Nvidia is pitching transaction foundation models as the next architectural shift in financial services, and the customer list is starting to look serious. Revolut, Mastercard, Adyen and Stripe are each training transformer-based models directly on proprietary payment data, replacing the sprawl of task-specific fraud, credit and recommendation systems banks have accumulated over the past decade. Nvidia's own 2026 State of AI in Financial Services report puts the context in numbers: 65% of institutions now use AI, nearly 90% are deploying or assessing it, and 42% are already using or assessing agentic AI.
The flagship deployment is PRAGMA, a family of transformer models Revolut built with Nvidia. PRAGMA was trained on 24 billion events across 26 million user records spanning over 100 countries, running on Nvidia Hopper GPUs with the cuDF library and Nemotron open models on Nebius cloud. A single foundation model outperformed Revolut's strong task-specific models across credit scoring, fraud detection and product recommendations, while cutting handcrafted feature engineering close to zero.
The pitch to data science teams is speed. Instead of maintaining one model per use case per market, teams train one representation of consumer behavior and fine-tune downstream.
“We move from weeks, or even in some cases months, in feature engineering to no time required for it at all.”— Tadas Kriščiūnas, Head of group credit data science at Revolut
Key facts
- 01Revolut's PRAGMA foundation model was trained on 24 billion events across 26 million user records spanning over 100 countries.
- 02Nvidia's 2026 State of AI in Financial Services report finds 65% of institutions now use AI and nearly 90% are deploying or assessing it.
- 03Adyen has processed $1 trillion in payments using transaction foundation models, with reinforcement learning tuning conversion and risk.
- 04Stripe blocked close to $112 billion in fraud last year and reports an average 38% reduction in fraud rates.
- 05GFT's Wynxx platform is in use at over 100 financial institutions; its Smaragd compliance engine cuts false positives by up to 75%.
Mastercard is taking the same approach at a larger scale. The company is developing a proprietary large tabular foundation model trained on billions of anonymized transactions today, designed to scale to hundreds of billions as it ingests fraud, authorization, chargeback, merchant location and loyalty data. The build uses Nvidia's NeMo AutoModel library and NeMo framework alongside AWS and Databricks. Early testing shows the unified model beating standard machine learning across cybersecurity, fraud detection, loyalty and portfolio optimization.
Adyen is the scale proof point. The payments company has processed $1 trillion in payments using transaction foundation models, applying reinforcement learning to maximize merchant conversion while holding down fraud loss.
That 0.1% figure is the entire argument for foundation models in payments: at trillion-dollar volumes, basis points of authorization lift dwarf the cost of GPU clusters.
“Even fractional improvements like a 0.1% uplift in authorization can translate to massive incremental gross merchandise value and substantial cost reductions.”— Dhruv Ghulati, Principal AI product manager at Adyen
Stripe is running the same play. Using the Nvidia and AWS platform, Stripe built foundation models that read full transactional context rather than scoring isolated signals, blocking close to $112 billion in fraud last year and posting an average 38% reduction in fraud rates. The bet is that proprietary transaction history — the data a competitor cannot buy or scrape — becomes the durable moat once everyone has access to the same open-weight models.
The agentic angle matters here too. With 42% of financial firms using or assessing agentic AI, the systems making and routing payments are increasingly software, not humans. A model that understands context across a customer's entire transaction history is better positioned to authorize, decline or escalate than a stack of single-purpose classifiers stitched together with rules.
Nvidia is also opening the architecture to non-flagship customers. A Build Your Own Transaction Foundation Model developer example is now available on AWS via Amazon SageMaker HyperPod and on Nebius AI Cloud, with Nebius supporting the full lifecycle through to managed inference on Token Factory. Services partners EXL, Infosys, GFT IT Consulting and Thoughtworks are wiring the example into their delivery practices — EXL through its EXLerate.ai platform, GFT through Wynxx, used by over 100 financial institutions, and its Smaragd compliance engine, which it says reduces false positives by up to 75% for major banks.
The skeptical read is that foundation-model economics in banking are still unproven outside the largest payment processors. A 0.1% authorization lift is meaningful at Adyen's scale; at a mid-size regional bank with fewer events, the GPU and engineering bill may swallow the benefit before it shows up in net income. Model governance, explainability for regulators, and concentration risk on a single Nvidia-anchored stack are also unresolved — every transformer the industry trains pulls more decisioning logic into systems that supervisors have not yet figured out how to examine.
For Nvidia, the financial services beat is shaping up like the manufacturing and telco beats: a reference architecture, a developer example, two or three lighthouse customers with eye-catching numbers, and a partner ecosystem to fan it out across the long tail. The interesting question is whether banks treat transaction foundation models as a shared infrastructure layer — the way they treat card networks — or as a proprietary edge each builds alone. The Mastercard and Revolut announcements suggest the latter, which is good news for Nvidia's GPU order book and bad news for any vendor hoping to sell a one-size-fits-all fraud model into this market over the next five years.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
