio Viola, Danilo J. Rezende, S. M. Ali Eslami and Yee Whye Teh, Neural Processes, arXiv:1807.01622, July 2018.
27. ^1 ^2Marta Garnelo, personal page, Department of Computing, Imperial College London, archived April 11, 2019.
28. ^1 ^2HumanX, Jeremy Fraenkel (speaker profile), accessed October 27, 2026.
29. ^1 ^2Global Fintech Series (Cision PRWeb), Arkifi Emerges From Stealth, Building Factual and Trustworthy Tools With Generative AI for the Finance Industry, August 25, 2023.
30. ^1 ^2 ^3Headline, Fundamental (portfolio page), accessed October 27, 2026.
31. ^1 ^2 ^3 ^4Greenfield Partners, Disruptors 60, Fundamental, Disruptors 60 Class of 2026, accessed October 27, 2026.
32. ^Fundamental, At the last World Cup, our model picked Argentina before the market. We're making that call again., June 19, 2026.
33. ^MUFG Innovation Partners, Bridging Silicon Valley and Japan: Stories of MUFG's open innovation efforts, accessed October 27, 2026.
34. ^Noah Hollmann, Samuel Muller, Katharina Eggensperger and Frank Hutter, TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second, arXiv:2207.01848, July 2022.
35. ^Citi Ventures, Large tabular models: A new approach to AI with significant promise, February 10, 2026.
36. ^Weihao Kong and Abhimanyu Das, Google Research, Introducing TabFM: A zero-shot foundation model for tabular data, June 30, 2026.
37. ^Debarghya Das, The Ultimate Neolab List, accessed October 27, 2026.