Revolut launches an in-house AI research unit built on PRAGMA, its own finance model trained with Nvidia

Revolut has created Revolut Research, a dedicated AI research division that will own the bank's proprietary AI and machine learning work, anchored on PRAGMA — a finance-specific foundation model built in partnership with Nvidia and trained on data from its 80 million customers.

The unit is a capital allocation decision rather than a transaction. No external investment, joint venture or equity stake has been disclosed, and the terms of the Nvidia partnership were not made public. Revolut is funding the build from its own balance sheet, keeping the resulting model assets wholly owned rather than licensing capability from third parties.

That is the point Revolut is making. The company argues most financial services firms deploy off-the-shelf AI tools and maintain separate models for separate problems — one for fraud, another for customer service — while PRAGMA is designed as a single foundation trained across the entire customer journey.

From that base, Revolut says the model supports fraud detection, credit decisions, product recommendations and customer service from shared learning rather than siloed systems.

The bank has published early testing figures. It reports 2.3 times higher accuracy in identifying credit default risk, 41 per cent more product recommendations, and an increase in fraud cases caught. These are internal test results and have not been independently verified.

The strategic argument rests on data scale. Revolut points to 80 million customers processing billions of cross-border transactions in real time, and says that as the dataset grows the models compound — producing what it describes as a proprietary intelligence advantage that conventionally structured banks struggle to match.

Revolut Research will also work with academic and technology institutions developing AI, which suggests a recruitment channel as much as a research one. Competition for senior model researchers in London remains the binding constraint on any bank attempting this in-house.

Pavel Nesterov, head of AI at Revolut, said the unit institutionalises a "build, don't bolt on" philosophy, and that training native foundation models on the company's global operational data gives engineering teams a faster route to deploying new features.

For the cap table, the read-across is about where the value sits. Owned models trained on proprietary transaction data are an intangible asset that accrues to existing shareholders, unlike vendor contracts, which convert the same spending into an operating cost.

Revolut has spent the year building out its regulated footprint, including securing a full French banking licence this month as it establishes a Western European hub in Paris. The AI unit extends the same pattern: control the infrastructure, own the licence, keep the stack in-house.

Original source: Tech.eu — https://tech.eu/2026/08/25/revolut-launches-ai-unit-as-it-proclaims-virtues-of-proprietary-ai/