Our philosophy at Fundamental is straightforward: customers shouldn't have to move their data to use a foundation model. The model should come to where the data already lives.

The structured data inside an enterprise (patient records, transaction histories, claims, underwriting files, customer accounts, supply chain telemetry etc) is the most sensitive and most regulated data the business holds. It's the data that drives the highest-value decisions and, if mishandled, it’s the data that creates the biggest exposure. Moving it into a vendor environment is a non-starter for most large enterprises.

Today we're taking another step toward solving that. NEXUS, our most powerful Large Tabular Model, is now available on Amazon SageMaker. AWS customers can deploy it as a single tenant, network isolated SageMaker instance from their own account that no other party can access.

Why this matters

That sensitive, high-stakes data is also where making predictions has been hardest to do efficiently and accurately. 

For decades, the only way to get predictions out of structured enterprise data was to spin up a bespoke ML project for every question. This required a team of data scientists, months of feature engineering, a separate training job, a separate model artifact and a separate endpoint. And then repeat the process for every use case. Most companies never got past the first few.

NEXUS replaces that pattern with a single foundation model trained on billions of real-world tables. It eliminates all the friction and inefficiency of the traditional ML approach: simply point it at your data and tell it what you want to predict. 

Every enterprise data leader we've spoken to quickly grasps why NEXUS changes everything. But then the same question always follows: “Can we run this against our real data, in a secure environment, without anything ever leaving it?” 

The data never leaves

Security and privacy are foundational considerations the SageMaker deployment is built on. Every architectural decision starts there, and every other capability follows from it.

NEXUS deploys as a dedicated, single-tenant endpoint inside the customer's own AWS account. Datasets stay in the customer's S3 buckets. Inputs, outputs, and model artifacts never transit a Fundamental-controlled environment. The container runs network-isolated, with no outbound calls during inference, so even the model itself can't exfiltrate data. IAM, VPC, encryption keys, logging, and audit all stay inside the security perimeter the customer has already built, hardened, and signed-off on.

For enterprises operating under healthcare privacy law, European data protection rules, payments industry standards, financial reporting controls, or any of the dozens of sector-specific frameworks that govern structured business data, this matters more than any benchmark. 

A foundation model that requires moving sensitive records into a third-party environment is, in practice, a foundation model most regulated enterprises cannot use. NEXUS on SageMaker removes that constraint without compromise: the same model running on the same dedicated infrastructure, inside the cloud account the enterprise already trusts.

One endpoint, many use cases

The architectural implications go beyond security.

Traditional ML pipelines treat every prediction problem as its own project: separate training, artifacts, infrastructures and endpoints. Run that pattern across an enterprise and you get sprawl: hundreds of pipelines, dozens of teams and no consolidated view of cost or performance.

A single NEXUS endpoint dynamically hosts multiple trained models, securely serving different internal use cases simultaneously on your dedicated compute. Demand forecasting, churn, fraud, underwriting, predictive maintenance - all from one deployment. That collapses infrastructure cost, simplifies governance, and lets data science teams move from one prediction problem to the next without standing up new infrastructure for each one.

For enterprises operating dozens or hundreds of predictive workflows, this is a meaningful change in the economics of running prediction at scale.

How it works

Customers subscribe to the NEXUS model package through AWS Marketplace, deploy a SageMaker asynchronous inference endpoint, and connect using the Fundamental Python SDK. From there the interface is the one data science teams are already familiar with such as fit(), predict(), and predict_proba().

Under the hood, the SDK handles serialization, endpoint invocation, asynchronous polling, and model management. The infrastructure is built for enterprise-scale workloads: gigabyte-scale datasets, dedicated GPU instances, asynchronous inference for long-running jobs, and S3-native data movement throughout.

The whole flow is designed so that a data scientist who already lives in the AWS ecosystem can go from raw data to a deployed prediction system in a session, not a quarter.

The bigger picture

We started Fundamental because we believed the same scaling story that played out for unstructured data with LLMs was going to play out for structured data with LTMs. The foundation model exists. The accuracy is there. The remaining question for most enterprises has always been deployment: can we actually run this against our real data, inside our real environment, under our real security model?

SageMaker is one of the most important answers to that question. It puts NEXUS directly inside the cloud environment most large enterprises already trust, with the security posture they've already built around.

This is one step toward making NEXUS-powered prediction the default for enterprise structured data.