Key points
- Revolut scales PRAGMA model for 80 million customers using NVIDIA Blackwell Ultra GPUs.
- Starling Bank trains agentic models on UK sovereign cloud infrastructure with proprietary data.
- Stripe achieves 42x faster data pipelines for fraud detection using GPU-accelerated tools.
The short version
- European financial institutions are shifting production AI workloads to open weight models.
- These deployments run on NVIDIA accelerated computing hardware within private infrastructure.
- Organizations cite cost control and data sovereignty as primary drivers for this change.
The details
European banks, payment processors and data providers demonstrated production-scale intelligence at GTC Berlin, according to NVIDIA blog posts. These organizations are building proprietary systems using open weight models that run on their own infrastructure. This approach allows them to keep financial data within their control while leveraging NVIDIA accelerated computing. The shift moves away from relying solely on external closed-model providers for core banking operations.
Revolut is rebuilding its technology stack around a single foundation model named PRAGMA. The system is powered by NVIDIA accelerated computing and uses the cuDF toolkit and Dynamo-Triton software. It runs on NVIDIA Blackwell Ultra GPUs and currently services 80 million customers. The architecture is engineered to scale three times its current capacity, according to the vendor’s research.
Starling Bank developed Starling Intelligence on Civo’s UK cloud using its own transaction data. It is the first bank to train an agentic model on UK sovereign cloud infrastructure. Lloyds Banking Group built ATUM, a transaction foundation model that processes billions of banking transactions. This system uses NVIDIA RTX PRO 6000 Blackwell GPUs to improve efficiency across bank-wide use cases.
SBS launched a platform allowing banks to run generative and agentic models without data leaving their walls. Neobank bunq runs its agentic assistant Finn on open weight models via Nebius. This deployment matches the quality of leading closed models at 60% of the cost per answer, according to NVIDIA. NatWest is building technology to recognize British accents and financial terminology using NVIDIA speech models.
Modulr is building specialized models for payment operations based on the NVIDIA Nemotron-3-Ultra-550B open model. Santander is using the NVIDIA Kumo family of open weight foundation models for structured financial data. These tools support core banking workloads with specialized intelligence.
The context
As TechSec Headlines reported on 2026-10-22, IBM CEO Arvind Krishna stated that open models offer 90% of top-tier performance at 10% of the cost. This economic argument is driving demand for local infrastructure among enterprises. The current deployments by European financial firms validate that claim in production environments.
LSEG is working with NVIDIA to analyze historical tick data up to 10x faster using the Polars GPU engine. This acceleration uses NVIDIA cuDF on RTX PRO 6000 Blackwell GPUs. RavenPack built a dataset generation pipeline on Bigdata.com using a fine-tuned NVIDIA Nemotron model. Investors use these datasets to identify and visualize market trends.
HSBC provides near-real-time client insights using Starburst Enterprise’s GPU Engine. This setup enables 5x faster data processing while minimizing risk and fraud. Stripe sees 42x faster data pipelines and 92% cost savings for its Radar fraud-detection platform. These results come from offline testing with NVIDIA GPU-accelerated Ray Data.
What to watch
- Expansion of open weight model usage beyond initial pilot programs.
- Performance benchmarks comparing open models to closed alternatives in banking.
- Regulatory responses to sovereign cloud AI deployments in Europe.
- Supply chain dependencies for NVIDIA Blackwell hardware in financial sectors.
What it means for organisations
Operations teams should assess whether current AI workloads can migrate to local infrastructure. Check if vendor lock-in risks are reduced by using open weight models. Ensure data governance policies align with on-premises processing requirements. Coordinate with procurement to evaluate hardware costs against cloud service expenses. Inform legal and compliance teams about data residency implications. Verify that internal security controls can manage model access and updates. Review supplier contracts for flexibility in adopting new AI architectures.
Background: Data residency
Data residency laws require organizations to store specific types of information within defined geographic boundaries. You must identify where your cloud providers place physical servers. Failure to comply can result in legal penalties and loss of customer trust. Map your data flows explicitly to meet these jurisdictional requirements.
Read the full guide: Data Residency Laws: How to Keep Data in the Right Jurisdiction
What to do and how to stay safe: NVIDIA
- Audit current AI dependencies to identify reliance on external closed models.
- Evaluate internal infrastructure capacity to support GPU-accelerated workloads.
- Review data residency requirements for AI training and inference processes.
- Monitor vendor announcements for updates on open weight model compatibility.
Step-by-step guide: Secure Home Routers for Small Business Networks
General security guidance from the TechSec Headlines newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which banks are using open weight models on NVIDIA hardware?
Revolut, Starling Bank, Lloyds Banking Group, bunq, NatWest, Modulr and Santander are deploying these systems.
What hardware is powering these AI deployments?
Most deployments use NVIDIA Blackwell Ultra GPUs or RTX PRO 6000 Blackwell GPUs.
Why are financial institutions switching to open weight models?
They offer cost savings, data sovereignty and proprietary intelligence built on owned data.
