Mastercard builds a tabular foundation model to fight payment fraud

Mastercard has introduced a new type of foundation model for payment infrastructure: a large tabular model (LTM) trained not on text but on arrays of transaction data. The company expects it to detect fraud more accurately and verify the authenticity of transactions in digital payments.

How the LTM works

Unlike LLMs, which work with unstructured data and predict the next token, an LTM is trained on tables with a large number of features. Mastercard's dataset includes billions of card transactions, and the company wants to grow it to hundreds of billions of records over time. Training used not only the payments themselves but also related fields: merchant geography, authorization chains, confirmed fraud cases, chargebacks and loyalty program activity. The model's job is to find behavioral links between features and spot anomalies that predefined rules do not cover.

Mastercard stresses that personal identifiers were removed from the data before training. By design, the model should analyze not the identity of the cardholder but the nature of behavior within the flow of transactions. This reduces some of the privacy risks that usually come with AI systems in finance, although it also removes some signals that could be useful for assessing risk. The company believes the loss of accuracy can be offset by the scale of the sample and the richness of context.

Nvidia and Databricks provided the technical platform for the project.

Where the model will be used

Cybersecurity and anti-fraud are the first deployment areas. Mastercard already has several systems that track suspicious transactions, but many of them depend on manual configuration: analysts define patterns such as a sharp rise in purchase frequency or transactions from different countries within a short time. The LTM is meant to complement these mechanisms and to see complex combinations of features better, without a rigid set of rules.

According to the company, the effect is most noticeable for rare and expensive purchases, which traditional models often flag as suspicious even when the transaction is legitimate.

The company also sees other uses for the model within payment infrastructure:

• analysis of loyalty program activity
• internal analytics on portfolios and transactions
• decision support in cybersecurity systems
• building new internal applications through APIs and SDKs

Mastercard does not plan to replace existing tools with a single model right away. Instead, the current plan is to build hybrid systems in which the LTM works alongside proven procedures and detectors. The cautious approach is understandable: errors are costly in the payments industry, and regulatory requirements are high. On the other hand, a single foundation model that can be fine-tuned for different scenarios could reduce the cost of training dozens of separate models, and of validating, monitoring and maintaining them.

Risks and limitations

The approach has weak spots. If a multipurpose model becomes widely used and starts making systematic errors, the consequences can hit several products or processes at once. That is why Mastercard is not yet making the LTM the sole arbiter of disputed transactions.

Explainability is another issue: in anti-fraud and credit processes, it is not enough for a business to receive a risk signal; it also has to show why the system reacted the way it did. Without transparency and auditability, such models are harder to defend before regulators and internal compliance.

There are also more practical questions that the market has not yet had independently answered. The claims about effectiveness come from Mastercard itself, so they cannot be treated as final proof that the LTM beats conventional ML approaches. It is not entirely clear how the model will behave under attack, how much it will cost to run over the long term after training, or how willing regulators will be to accept this class of systems in critical financial infrastructure. These factors, not just benchmark quality, will determine the pace of adoption.

What this means

Mastercard shows that the next wave of AI in finance may be built not around chatbots but around foundation models for tabular data. If the LTM really reduces false positives and simplifies work across several anti-fraud scenarios, banks and payment providers will start investing more actively in such systems instead of a collection of narrow models for each task.