NVIDIA Says Financial Firms Are Moving From Siloed Fraud and Credit Models to Unified “Transaction Foundation” Models

NVIDIA is arguing that financial institutions are outgrowing task-by-task AI, and is pushing a new model approach meant to unify decisioning across credit, fraud, payments, and recommendations. In an update dated June 1, 2026, the company says many institutions have built effective AI over the years, but that the resulting “sprawl” of purpose-built models runs into a structural problem: customer and transaction data is too often siloed, limiting what AI can understand about a person’s behavior across time, channels, and contexts.

According to NVIDIA, the scale of adoption makes the architecture challenge more urgent. Its “State of AI in Financial Services” survey data is cited in the post: 65% of institutions are using AI, nearly 90% are deploying or assessing it, and almost all are maintaining or increasing spend. As these systems scale, NVIDIA says complexity and fragmented model architectures become the limiting factor, rather than the availability of models for individual use cases.

NVIDIA’s thesis is that transformer-based “transaction foundation models” can learn a single representation of customer behavior from proprietary transaction events such as payments, transfers, product interactions, and behavioral indicates. Unlike a traditional fraud model that evaluates isolated indicates, NVIDIA says a foundation model interprets behavior in context, where timing, device, location, and prior activity change the meaning of what would otherwise look like a similar event. NVIDIA gives an example in which a payment made at midnight can announcement something different if it is the fourth transaction in 10 minutes, from an unfamiliar device, in a city where the customer has no prior history.

The post points to Revolut as an example of using this architecture at scale. In collaboration with NVIDIA, Revolut built PRAGMA, described as a family of transformer-based foundation models trained on 24 billion events across 26 million user records spanning over 100 countries. NVIDIA says the training and deployment stack included NVIDIA Hopper GPUs, NVIDIA cuDF for GPU-accelerated data processing, and NVIDIA Nemotron open models, running on Nebius cloud. Revolut’s head of group credit data science, Tadas Kriščiūnas, is quoted saying teams can move from weeks or months of feature engineering to essentially no time required for that work.

To encourage adoption, NVIDIA is also offering a “Build Your Own Transaction Foundation Model” developer example. The blueprint, hosted, describes the tool as a starting point for creating transformer-based embeddings from tabular transaction data. “Embeddings” are numeric representations that capture patterns in structured data, which can then be used as features for other models. The blueprint says it leverages NVIDIA CUDA-X libraries for data processing and NVIDIA NeMo for model fine-tuning, and it emphasizes using embeddings to improve accuracy or reduce false positives on top of existing machine learning systems. citeturn1view1

NVIDIA says the developer example is designed to fit into existing pipelines without requiring teams to rebuild everything from scratch. In the blog post, the company says institutions can run the example on Amazon Web Services (AWS), deployed with Amazon SageMaker HyperPod, or on Nebius AI Cloud. NVIDIA also describes Nebius AI Cloud as supporting the full lifecycle, including multi-node training and managed inference, and says services partners can help apply the approach in more complex environments.

Beyond Revolut, the post names multiple payments and banking players testing or building foundation-style approaches. Mastercard is described as developing a proprietary large tabular foundation model for payments that NVIDIA says is trained on billions of anonymized transactions and intended to scale to hundreds of billions of additional data types, including fraud and authorization information, chargebacks, merchant location, and loyalty data. Adyen is cited as processing $1 trillion in payments using transaction foundation models at scale, with reinforcement learning aimed at maximizing conversion while minimizing risk for merchants.

NVIDIA also ties the push for unified transaction models to the rise of agentic AI, or AI systems that can take actions rather than only generate outputs. The company says 42% of financial firms are already using or assessing agentic AI, and that as these systems execute transactions such as managing subscriptions or routing payments, the nature of financial behavior shifts. Stripe is cited as using the NVIDIA and AWS platform to build foundation models that focus on the full context of transactional behavior, with NVIDIA stating the approach blocked close to $112 billion in fraud last year and delivered an average 38% reduction in fraud rates.

Still, the public material leaves gaps that institutions would likely need to address internally. NVIDIA’s blog and blueprint describe architecture and example capabilities, but they do not provide detailed methodology for how model quality is measured across different institutions, how privacy and governance controls are implemented end to end, or what total cost and operational overhead look like versus a portfolio of specialized models. The outcomes attributed to specific firms, including fraud and credit improvements, appear to come from company-reported results rather than independent benchmarking, and financial regulators may require additional documentation before these systems can be deployed widely.