European Finance Firms Deploy NVIDIA Open AI Models on Sovereign Infrastructure
European finance firms deploy NVIDIA-powered open models, cutting AI costs, accelerating research and improving fraud detection on sovereign infrastructure.
Summary
At GTC Berlin, NVIDIA detailed European financial AI deployments using open weight models, accelerated computing and institution-controlled infrastructure. Revolut’s PRAGMA uses cuDF and Dynamo-Triton on Blackwell Ultra GPUs, initially serving 80 million customers with capacity engineered to triple while processing billions of transactions. AI head Pavel Nesterov said the foundation model spans banking domains. Starling Bank trained agentic model Starling Intelligence on its transaction data in Civo’s UK sovereign cloud. Lloyds Banking Group’s ATUM converts billions of transactions into reusable intelligence on RTX PRO 6000 Blackwell GPUs, while SBS keeps generative and agentic models inside banks’ infrastructure.
bunq’s Finn, deployed on Nebius, matches leading closed models at 60% of the cost per answer. NatWest is developing British accent, dialect and financial terminology recognition with NVIDIA speech models including Nemotron 3.5 ASR. Modulr is building payment operations models from Nemotron-3-Ultra-550B, and Santander is an early adopter of Kumo open weight models for structured core banking data.
LSEG combines the Polars GPU engine, cuDF and RTX PRO 6000 Blackwell GPUs to analyze historical tick data up to 10x faster. RavenPack’s Bigdata.com pipeline uses a fine-tuned Nemotron model to turn plain-language hypotheses into labeled, backtest-ready datasets for identifying, analyzing and visualizing market trends. RTX PRO 6000 Server Edition GPUs and cuDF accelerate multi-terabyte, siloed regulated data. HSBC achieves 5x faster processing with Starburst Enterprise’s GPU Engine for near-real-time client decisions and fraud reduction. In offline testing, NVIDIA Ray Data made Stripe Radar pipelines 42x faster and cut costs 92%, supporting continuous training optimization and improved fraud detection.
Positives
- Revolut’s PRAGMA initially serves 80 million customers and is engineered to support three times that scale.
- bunq’s Finn matches leading closed-model quality on Nebius at 60% of the cost per answer.
- LSEG accelerates historical tick-data analysis by up to 10x using Polars, cuDF and RTX PRO 6000 Blackwell GPUs.
- HSBC achieves 5x faster processing through Starburst Enterprise’s GPU Engine for near-real-time insights and fraud reduction.
- Stripe Radar’s offline tests delivered 42x faster data pipelines and 92% cost savings with NVIDIA Ray Data.
- Starling Bank trained its agentic model on UK sovereign cloud infrastructure using its own transaction data.
Risks & concerns
- Stripe’s 42x speed gain and 92% savings come from offline testing, with production results not provided.
- Revolut’s threefold scalability figure describes engineered capacity beyond its initial 80 million customers, not a reported live deployment.
- NatWest and Modulr are still building their systems, with no production performance or cost figures disclosed.
- Financial institutions must still process multi-terabyte data distributed across siloed, regulated sources.