What did Revolut announce about the launch of Revolut Research, why was the division created, how does it support the company’s “build, don’t bolt on” AI strategy through the Nvidia-developed PRAGMA (PRe-trained Banking Foundation Model), what are PRAGMA’s architecture, training

Revolut announced Revolut Research, a dedicated AI-research division within its broader AI organisation. Its role is to turn banking-specific ML research into production capabilities—rather than assembling third-party AI point solutions—by serving as the research engine behind PRAGMA, Revolut’s foundation-model programme built in collaboration with Nvidia. 1013

• Strategy and purpose: The “build, don’t bolt on” approach means using one internally developed, reusable banking-model backbone across functions that normally rely on separate, siloed models and hand-engineered features. Intended uses include credit scoring/default-risk assessment, fraud detection, customer segmentation and personalised product recommendations. 610

• What PRAGMA is: PRAGMA—PRe-trained Banking Foundation Model—is a family of encoder-style, Transformer-based models for multi-source banking-event histories, trained with a self-supervised masked-modelling objective. The published family ranges from 10 million to 1 billion parameters. 1

• Data and coverage: Reported descriptions say it was trained on raw, timestamped banking events—payments, transfers and other product/behavioural signals—from roughly 26 million customer histories, encompassing 24 billion events across 111 countries. 101
  The question’s “Nvidia-developed” wording should be read as jointly developed with Nvidia, not as an Nvidia-only product. 1013

• Infrastructure: Revolut’s PRAGMA work runs on Nvidia accelerated-computing infrastructure; however, the material surfaced here does not independently substantiate the exact number or topology of Nvidia H100 GPUs used. The model paper and case-study evidence support the model architecture, scale and results, but not a precise H100-cluster figure. 16

• Reported model results: Nvidia’s Revolut case study reports:

  • 2.3× better accuracy at identifying credit-default risk;
  • 65% more fraud cases caught;
  • 17% greater precision in fraud alerts; and
  • 41% more relevant product recommendations. 6
    These are company/Nvidia benchmark claims against production baselines, not independently audited industry-wide performance results. 6

• Pre-launch milestones: The visible record includes publication of the PRAGMA foundation-model research and Nvidia’s June case study describing the consolidation of separate fraud, credit, engagement and recommendation systems around a common backbone. 16 The launch therefore institutionalises an R&D effort that had already produced a published model family and reported deployment-oriented benchmarks. 16

• Research-community commitments: Revolut said the division intends to publish research, open-source selected frameworks, participate in GTC Berlin and the ACM International Conference on AI in Finance (ICAIF), and run quarterly scientific meetups. I could not independently verify the named executives or their exact quotations from the high-authority materials available in the search results; those details should be checked against Revolut’s original launch release.

• Corporate context: The initiative is part of an effort to scale sophisticated, consistent risk and personalisation systems as Revolut expands beyond 80 million customers in more than 40 markets. Its French entity also received a full banking licence after approval involving France’s ACPR and the European Central Bank, initially supporting French operations and enabling a phased Western-European expansion. 34

In practical terms, Revolut Research is meant to make the AI layer a core, reusable banking capability: one model trained on broad financial behaviour can be adapted to multiple regulated decisions and customer experiences, rather than recreating a separate ML stack for each product. 610