European Banks Turn to NVIDIA Open Models for Proprietary AI

At GTC Berlin, NVIDIA described how European financial institutions are using open weight models and accelerated computing for proprietary AI systems. Its account spans production transaction models, bank-owned data, sovereign cloud deployments and projects under development. For finance leaders, the common thread is control over where workloads run and how institution-specific data is made useful, with benefits attributed to the vendor and its named customers. NVIDIA Newsroom report.

Revolut rebuilds stack on PRAGMA

NVIDIA says Revolut is rebuilding its technology stack on one foundation model, PRAGMA, powered by NVIDIA accelerated computing. The vendor places PRAGMA in production for an initial customer base and says it is engineered to scale, using cuDF, Dynamo-Triton and Blackwell Ultra GPUs.

NVIDIA cites Revolut AI head Pavel Nesterov on transaction volume and cross-domain operation. It says PRAGMA was built in collaboration with NVIDIA and presents it as evidence that research models can reach production banking systems. That is a vendor characterization, not an independent benchmark.

Related coverage: related Fintech reporting.

Sovereign cloud and bank models

NVIDIA says Starling Bank developed its Starling Intelligence layer on Civo's U.K. cloud using its own transaction data, calling it the first bank to train an agentic model on U.K. sovereign cloud infrastructure. Lloyds built ATUM, a transaction foundation model, accelerated with NVIDIA RTX PRO 6000 Blackwell GPUs. SBS launched a banking sovereign intelligence platform.

NVIDIA presents these as institution-specific efforts. It says the SBS platform lets banks run generative and agentic models without data leaving their walls. Our assessment is that buyers should evaluate each named deployment separately, checking which data, model and hosting choices apply to the particular workload they intend to develop.

Related coverage: related Fintech reporting.

Open weight assistants

NVIDIA says open weight models give institutions proprietary intelligence on data they own at costs they control. It cites bunq's assistant Finn, now on open weight models and deploying on Nebius, and says it matches closed models on quality. That comparison is NVIDIA's account, not independent evaluation.

NVIDIA describes NatWest building technology for British accents, dialects and financial terminology with speech models including Nemotron 3.5 ASR. It says Modulr is building payment-operations models on the Nemotron-3-Ultra-550B open model, and Santander is an early adopter of the Kumo family for structured financial data. These are builds or early adoption.

Related coverage: related Fintech reporting.

Research and fraud pipelines

NVIDIA says LSEG is working with NVIDIA to help institutions analyze historical tick data faster using the Polars GPU engine, accelerated by cuDF on NVIDIA RTX PRO 6000 Blackwell GPUs. The source frames this as faster analysis, not an independent benchmark. RavenPack built a dataset pipeline on Bigdata.com using a fine-tuned NVIDIA Nemotron model.

NVIDIA says RavenPack's pipeline returns backtest-ready datasets for investors. It also says HSBC uses Starburst Enterprise's GPU engine for client insights and on-demand decisions while minimizing risk and fraud, and that Stripe tested GPU-accelerated Ray Data for Stripe Radar in offline testing. The account describes vendor-built or tested workflows, not independent validation or universal adoption.

Related coverage: related Fintech reporting.

What This Means for Practitioners

Buyers should ask whether a named model is in production, in pilot or still being built, and which institution-specific data governs it. They should ask whether the model is generally available, whether it runs on the buyer's infrastructure, and whether NVIDIA technology covers every function or selected workloads. They should require software product information, safety and governance evidence.

Researchers should ask whether any efficiency or quality claim has been independently measured, and under what data, scale and risk controls. They should ask whether fraud, research and assistant workflows were tested outside vendor evaluations, and whether cost control holds when models are retrained or expanded. These are diligence conditions, not forecasts.