European financial firms adopt open models for proprietary AI

News

NVIDIA says banks, payment firms and data providers in Europe are using open weight models and accelerated computing for proprietary AI systems.

At GTC Berlin, NVIDIA said European banks, payment companies and financial data providers are moving production AI workloads onto their own infrastructure, using open weight models and accelerated computing to build proprietary intelligence. The examples cited by NVIDIA on 22 Oct. 2026 range from customer-service assistants and transaction models to fraud-detection pipelines and financial research tools.

Evidence and context

NVIDIA said Revolut is rebuilding its technology stack on a single foundation model, PRAGMA, using NVIDIA cuDF and Dynamo-Triton software on Blackwell Ultra GPUs. The company said the system is serving an initial 80 million customers and is engineered to scale 3x. Pavel Nesterov, head of AI at Revolut, said Revolut processes “billions of transactions” and that the question was whether foundation models could operate at that scale across domains; he said PRAGMA answers that and was built in collaboration with NVIDIA.

Other institutions cited by NVIDIA included Starling Bank, which it described as the first bank to train an agentic model on U.K. sovereign cloud infrastructure, using its own transaction data on Civo’s U.K. cloud. Lloyds Banking Group built ATUM, a transaction foundation model that converts billions of banking transactions into reusable intelligence, and NVIDIA said it was accelerated with RTX PRO 6000 Blackwell GPUs. NVIDIA also said SBS launched a banking sovereign intelligence platform designed so banks can run generative and agentic models without their data leaving their own walls.

NVIDIA framed the deployments as evidence that open weight models can support proprietary intelligence at controlled cost. It said bunq’s agentic assistant Finn now runs on open weight models and that deploying on Nebius matches the quality of leading closed models at 60% of the cost per answer. NVIDIA also said NatWest is building technology to recognize British accents, dialects and financial terminology using its speech models, including Nemotron 3.5 ASR, while Modulr is building specialized models for payment operations on the NVIDIA Nemotron-3-Ultra-550B open model. Santander, NVIDIA said, is an early adopter of the Kumo family of open weight foundation models for structured financial data.

In research and data processing, NVIDIA said LSEG is working with it to analyze large volumes of historical tick data up to 10x faster with the Polars GPU engine, accelerated by cuDF on RTX PRO 6000 Blackwell GPUs. It also said RavenPack built a dataset-generation pipeline on Bigdata.com that takes a plain-language research hypothesis and returns a labeled, backtest-ready dataset, running on a fine-tuned NVIDIA Nemotron model.

For fraud and payments operations, NVIDIA said RTX PRO 6000 Blackwell Server Edition GPUs and cuDF are helping accelerate data processing across multi-terabyte, siloed, regulated sources. It said HSBC uses 5x faster data processing on Starburst Enterprise’s GPU Engine to provide near-real-time client insights, while Stripe reported 42x faster data pipelines and 92% cost savings for its Stripe Radar fraud-detection platform in NVIDIA’s offline testing. Those figures are vendor-reported or company-reported performance claims and were not independently verified in the source material.

The announcement adds to a sequence of NVIDIA-financial-services disclosures in 2026, including a 21 Aug. 2026 technical blog on GPU-accelerated clustering for financial instruments, a 9 July 2026 post on synthetic data generation for financial AI research and a 27 May 2026 post on Blackwell setting a STAC-AI record for LLM inference in finance. In January 2026, NVIDIA also said a survey of more than 800 financial-services professionals found that 65% of respondents’ companies actively used AI, up from 45% in the preceding report; that survey described respondents, not the full industry.

The source material does not provide independent audit data, market share, investment totals or a single cross-company benchmark. The claims cited here should be read as company- or vendor-reported deployment descriptions and performance results, with dates attached where supplied.