Financial services
How banks, payment companies and trading firms use NVIDIA systems and software for fraud scoring, forecasting, trading research and employee agents, and how model risk rules, customer data protection and fair lending obligations shape each design.
The problem
Banks are squeezed between thin margins and rising operating, technology and compliance costs, while the data that could help them sits in separate systems for cards, deposits, payments and markets. The stakes in card fraud are large: a Payments Dive estimate cited by NVIDIA puts worldwide card transaction fraud losses above 403 billion dollars over the next decade, and fixed rules catch only the patterns someone has already written down.
Trading desks face a different version of the same problem: noisy market data, conditions that shift within a day, and execution costs that eat into returns. Every model that informs a decision about money is itself a source of risk. In the United States, the Federal Reserve, OCC and FDIC replaced the 2011 model risk guidance with SR 26-2 in April 2026, aimed mainly at banks above 30 billion dollars in assets, and in the European Union credit scoring that can deny someone a loan counts as a high-risk AI use.1234
The approach
For fraud, NVIDIA publishes an AI Blueprint that builds features with RAPIDS (part of CUDA-X Data Science), adds graph neural network embeddings as extra inputs to an XGBoost classifier to cut false positives, and serves the model in real time with Dynamo-Triton. NVIDIA says it runs on AWS and HPE and through partners such as Cloudera, EXL, Infosys and SHI International, and that CUDA-X Data Science and Dynamo-Triton are part of NVIDIA AI Enterprise.
Some banks own their compute. BNY installed a DGX SuperPOD with dozens of DGX H100 systems in 2024 to support forecasting of deposits and cash positions, payment automation and trading analytics. NVIDIA also promotes transaction foundation models, trained on large volumes of tabular payment data, as a successor to rules-based detection, and agentic AI that draws on data from across the bank.
A simpler path is often enough. Gradient-boosted tree models on CPUs remain a sound baseline for fraud and credit scoring at moderate volumes, buying fraud scores from a payment processor can cover many needs, and a smaller bank gains more from clean, joined-up data and a documented validation process than from new hardware. GPUs earn their place when data volume, graph size or latency targets outgrow CPU pipelines.125
Conceptual architecture
Diagram as a list
Applications & solutions
- Employee assistants and workflow agents on NVIDIA AI EnterpriseSummarize cases and gather evidence for analysts and relationship staffConnects to Fraud operations, credit and payments decisions with human review
- Fraud operations, credit and payments decisions with human reviewCombine scores with policy and analyst judgment and record the outcomeConnects to Model risk management (inventory, validation, ongoing monitoring)
Models & frameworks
- Scoring and forecasting models (GNN embeddings with XGBoost, transaction foundation models)Produce risk scores and forecasts rather than final decisionsConnects to Real-time inference service (Dynamo-Triton)
Inference & runtime software
- Real-time inference service (Dynamo-Triton)Scores each transaction within the payment authorization windowConnects to Fraud operations, credit and payments decisions with human review
Operations & orchestration
- Transaction, account and market data (core banking, card, payments and market feeds)Supplies labeled history and live events under access and retention rulesConnects to GPU feature engineering and graph building (CUDA-X Data Science)
- Model risk management (inventory, validation, ongoing monitoring)Approves models before use and tracks drift, bias and performanceConnects to Scoring and forecasting models (GNN embeddings with XGBoost, transaction foundation models)
Accelerated computing
- GPU feature engineering and graph building (CUDA-X Data Science)Turns raw records into features and account-merchant graphs at scaleConnects to Model training on owned or rented GPUs (DGX systems or cloud instances)
- Model training on owned or rented GPUs (DGX systems or cloud instances)Trains fraud, forecasting and transaction models on sensitive dataConnects to Scoring and forecasting models (GNN embeddings with XGBoost, transaction foundation models)
Technologies and their roles
NVIDIA CUDA-X Data Science2
GPU data preparation for fraud and risk models
NVIDIA's fraud detection blueprint uses RAPIDS, part of CUDA-X, to process transaction data and build features, and CUDA-X libraries to produce graph neural network embeddings.
NVIDIA Dynamo-Triton2
Real-time model serving
NVIDIA names Dynamo-Triton in the fraud blueprint for real-time inference with attention to throughput, latency and utilization.
NVIDIA Blueprints2
Reference workflow for card fraud detection
NVIDIA's AI Blueprint for financial fraud detection supplies reference code, deployment tools and an architecture, currently tuned for credit card fraud.
NVIDIA AI Enterprise26
Supported software layer for bank AI applications
NVIDIA says AI Enterprise includes CUDA-X Data Science and Dynamo-Triton, and BNY said it planned to use AI Enterprise to build AI applications and manage its infrastructure.
NVIDIA DGX6
Owned training and inference infrastructure
BNY runs a SuperPOD made of DGX H100 machines for workloads such as deposit forecasting, payment automation and trade analytics.
What you need first
- A model inventory and validation process that covers machine learning and generative models, not only traditional statistical ones
- Labeled fraud and dispute history with consistent definitions across card, account and payment channels
- A latency budget for each scoring point, such as card authorization or instant payment release, agreed with operations
- Clear rules on which customer data may leave the bank, be used for training, or be sent to an external model
- Explainability and adverse action processes for any model that influences credit or account decisions
- A cost comparison of CPU, cloud GPU and owned GPU options based on measured workloads, not vendor estimates
- Fraud analysts and risk owners involved in setting score thresholds and reviewing false positives
Risks and how to reduce them
- Decisions driven by unvalidated or drifting models
- Put every model through the bank's model risk process before use, monitor performance and drift continuously, and keep a documented fallback such as the previous model or rules.
- Unfair or unexplainable credit outcomes4
- Test models for disparate impact before and after launch, keep reason codes for adverse actions, and plan for EU high-risk obligations such as data quality, logging and human oversight where they apply.
- Customer data exposure through generative AI tools
- Keep sensitive data in controlled environments, filter prompts and outputs, log access, and approve each external model or service through vendor risk review.
- Legitimate customers blocked by false positives
- Treat model output as a score, tune thresholds with fraud operations, measure customer friction alongside caught fraud, and give customers a fast way to clear a block.
Documented examples
BNY (The Bank of New York Mellon Corporation) · Banking and financial services
BNY: an on-premises DGX SuperPOD with DGX H100 for banking AI
BNY installed an NVIDIA DGX SuperPOD built from dozens of DGX H100 systems in March 2024, the first major bank to do so by NVIDIA's account. Named workloads include deposit forecasting and payment automation. No performance or business results have been published for the system.
In production
Related
Sources
- AI Solutions for Finance Industries (opens in a new tab)
- Bring Receipts: New NVIDIA AI Blueprint Detects Fraudulent Credit Card Transactions With Precision (opens in a new tab)
- SR 26-2: Revised Guidance on Model Risk Management (opens in a new tab)
- AI Act: regulatory framework for AI (opens in a new tab)
- BNY Mellon, First Global Bank to Deploy AI Supercomputer Powered by NVIDIA DGX SuperPOD With DGX H100 (NVIDIA blog post reprinted by BNY) (opens in a new tab)
- BNY First Global Bank to Deploy AI Supercomputer Powered by NVIDIA DGX SuperPOD With DGX H100 (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.