Revolut PRAGMA Powers Bank's Unified AI Stack

One model to run the bank: that is the wager behind the Revolut PRAGMA launch. Revolut is betting its future on one model. The London-based fintech revealed at NVIDIA's GTC Berlin conference that it is consolidating its technology stack onto PRAGMA, a single Revolut PRAGMA foundation model that will serve the company's customer base and handle the billions of transactions flowing through its platform each year.

The announcement marks one of the most aggressive AI infrastructure commitments in banking: instead of stitching together dozens of specialized models for fraud detection, support, and personalization, Revolut is converging on a unified intelligence layer built with NVIDIA's cuDF toolkit and Dynamo-Triton software, running on NVIDIA Blackwell Ultra GPUs. The Revolut PRAGMA foundation model is engineered to scale threefold from its initial deployment, according to reporting on the reveal.

Revolut's head of AI, Pavel Nesterov, framed the move as proof that foundation models can now operate at banking scale across domains, arguing that the collaboration with NVIDIA closes the gap between research breakthroughs and production banking systems. That is a striking claim from a company processing transactions for tens of millions of customers — and a direct challenge to the conventional wisdom that a bank must license closed models from Silicon Valley labs.

Open Weights, Own Infrastructure

The Revolut PRAGMA foundation model rollout is part of a broader pattern on display at GTC Berlin: European financial institutions are building proprietary intelligence on open weight models running on their own infrastructure, with NVIDIA accelerated computing underneath. Starling Bank disclosed that it trained an agentic model, "Starling Intelligence," on U.K. sovereign cloud infrastructure using its own transaction data — reportedly the first bank to do so. The open-weight momentum also reaches agentic AI: NVIDIA-backed sovereign agent platforms are proliferating across Europe (see our coverage of deepset's sovereign agents with NVIDIA). Lloyds Banking Group built ATUM, a transaction foundation model converting billions of banking transactions into reusable intelligence, accelerated with NVIDIA RTX PRO 6000 Blackwell GPUs. SBS launched a sovereign banking intelligence platform so banks can run generative and agentic models without data leaving their own walls.

For European banks, the motivation is not purely technical. GDPR and data-sovereignty rules make piping customer financial data through third-party AI services legally risky, and the strategic calculation is simple: if AI becomes the core of banking infrastructure, banks want to own it outright. Retailers are making a parallel calculation about agentic commerce (Walmart opened checkout to AI shopping agents earlier this week). Open weight models give institutions proprietary intelligence built on data they already own, at costs they control.

Why One Model Changes the Economics

The technical case for the Revolut PRAGMA foundation model is consolidation. Maintaining a fleet of specialist models means duplicated training pipelines, separate evaluation regimes, and fragmented monitoring. A single foundation model lets Revolut route fraud detection, customer service, credit decisions, and product recommendations through one stack — with shared tooling and one audit trail. The company's subscription revenue, the line most tied to AI-driven personalization, reached roughly $966 million in 2025, up 67 percent year-on-year, suggesting the personalization payoff is already real.

There are open questions. Regulators increasingly demand legible decision trails for credit and fraud judgments, and it remains to be seen how a large foundation model satisfies those explainability requirements — a question the underlying research has not yet fully answered. Running proprietary infrastructure at this scale also concentrates operational risk: a single-model stack is only as resilient as its deployment.

Still, the signal from GTC Berlin is unmistakable. Europe's financial sector is no longer renting intelligence from closed providers; it is manufacturing it. If the Revolut PRAGMA foundation model delivers on its scale promises, expect every major bank in the region to ask whether their stack should converge too.

The Revolut PRAGMA story also sets up a longer research arc: Revolut Research is expected to present further findings at ICAIF in November 2026, and the division has signaled plans to publish research and open-source technical frameworks on an ongoing basis. Those publications will let the industry judge whether a unified banking model is a genuine efficiency breakthrough or an elegant demo that works best on its home turf.

Sources: NVIDIA Blog — European Financial Institutions Embrace Open Models; NVIDIA GTC Berlin 2026 live updates; Tech Times — Revolut Builds Its Own Banking AI; The Tech Buzz — Europe's Banks Ditch Closed AI.