NVIDIA highlights European finance’s shift toward open AI models built on company-owned data

NVIDIA used GTC Berlin to spotlight how leading European financial services organizations are moving from experiments to production-grade artificial intelligence, with a common theme: open weight models paired with computing systems that let companies run workloads on their own infrastructure. The demonstrations, described in an NVIDIA blog, focus on building what participants call proprietary intelligence, using data they control and keeping costs and governance within their boundaries.

Several case studies centered on foundation models that can be adapted for financial tasks, including fraud, transaction intelligence and speech recognition. Revolut said it is rebuilding its technology stack on a single foundation model called PRAGMA, powered by NVIDIA accelerated computing. The company described using NVIDIA’s cuDF data processing toolkit and Dynamo-Triton software on NVIDIA Blackwell Ultra GPUs to scale PRAGMA in production, starting with an initial user base of 80 million customers and engineered to scale 3x.

Revolut’s leadership framed the effort as a test of whether foundation models can handle real-world banking volume. Pavel Nesterov, head of AI at Revolut, said that PRAGMA answers whether foundation models can operate at scale and across domains, and he characterized the collaboration with NVIDIA as closing the gap between research breakthroughs and production banking systems.

Other institutions emphasized sovereignty and deployment controls. Starling Bank, described as the first bank to train an agentic model on U.K. sovereign cloud infrastructure, developed “Starling Intelligence,” its intelligence layer for AI capabilities. The bank said this layer was built on Civo’s U.K. cloud using Starling’s own transaction data. SBS also highlighted a “banking sovereign intelligence platform” designed to run generative and agentic models without customer or bank data leaving its own walls.

Larger banks and financial platforms focused on turning high volumes of structured and transactional data into reusable intelligence. Lloyds Banking Group said it built ATUM, a transaction foundation model that converts billions of banking transactions into reusable intelligence. The company tied ATUM’s acceleration to NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and said it is improving efficiency across bank-wide use cases.

NVIDIA’s blog also described additional deployments across the sector. bunq’s agentic assistant “Finn” runs on open weight models, with the company stating that using Nebius alongside NVIDIA accelerated computing delivered quality comparable to leading closed models at 60% of the cost per answer. NatWest said it is building technology to recognize a wide range of British accents, dialects and financial terminology, developed using NVIDIA speech models including Nemotron 3.5 ASR. Modulr described specialized models for distinct payment operations built on NVIDIA’s Nemotron-3-Ultra-550B open model. Santander said it is adopting NVIDIA’s new Kumo family of open weight foundation models for structured financial data to build specialized intelligence for core banking workloads.

Data and market analytics firms also featured in NVIDIA’s roundup. LSEG said it is working with NVIDIA to analyze large volumes of historical tick data up to 10x faster, using the Polars GPU engine, accelerated by NVIDIA cuDF on RTX PRO 6000 Blackwell GPUs. RavenPack described a dataset generation pipeline that takes a plain-language research hypothesis and returns a labeled dataset that is backtest-ready, running on a fine-tuned NVIDIA Nemotron model, with the goal of helping investors identify, analyze and visualize market trends. NVIDIA further noted that accelerated data processing using RTX PRO 6000 Blackwell Server Edition GPUs and cuDF can reduce the time from data to decision for scenarios such as fraud, across multi-terabyte-scale regulated sources.

The blog also offered examples of fraud and customer decision support. HSBC said it provides near-real-time client insights to support customer processes and on-demand decision-making while minimizing risk and fraud, powered by 5x faster data processing on Starburst Enterprise’s GPU Engine. Stripe said it observed 42x faster data pipelines and 92% cost savings for its Stripe Radar fraud-detection platform in offline testing, attributing the improvements to NVIDIA GPU-accelerated Ray Data, and describing continuous optimization of model training to improve fraud detection.

What the showcases do not fully disclose is the detailed performance trade-off between different model sizes, deployment environments, or governance configurations, beyond the headline metrics each company chose to emphasize. The blog also does not provide benchmark methodology, latency targets, or model accuracy figures that would allow apples-to-apples comparisons across institutions. Still, the repeated emphasis on open weight models and on-prem or sovereign cloud deployment reflects a broader industry push to balance performance with regulatory and operational control.

Looking ahead, investors and industry observers will likely watch whether open model deployments continue to scale across more banking functions, and how quickly organizations standardize tooling for data preparation, evaluation, and monitoring. NVIDIA’s demonstrations suggest a near-term focus on speeding up data pipelines and integrating foundation models into transaction, speech and market analytics workflows. The next bar, judging by the companies’ own framing, will be proving that these systems remain reliable under production volume, not just capable in controlled demonstrations.