Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence

At GTC Berlin, NVIDIA is pitching a very specific future for finance: banks and payments firms running production AI on their own infrastructure, using open weight models and accelerated computing rather than handing the whole stack to a closed service. The pitch isn’t abstract. It comes with named systems, real workloads, and claims tied to actual bank operations.

Revolut is rebuilding around a single foundation model called PRAGMA. It runs on NVIDIA accelerated computing, uses cuDF and Dynamo-Triton software, and is deployed on NVIDIA Blackwell Ultra GPUs. Revolut says PRAGMA is already serving its initial 80 million customers and was built to scale 3x. That’s a serious bet for a bank that processes billions of transactions. Pavel Nesterov, Revolut’s head of AI, says the model shows research can be turned into a production banking system.

Other banks are taking their own route, but the pattern is the same. Starling Bank says it is 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 turns billions of banking transactions into reusable intelligence, with NVIDIA RTX PRO 6000 Blackwell GPUs helping with efficiency across bank-wide use cases. SBS has launched a sovereign intelligence platform so banks can run generative and agentic models without their data leaving their own walls.

The economics matter too. bunq says its Finn assistant now runs on open weight models and, on Nebius, matches leading closed models at 60% of the cost per answer. NatWest is using NVIDIA speech models, including Nemotron 3.5 ASR, to handle British accents, dialects and financial terms. Modulr is building specialist models for payment operations on NVIDIA Nemotron-3-Ultra-550B, and Santander is using the new Kumo family for structured financial data.

The same stack is reaching research and fraud work. LSEG says it can analyze historical tick data up to 10x faster with the Polars GPU engine, accelerated by NVIDIA cuDF on RTX PRO 6000 Blackwell GPUs. RavenPack’s Bigdata.com pipeline turns a plain-language research hypothesis into a labeled, backtest-ready dataset using a fine-tuned NVIDIA Nemotron model. And on the fraud side, HSBC says it gets near-real-time client insights with 5x faster data processing, while Stripe says offline testing of its GPU-accelerated Ray Data setup delivers 42x faster pipelines and 92% cost savings for Stripe Radar.