European financial institutions are deploying open weight artificial intelligence models on their own infrastructure and sovereign cloud systems, according to an announcement from Nvidia at GTC Berlin.
Revolut is rebuilding its technology stack around a single foundation model named PRAGMA, Nvidia reported. The digital bank runs the model using Nvidia Blackwell Ultra GPUs alongside the company's cuDF toolkit and Dynamo-Triton software. Revolut engineered PRAGMA to serve an initial customer base of 80 million people, with capacity designed to scale threefold.
“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.”
Sovereign infrastructure and banking models
Other European lenders are keeping model training on local hardware. Starling Bank trained an agentic model on U.K. sovereign cloud infrastructure provided by Civo, using its own transaction records to create an intelligence layer called Starling Intelligence. Lloyds Banking Group built ATUM, a transaction foundation model accelerated by Nvidia RTX PRO 6000 Blackwell GPUs to convert banking transactions into intelligence across bank-wide operations. Software vendor SBS launched a platform designed to let banks run generative and agentic models without data leaving their premises.
Neobank bunq shifted its agentic assistant, Finn, to open weight models deployed on Nebius infrastructure with Nvidia computing. That setup matched the response quality of leading closed models at 60 percent of the cost per answer, Nvidia said. NatWest is using Nvidia speech models, including Nemotron 3.5 ASR, to recognize British accents, dialects and financial terms. Modulr is developing models for payment workflows based on the Nemotron-3-Ultra-550B model, while Santander adopted Nvidia's Kumo family of open models to analyze structured financial records for core banking tasks.
Data processing and fraud prevention
Financial data providers are using the hardware to process market activity. The London Stock Exchange Group is working with Nvidia to query historical tick data up to 10 times faster by running the Polars GPU engine with cuDF on RTX PRO 6000 Blackwell GPUs. RavenPack created a pipeline on Bigdata.com that converts written research hypotheses into labeled backtesting datasets using a fine-tuned Nemotron model.
Nvidia reported that its RTX PRO 6000 Blackwell Server Edition GPUs and cuDF library accelerate multi-terabyte data workflows across regulated databases. HSBC increased data processing speeds fivefold on Starburst Enterprise's GPU Engine to deliver near-real-time client information and limit fraud risk. In offline testing for its Stripe Radar fraud detection platform, payment company Stripe achieved 42 times faster data pipelines and a 92 percent reduction in costs using Nvidia GPU-accelerated Ray Data.
