European Banks Are Building Their Own Models on Open Weights and Transaction Data
Revolut, Lloyds, bunq and others are training models on their own transaction data and running open weights on their own infrastructure, with NVIDIA GPUs at the core.
The most interesting thing about the European financial services announcements from GTC Berlin isn't any single model. It's the pattern. Banks, payment companies and financial data providers are running production systems on open weight models, trained or tuned on data they own, and hosted on infrastructure they control. The compute underneath is NVIDIA accelerated computing.
Foundation Models Trained on Transactions
Revolut is the most ambitious case. The company is rebuilding its technology stack around a single foundation model, PRAGMA, running on NVIDIA accelerated computing. The production setup uses the NVIDIA cuDF toolkit and Dynamo-Triton software on NVIDIA Blackwell Ultra GPUs. It currently serves an initial 80 million customers and is engineered to scale 3x.
Pavel Nesterov, head of AI at Revolut, framed the problem this way: "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." He added that the work was "the clearest proof yet that the gap between a research breakthrough and a production banking system can be closed."
Lloyds Banking Group has taken a similar route. Its ATUM model is a transaction foundation model that turns billions of banking transactions into reusable representations. It runs on NVIDIA RTX PRO 6000 Blackwell GPUs and is being applied across several bank-wide use cases. The reusable-embedding idea matters here: one model trained once on transaction history can feed many downstream tasks, instead of each team building its own narrow classifier from scratch.
Starling Bank is working on a different axis. It's described as the first bank to train an agentic model on U.K. sovereign cloud infrastructure. Its "Starling Intelligence" layer, which underpins its AI capability, was developed on Civo's U.K. cloud using the bank's own transaction data. SBS, meanwhile, has launched a banking sovereign intelligence platform designed so that banks can run generative and agentic models without their data leaving their own environment.
Open Weights and the Cost Equation
The economic argument is the one that will land with most engineering leads. Open weight models let a bank build proprietary intelligence on data it already holds, at costs it can model and control. Closed API pricing scales with usage in ways that are hard to forecast, while self-hosted open weights shift the question toward GPU capacity and engineering time.
bunq offers the clearest number. Its agentic assistant, Finn, now runs on open weight models. Deployed on Nebius with NVIDIA accelerated computing, the company reports quality that matches leading closed models at 60% of the cost per answer. That's a claim about per-answer economics on bunq's workload, so the figure is most useful as a benchmark for a similar assistant, not a universal price cut.
Other deployments point at different model types. NatWest is building speech technology meant to recognize the full range of British accents, dialects and financial terminology. It's developed with NVIDIA's speech models, including Nemotron 3.5 ASR. Speech recognition tuned to regional accents is a practical problem for any bank running voice channels, and it's a good example of where a general-purpose model tends to underperform without domain adaptation.
Modulr is building specialized models for distinct parts of payment operations, based on the NVIDIA Nemotron-3-Ultra-550B open model. The approach here is to start from a large open base and specialize it for narrow tasks, rather than train a single generalist. Santander, for its part, is an early adopter of the NVIDIA Kumo family of open weight foundation models for structured financial data. It's using them to build specialized intelligence for core banking workloads. Structured tabular data is still where many banks have the most valuable and least-exploited information, so a model family aimed at that format is worth watching.
Research Workloads and Data Pipelines
Not all of this is about customer-facing AI. LSEG is working with NVIDIA to help financial institutions analyze large volumes of historical tick data up to 10x faster. The method uses the Polars GPU engine, accelerated by NVIDIA cuDF on NVIDIA RTX PRO 6000 Blackwell GPUs. For quant teams, backtesting speed often determines how many hypotheses get tested in a given week, so a 10x gain changes research throughput directly.
RavenPack built a dataset generation pipeline on Bigdata.com. It takes a plain-language research hypothesis and returns a labeled, backtest-ready dataset, running on a fine-tuned NVIDIA Nemotron model. Investors can use these datasets to identify, analyze and visualize market trends. The interesting design choice is the interface: a natural-language hypothesis as the input, with data preparation handled by the model. Whether that reliably produces sound datasets is the kind of question backtesting discipline has to answer, and the announcement doesn't address it in detail.
Faster Pipelines for Fraud Detection
Fraud is where data latency becomes a business constraint. NVIDIA says RTX PRO 6000 Blackwell Server Edition GPUs and cuDF help accelerate processing across multi-terabyte, siloed and regulated data sources, shortening the time from raw data to decision.
HSBC provides near-real-time client insights that support customer processes and on-demand decisions while minimizing risk and fraud. It reports 5x faster data processing on Starburst Enterprise's GPU Engine.
Stripe reports the largest figures in the group. For its Stripe Radar fraud-detection platform, it sees 42x faster data pipelines and 92% cost savings using NVIDIA GPU-accelerated Ray Data. Those results come from offline testing, not live production traffic, and that distinction matters. Offline gains in pipeline throughput don't always carry over unchanged once real-time serving, feature freshness and model-retraining cadence enter the picture. What the figures do show is that Stripe is treating pipeline speed as a lever for model quality: faster processing lets its Radar machine learning pipelines retrain more continuously.
What to Watch
Most of the numbers here are vendor-reported and drawn from specific workloads, so they're directional rather than transferable. The 60% cost figure depends on bunq's prompts, traffic and quality bar. The 42x and 92% figures come from offline testing. Teams evaluating a similar move should replicate the comparison on their own data before drawing conclusions. The licensing and operational questions are real, too. Open weights shift responsibility for serving, updating, evaluating and securing the model onto the bank. Each open model also carries its own license terms, and those need to be read against the intended use. A bank that self-hosts gains control over data flow but takes on the MLOps burden that a managed API would otherwise absorb. Still, the direction is clear. The question for most institutions is no longer whether a foundation model can work on transaction or market data. It's how much of the stack they want to own, and how quickly they can measure whether a specialized open model beats a general one on the tasks that actually pay the bills.
