Revolut builds AI transaction model for 80 million customers

Revolut is rebuilding its technology stack around PRAGMA, a single-transaction foundation model now deployed in production for an initial 80 million customers.

PRAGMA runs on NVIDIA-accelerated computing, using the cuDF toolkit, Dynamo-Triton software, and NVIDIA Blackwell Ultra GPUs. The system has been engineered to scale threefold as transaction volumes grow.

The move puts Revolut among a wider group of banks, payment companies, and financial data providers in Europe that are putting artificial intelligence models into live financial workloads on their own infrastructure. Several are using open-weight models trained or adapted with internal datasets, rather than relying only on closed third-party systems.

According to the companies involved, the work spans retail banking, payment operations, speech recognition, market research and fraud detection. The common thread is a push to use models trained on proprietary financial data inside regulated environments.

"Revolut processes billions of transactions - the opportunity was always there, but the question was whether foundation models could operate at that scale and across domains. PRAGMA answers that," said Pavel Nesterov, Head of AI at Revolut.

"Built in collaboration with NVIDIA, it's the clearest proof yet that the gap between a research breakthrough and a production banking system can be closed," Nesterov said.

Banking models

Other banks are building similar model layers around transaction data. Lloyds Banking Group has developed ATUM, a transaction foundation model that turns billions of banking transactions into reusable intelligence for use across the group. It is using NVIDIA RTX PRO 6000 Blackwell GPUs for that work.

Starling Bank has trained what it describes as an agentic model on UK sovereign cloud infrastructure, using its own transaction data on Civo's UK cloud. The bank describes Starling Intelligence as the layer underpinning its internal AI work.

SBS has launched a sovereign intelligence platform for banks that want to run generative and agentic models without moving data outside their own environments. The move reflects a broader industry focus on data residency and control as financial groups test larger models in regulated settings.

Open models

European financial institutions are also testing whether open-weight models can cut costs while keeping tighter control over data and deployment. Bunq said its assistant Finn now runs on open-weight models, and that deployment on Nebius matches the quality of leading closed models at 60% of the cost per answer.

NatWest is building speech technology intended to recognise British accents, dialects and specialist financial terminology. It is developing the work with NVIDIA speech models, including Nemotron 3.5 ASR.

Payments company Modulr is building specialised models for separate parts of payment operations based on the NVIDIA Nemotron-3-Ultra-550B open model. Santander, meanwhile, is using the Kumo family of open-weight foundation models for structured financial data in core banking workloads.

Research workloads

Market data and research groups are also applying the same infrastructure to analytical tasks. LSEG is working with NVIDIA to help financial institutions analyse large volumes of historical tick data up to 10 times faster with the Polars GPU engine, accelerated by cuDF on NVIDIA RTX PRO 6000 Blackwell GPUs.

RavenPack has built a dataset generation pipeline on Bigdata.com that takes a plain-language research hypothesis and produces a labelled, backtest-ready dataset. The pipeline runs on a fine-tuned NVIDIA Nemotron model, allowing investors to identify, analyse and visualise market trends.

Fraud focus

Fraud detection remains one of the clearest commercial uses for faster data processing in finance. NVIDIA said RTX PRO 6000 Blackwell Server Edition GPUs and cuDF can speed work across multi-terabyte and siloed regulated data sources, shortening the time between data processing and a fraud decision.

HSBC said it is using Starburst Enterprise's GPU engine to provide near-real-time client insights that support customer processes and on-demand decision-making while reducing risk and fraud. The bank said this has made data processing five times faster.

Stripe reported 42 times faster data pipelines and 92% cost savings in offline testing for its Stripe Radar fraud detection platform with GPU-accelerated Ray Data. It said this would allow machine learning pipelines to update model training continuously and improve fraud detection.

Together, the projects show how large financial institutions and fintech groups are moving beyond pilot systems and applying AI models to transaction processing, customer service, market analysis and risk controls in production environments. In many cases, the emphasis is less on public chatbots and more on using internal data, sovereign infrastructure and model tuning to address tightly defined financial tasks.