The deep learning revolution has a strange blind spot: spreadsheets. While large language models (LLMs) have mastered the nuances of human prose and image generators have conquered the digital canvas, the structured, relational data that underpins the global economy – the rows and columns of ERP systems, CRMs and financial ledgers – has until now been treated as just another file format, akin to text or PDF.
This has left enterprises to forecast business outcomes using a custom, labor-intensive data science process typical of manual feature engineering and classic machine learning algorithms that predates modern deep learning.
But now Fundamental, a San Francisco-based AI firm co-founded by DeepMind alumni, is launching today with a total of $255 million in funding to bridge this gap.
Emerging in secret, the company is introducing NEXUS, a large tabular model (LTM) designed to treat business data not as a simple sequence of words, but as a complex web of non-linear relationships.
Technique: Proceeding from Sequential Reasoning
Most current AI models are built on sequential reasoning – predicting the next word in a sentence or the next pixel in a frame.
However, enterprise data is inherently non-sequential. A customer’s churn risk isn’t just a timeline; It is a multidimensional intersection of transaction frequency, support ticket sentiment, and regional economic shifts. Existing LLMs struggle with this because they are poorly suited for the size and dimensional constraints of enterprise-scale tables.
"The most valuable data in the world lives in tables and until now there has been no good base model specifically designed to understand it," said Jeremy Frankel, CEO and co-founder of Fundamental.
In a recent interview with VentureBeat, Frankel emphasized that while the AI world is obsessed with text, audio, and video, tables remain the biggest tool for enterprises. "LLMs can’t really handle this type of data very well," he explained, "And enterprises currently rely on very old-school machine learning algorithms to make predictions."
NEXUS was trained on billions of real-world tabular datasets using Amazon SageMaker Hyperpod. Unlike traditional
It identifies latent patterns in columns and rows that human analysts might miss, effectively reading the hidden language of the grid to understand non-linear interactions.
tokenization mesh
One of the primary reasons traditional LLMs fail with tabular data is because of how they process numbers. Frankel explains that LLMs tokenize numbers in the same way they tokenize words, breaking them into smaller pieces. "The problem is that they apply the same thing to numbers. The tables are, by and large, all numeric," Frankel noted. "If you have a number like 2.3, ‘2’, ‘.’, and ‘3’ are seen as three separate tokens. This essentially means that you lose understanding of the distribution of numbers. It’s not like a calculator; You don’t always get the right answer because the model doesn’t natively understand the concept of numbers."
Furthermore, tabular data is order-invariant in a way that the language is not. Frankel uses a health care example to illustrate this: "If I give you a table of hundreds of thousands of patients and ask you to guess which of them has diabetes, it doesn’t matter whether the first column is height and the second is weight, or vice versa."
While LLMs are highly sensitive to the order of words in a signal, Nexus is designed to understand that changing the position of a column should not affect the underlying prediction.
Operating on predictive layer
Recent high-profile integrations, such as Anthropic’s cloud displays directly into Microsoft Excel, have suggested that LLMs are already solving tables.
However, Frankel distinguishes the work of Fundamentals as working on a fundamentally different layer: the predictive layer. "What they are doing is basically at the sutra level—sutras are texts, they are like codes," He said. "We are not trying to allow you to create financial models in Excel. We’re helping you forecast."
NEXUS is designed for split-second decisions where no human is in the loop, like a credit card provider determining whether a transaction is fraudulent as you swipe.
Whereas tools like the cloud can summarize a spreadsheet, Nexus is built to predict the next line of action – whether it’s equipment failure in a factory or the likelihood of a patient being readmitted to a hospital.
Architecture and availability
Fundamental’s core value proposition is a radical reduction in time-to-insight. Traditionally, building a predictive model could take several months of manual labor.
"You have to hire an army of data scientists to build all those data pipelines to process and clean the data," Frankel explained. "If there are missing values or inconsistent data, your model will not work. You have to build those pipelines for each use case."
The fundamental claim is that NEXUS replaces this entire manual process with just one line of code. Because the model is pre-trained on one billion tables, it does not require the same level of task-specific training or feature engineering as traditional algorithms.
As Fundamental moves from its latent phase into the mass market, it does so with a commercial structure designed to bypass the traditional friction of enterprise software adoption.
The company has already secured multiple seven-figure contracts with Fortune 100 organizations, a feat facilitated by a strategic go-to-market architecture where Amazon Web Services (AWS) serves as the vendor of record on the AWS Marketplace.
This allows enterprise leaders to purchase and deploy Nexus using existing AWS credits, effectively treating compute and storage as well as predictive intelligence as a standard utility. For the engineers doing the implementation work, the experience is high-impact but low-friction; NEXUS operates through a Python-based interface at a purely predictive layer, rather than a conversational interface.
Developers connect raw tables directly to the model and label specific target columns – such as credit default probability or maintenance risk score – to trigger forecasts. The model then returns the regression or classification directly to the enterprise data stack, acting as a silent, high-speed engine for automated decision making rather than a chat-based assistant.
Social Stakes: Beyond the Bottom Line
While the business implications of demand forecasting and price forecasting are clear, the fundamentals are emphasizing the societal benefits of predictive intelligence.
The company highlights key areas where NEXUS can prevent devastating outcomes by identifying signals hidden in structured data.
By analyzing sensor data and maintenance records, NEXUS can predict failures such as pipe corrosion. The company points to the Flint water crisis – which has cost more than $1 billion to repair – as an example where predictive monitoring could have prevented life-threatening pollution.
Similarly, during the COVID-19 crisis, PPE shortages cost hospitals $323 billion in a single year. Fundamental argues that by using manufacturing and epidemiological data, Nexus can predict shortages 4-6 weeks before peak demand, allowing emergency manufacturing to begin in time to save lives.
On the climate front, Nexus aims to provide 30-60 day forecast of floods and droughts, such as 2022. Pakistan floods caused losses of $30 billion.
Finally, the model is being used to predict hospital readmission risks by analyzing patient demographics and social determinants. As the company says: "A single mom working two jobs should not end up back in the ER because we failed to anticipate that she would need follow-up care."
performance vs latency
In the enterprise world, the definition of better varies by industry. For some, it’s the speed; For others, it’s raw accuracy.
"In terms of latency, it depends on the use case," Frankel explains. "If you’re a researcher trying to understand which drugs to give to a patient in Africa, latency doesn’t matter that much. You are trying to make more accurate decisions that can save the most lives possible."
In contrast, for a bank or hedge fund, even a small increase in accuracy translates into massive value.
"Increasing forecast accuracy by half a percent is worth billions of dollars to a bank," Frankel says. "For different use cases, the magnitude of the percentage increase varies, but we can provide you with better performance than you currently have."
Ambitious vision gets big support
The $225 million Series A, led by Oak HC/FT with participation from Salesforce Ventures, Valor Equity Partners, and Battery Ventures, signals high-conviction belief that tabular data is the next great frontier.
Notable angel investors, including leaders from Perplexity, Viz, Brex, and Datadog, further validate the company’s pedigree.
Anne Lamont, co-founder and managing partner of Oak HC/FT, echoed this sentiment: "It is difficult to overstate the importance of fundamental models – the benefits of the deep learning revolution have yet to be seen in structured, relational data."
Fundamentals is establishing itself as not just another AI tool, but as a new category of enterprise AI. With a team of about 35 people based in San Francisco, the company is moving away from the custom model era toward the foundation model era for tables.
"Those traditional algorithms have been the same for the last 10 years; They are not improving," Frankel said. "Our models keep improving. We’re doing for tables what ChatGPT did for text."
Partnership with AWS
Through a strategic partnership with Amazon Web Services (AWS), Nexus is integrated directly into the AWS dashboard. AWS customers can deploy the model using their existing credits and infrastructure. Frankel describes it as "Very unique agreement," Noting Fundamental is one of only two AI companies to establish such a deep, multi-layered partnership with Amazon.
One of the most significant barriers to enterprise AI is data privacy. Companies are often unwilling to transfer sensitive data to third party infrastructure.
To solve this, Fundamental and Amazon achieved a major engineering feat: the ability to deploy a fully encrypted model – both architecture and weights – directly into the customer’s own environment. "Customers can be assured that the data belongs to them," Frankel said. "We are the first and currently the only company to create such a solution."
The emergence of Fundamentals is an attempt to redefine the OS for business decisions. If NEXUS performs as advertised – handling financial fraud, energy prices and supply chain disruptions with a single, generalized model – it will mark the moment where AI finally learns to read the spreadsheets that actually run the world. The power of prediction is no longer about seeing what happened yesterday; It’s about uncovering the hidden language of the tables to determine what will happen tomorrow.
<a href