Nvidia publishes European banks’ open-model AI deployments

Nvidia published a list of European banks, payment firms and data providers that it says run open weight models on its accelerated computing, led by Revolut’s PRAGMA model for an initial 80 million customers. For institutional readers, the document shows where bank AI spending is going, and that most of the named deployments depend on Nvidia hardware or software even when the models are open.

• Revolut is scaling its PRAGMA foundation model on Blackwell Ultra GPUs, serving an initial 80 million customers.
• Nvidia says bunq’s Finn assistant, on open weight models, matches closed-model quality at 60% of cost per answer.
• The release gives no baselines for the cost and speed claims and no dates for further rollouts.

• 80M Revolut customers served by PRAGMA initially, engineered to scale 3x
• 60% bunq Finn cost per answer versus leading closed models, at matched quality
• 42x Faster Stripe Radar data pipelines in offline testing, baseline unstated
• 92% Stripe Radar cost savings in offline testing, same test as 42x speedup

In a release, Nvidia said banks, payment companies and financial data providers are running production AI “built on NVIDIA accelerated computing, powered by open weight models and running on their own infrastructure.” The release is tied to GTC Berlin and is a vendor document, not a filing or a joint disclosure.

Only one customer is quoted. The rest of the claims are Nvidia’s own summaries of its partners’ work.

Revolut, Starling and Lloyds build on their own transaction data

Revolut is rebuilding its technology stack on a single foundation model, PRAGMA, using the cuDF toolkit and Dynamo-Triton software on Blackwell Ultra GPUs. Nvidia says it is serving an initial 80 million customers and is “engineered to scale 3x.” Applied to the 80 million figure, that implies capacity for about 240 million, though the release does not state a target or a timeline.

Pavel Nesterov, head of AI at Revolut, is the only named executive in the release.

“Built in collaboration with NVIDIA, it’s the clearest proof yet that the gap between a research breakthrough and a production banking system can be closed.”

Pavel Nesterov, head of AI at Revolut, quoted in the Nvidia release

Starling Bank, described by Nvidia as the first bank to train an agentic model on U.K. sovereign cloud infrastructure, built “Starling Intelligence” on Civo’s U.K. cloud using its own transaction data. Lloyds Banking Group built ATUM, a transaction foundation model, accelerated with RTX PRO 6000 Blackwell GPUs. SBS launched a banking sovereign intelligence platform so banks can run generative and agentic models without data leaving their own walls.

Open models are cheaper, but the speed claims lack baselines

The release’s central argument is that “Open weight models give European financial institutions proprietary intelligence built on data they own, at costs they control.” The one hard cost figure comes from bunq, whose Finn assistant now runs on open weight models deployed on Nebius. Nvidia says that matches “leading closed models” at 60% of the cost per answer, which is a 40% saving if the comparison holds.

Other named uses cover models, speech and data. Modulr is building specialized payment models on the Nemotron-3-Ultra-550B open model. NatWest is working on speech technology for British accents and financial terminology with Nvidia’s speech models, including Nemotron 3.5 ASR. Santander is an early adopter of the new Kumo family of open weight models for structured financial data.

On data processing, LSEG is working with Nvidia to analyze historical tick data “up to 10x faster” with the Polars GPU engine and cuDF. HSBC cites 5x faster data processing on Starburst Enterprise’s GPU Engine. Stripe reports 42x faster pipelines and 92% cost savings for Stripe Radar with GPU-accelerated Ray Data, in offline testing only.

RavenPack built a pipeline on Bigdata.com, running on a fine-tuned Nemotron model, that turns a plain-language research hypothesis into a labeled, backtest-ready dataset.

What the document leaves open

The release does not name the closed models bunq compared against, define “cost per answer,” or say how quality was measured. It does not give the baseline hardware for the LSEG, HSBC or Stripe speedups, and the Stripe numbers come from offline testing, not live traffic. It also does not say which accelerators Starling or SBS use, give any contract values, or state Kumo’s size, licensing or availability.

It does not say whether PRAGMA’s 80 million customers are all live today or how many transactions it scores. Nvidia’s “initial” qualifier leaves the current footprint undefined.

What changes for banks buying AI

For bank technology buyers, the pitch is about control: models trained on data the bank owns, run on infrastructure it picks, with unit costs it can measure. The deployments span a sovereign U.K. cloud (Civo), a neocloud (Nebius) and in-house setups. That suggests the model layer is portable across venues.

The acceleration layer is less portable. The named tools include cuDF, Dynamo-Triton, Ray Data and the Polars GPU engine, plus Blackwell Ultra and RTX PRO 6000 GPUs. A bank that swaps an open model for another still depends on Nvidia’s software and silicon for the speedups the release advertises. Open weights lower the cost of leaving a model vendor but not necessarily a GPU vendor.

The BlockWest read. We think the competitive point is not open versus closed models but who owns the serving stack. Banks gain leverage over model vendors, and Nvidia converts that leverage into demand for its libraries and GPUs. The 60% cost figure will pressure closed-model vendors on price, and sovereignty claims will push buyers toward regional clouds. The winner is whoever controls inference economics, and the release shows Nvidia positioned there.

No dates are given. The open question is for Revolut and Nvidia: when will PRAGMA’s “initial 80 million customers” grow toward the 3x capacity, and will bunq or Stripe publish the baselines behind the 60%, 42x and 92% figures?

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.