Mastercard Builds a Large Tabular Foundation Model on Anonymised Transaction Data, Starting with Fraud and Cybersecurity

Mastercard has developed a large tabular model, or LTM, a foundation model trained on transaction data rather than text or images to help address security and authenticity issues in digital payments, AI News reported in March 2026. The model has been trained on billions of card transactions, with plans to expand to hundreds of billions, using payment events and related data such as merchant location, authorisation flows, fraud incidents, chargebacks and loyalty activity. Mastercard says personal identifiers were removed before training and that the model learns behavioural patterns rather than individual identities. Unlike a large language model, which predicts the next token in unstructured text, the LTM examines relationships between fields in multi-dimensional data tables and learns which relationships are predictable, so it can flag anomalous patterns that predefined rules miss. Mastercard describes it as an insights engine to be used inside existing products and workflows. Nvidia provides the computing platform, and Databricks handles data engineering and model development. Cybersecurity is the first area of active deployment. Mastercard's existing fraud systems rely on human-defined rules, such as sudden spikes in transaction frequency or purchases in distant places within a short time. In early results, the company says the new model distinguishes legitimate high-value, low-frequency purchases, which traditional models tend to flag, more accurately than its counterparts. Mastercard plans hybrid systems that combine established procedures with the LTM rather than replacing them, partly because of regulatory scrutiny and the risk that a failure in a widely used model could have system-wide effects. The company says the model could also be used for loyalty programmes, portfolio management and internal analytics, potentially replacing many task-specific models with one fine-tuned foundation model, and it plans API access and SDKs for internal teams. AI News notes that the evidence so far comes from vendor reports.