European banks deploy open weight models on NVIDIA Blackwell infrastructure

European financial services organisations are deploying open weight artificial intelligence models on proprietary infrastructure, according to research published by NVIDIA. Institutions including banks, payment processors and financial data providers demonstrated production-scale intelligence at GTC Berlin. The shift relies on NVIDIA accelerated computing platforms to run these models on internal hardware or sovereign cloud services.

Operational Shift

The adoption of open weight models allows financial institutions to build proprietary intelligence using their own data, according to the NVIDIA blog post. This approach gives organisations control over costs and data sovereignty. Revolut rebuilt its technology stack on a single foundation model called PRAGMA to address the challenge of operating foundation models at the scale of billions of transactions.

Implementation Details

Operational scaling is achieved through specific software toolkits and hardware configurations. Revolut uses the NVIDIA cuDF toolkit and Dynamo-Triton software on Blackwell Ultra GPUs to scale PRAGMA, which services an initial 80 million customers. Lloyds Banking Group accelerates its ATUM transaction foundation model using RTX PRO 6000 Blackwell GPUs. Starling Bank trained its agentic model on UK sovereign cloud infrastructure using its own transaction data.

Technology Stack

The implementations involve specific NVIDIA hardware and software versions. Revolut uses Blackwell Ultra GPUs. Lloyds and LSEG utilise RTX PRO 6000 Blackwell GPUs. HSBC employs Starburst Enterprise’s GPU Engine for data processing. Stripe uses NVIDIA GPU-accelerated Ray Data. NatWest developed speech recognition technology with NVIDIA Nemotron 3.5 ASR. Modulr builds on the Nemotron-3-Ultra-550B open model.

Performance Outcomes

• Revolut engineered PRAGMA to scale 3x beyond its initial 80 million customers.
• bunq matches the quality of leading closed models at 60% of the cost per answer.
• LSEG analyzes historical tick data up to 10x faster with the Polars GPU engine.
• Stripe sees 42x faster data pipelines and 92% cost savings for fraud detection.

What it means for organisations

Engineering and risk teams should evaluate whether open weight models align with data governance policies. Organisations need to determine if in-house GPU infrastructure can support the computational demands of foundation models. Leadership should review supplier contracts to understand support levels for proprietary AI stacks. Compliance officers must verify that training data remains within legal jurisdictions. This shift requires coordination between data engineering, security and infrastructure teams to ensure models operate reliably at scale without exposing sensitive financial records to external vendors.

What to do and how to stay safe: NVIDIA

• Audit current AI model deployments to identify reliance on external inference endpoints.
• Review infrastructure capacity to handle GPU-accelerated workloads for large language models.
• Establish clear metrics for model performance and cost before full production rollout.
• Verify that data processing pipelines comply with relevant financial regulations.

General security guidance from the TechSec Observer newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Which banks are using NVIDIA technology for AI models?

Revolut, Starling Bank, Lloyds Banking Group, NatWest, Santander and HSBC are using NVIDIA technology.

What hardware is Revolut using for its PRAGMA model?

Revolut uses NVIDIA Blackwell Ultra GPUs and the cuDF toolkit.

How much faster are Stripe's data pipelines with GPU acceleration?

Stripe sees 42x faster data pipelines in offline testing.