Europe’s banks are training their own AI on open weights, with Nvidia underneath

At GTC Berlin, Nvidia published a long list of European financial firms building AI on open-weight models and their own infrastructure. Revolut says its PRAGMA foundation model is in production serving an initial 80 million customers. Lloyds built ATUM, a transaction foundation model. Starling Bank trained what it calls Starling Intelligence on U.K. sovereign cloud using its own transaction data. Santander is an early adopter of Nvidia’s new Kumo open models for structured financial data. NatWest is building speech recognition for British accents and financial terms with Nemotron 3.5 ASR, Modulr is developing payment-operations models on Nemotron-3-Ultra-550B, SBS launched a platform for banks to run generative and agentic models inside their own walls, and LSEG is working with Nvidia on analyzing historical tick data up to 10x faster.

The common thread is control: models trained on data banks own, running where regulators can see them, at costs banks set. It fits Nvidia’s wider European pitch around sovereign, open models across industries (Nvidia’s European open models).

Banks hold proprietary transaction data that general-purpose chatbots never see, and they operate under rules about where data can live. Open weights let them train on that data without sending it to a model vendor. The bunq example makes the cost case: Nvidia says its Finn assistant, now on open models deployed via Nebius, matches leading closed models on quality at 60% of the cost per answer. If that holds broadly, it puts pricing pressure on closed API providers selling into finance.

The messenger matters, though. The post was written by Nvidia’s regional financial-services sales lead, and every customer story is framed around Nvidia hardware and software. Figures such as Stripe’s reported 42x faster data pipelines and 92% cost savings for Radar fraud detection are presented without methodology, and Stripe’s numbers come from offline testing. They are directional until the institutions publish more.

The word open carries a tension here. These banks are reducing dependence on closed model vendors, but many of the models in the post, from Nemotron to Kumo, come from Nvidia and are tuned for its GPUs. Whether open weights translate into real portability across hardware is the test; independent chipmakers such as Tenstorrent are betting that serving open-weight models cheaply on other silicon is a big market (Tenstorrent’s raise).

The stakes for customers are real either way. Transaction foundation models can catch fraud faster, and HSBC and Stripe both point to fraud work. The same models can drive credit decisions, pricing and account flags at enormous scale, where an error becomes systematic rather than individual. Banks that build their own models also own the duty to validate and explain them to supervisors and customers. More in our Nvidia coverage, finance coverage and enterprise coverage.