At GTC Berlin, Revolut rebuilds on PRAGMA as European finance turns to open models

At GTC Berlin, NVIDIA gathered a set of production-scale AI deployments from European finance: banks, payment companies and financial data providers running open-weight models on their own infrastructure, building what the company calls proprietary intelligence from data they own.

The headline case is Revolut. The digital bank is rebuilding its technology stack on a single foundation model, PRAGMA, running on NVIDIA accelerated computing with the cuDF toolkit and Dynamo-Triton software on Blackwell Ultra GPUs. NVIDIA says PRAGMA is already serving an initial 80 million customers in production and is engineered to scale three times that.

Pavel Nesterov, Revolut's head of AI, is quoted saying the company processes billions of transactions and that the question was never whether the opportunity existed, but whether foundation models could operate at that scale across domains. That gets at the motive: a general-purpose chat model struggles to satisfy throughput, cost and data-boundary requirements at once.

Other examples point the same way. Starling Intelligence is described as the intelligence layer built by the first bank to train an agentic model on UK sovereign cloud infrastructure. Another institution's assistant, Finn, moved to open-weight models and now matches leading closed models at roughly 60 percent of the cost per answer. Banks are also using Nemotron 3.5 ASR for British accents and financial terminology, and building payment-operations models on Nemotron-3-Ultra-550B.

NVIDIA also said early customers have adopted Kumo, a new family of open-weight foundation models for structured financial data, to build intelligence for core banking workloads. That is the quieter but longer-lived signal: foundation models aimed at tables and transaction data are becoming a category of their own.

Why open weights? For European institutions the answer comes down to keeping data inside their own perimeter, controlling cost, and being able to explain the system to regulators. Holding the weights in your own data center, alongside sovereign cloud and private deployment, is what makes internal risk review and supervisory scrutiny tractable.

The trade-off is equally clear: you operate the model, you evaluate it, and you own the risk of every version change — work that cloud providers used to absorb. The real barrier to these deployments is therefore not GPUs but engineering and governance headcount.

What to watch is whether these projects move from first customers into core banking systems: whether PRAGMA really holds at three times today's scale, whether Kumo opens to more institutions, and whether European sovereign-cloud policy pushes localization further.