Mastercard is training a foundation model on payments rather than language. Its large tabular model, or LTM, learns relationships across billions of card transactions, merchant locations, authorisation flows, fraud incidents, chargebacks and loyalty activity.
The objective is to detect behavioural patterns that fixed rules and smaller task-specific models may miss. Mastercard plans eventually to expand training to hundreds of billions of transactions, positioning structured financial data as the basis for a new kind of reusable intelligence.
Personal identifiers were removed before training. The company argues that enormous volumes of behaviour can preserve useful signals while reducing privacy exposure.
A table demands different intelligence
Large language models learn from sequences of unstructured tokens. Mastercard’s LTM examines relationships among fields in multidimensional tables. It can infer which combinations are predictable and identify anomalies that were not explicitly described in a rule.
That matters in fraud detection. Conventional systems often begin with human-defined indicators such as unusual transaction frequency or rapid purchases in distant locations. Those rules remain useful but can flag legitimate behaviour when the context is rare.
Mastercard cites high-value, low-frequency purchases as an early example. Traditional models may treat them as suspicious simply because they are unusual, while the LTM appears better able to recognise when the surrounding pattern is legitimate. The evidence currently comes from the company, so performance still needs broader validation.
Nvidia supplies the computing platform and Databricks supports data engineering and model development. Mastercard describes the result as an insights engine that augments existing products rather than a customer-facing assistant.
One foundation can support many financial tasks
Payments companies often train separate models for fraud, loyalty, portfolio analysis and internal reporting. Each one requires development, validation, monitoring and maintenance.
A foundation model fine-tuned for several structured-data tasks could reduce that duplication. Mastercard plans APIs and software-development kits so internal teams can build applications on the shared base. Loyalty activity and portfolio management are among the potential uses.
The approach also concentrates risk. If many products depend on one model, a hidden weakness could spread across operations. Mastercard therefore plans hybrid deployments that combine the LTM with established detection procedures. No single model is expected to perform best in every scenario.
Hybrid design provides comparison and fallback. Existing systems can continue handling cases they understand well, while the LTM adds broader pattern recognition. Disagreements can be routed for review instead of allowing a new model to silently replace proven controls.
Scale makes governance more important
Removing personal identifiers reduces some privacy concerns but does not eliminate responsibility. A system influencing fraud outcomes must remain explainable and auditable, particularly when a legitimate payment is blocked or an illicit one passes.
Large behavioural datasets may also contain historical patterns that do not remain stable. Fraud tactics change, merchants evolve and consumer behaviour shifts. Monitoring after deployment is therefore as important as training scale.
Adversarial robustness will be a defining test. Attackers actively probe payment controls and adapt when patterns become visible. Mastercard must understand how the LTM behaves under manipulation, how quickly it can be updated and what operating costs continue after training.
The project shows that foundation models need not be conversational. Highly structured domains can support models whose intelligence lies in relationships among rows and columns rather than words and images.
For payments, that may be a better match than forcing every problem through an LLM. The LTM can operate quietly inside existing workflows, adding context without speaking to the customer.
Its success will not be measured by a compelling demo. It will be measured by fewer incorrect fraud decisions, controlled operating costs and the ability to explain why a system trusted one unusual transaction and challenged another.
