The shift at GTC Berlin
At GTC Berlin, multiple European banks, payment companies, and financial data providers described live deployments of large-scale intelligence built on open-weight models. The systems run on NVIDIA accelerated computing inside the organizations’ own data centers. Revolut is rebuilding its core technology stack around a single foundation model named PRAGMA that is powered by the same NVIDIA hardware.
Until recently, most financial institutions relied on closed models hosted by external vendors. That approach limited control over data residency and model behavior. The shift described at the event moves training and inference onto open-weight models that organizations can fine-tune and operate on premises. NVIDIA’s role is supplying the accelerated computing layer that makes those workloads practical at production scale.
Revolut’s PRAGMA deployment
Revolut’s PRAGMA model is the clearest example given. The company is consolidating multiple prior systems onto this single foundation model. NVIDIA cuDF is cited as part of the data-processing pipeline that supports the effort. Other unnamed banks and payment processors reported similar patterns: open models chosen for their transparency, combined with NVIDIA GPUs or DGX-class systems kept inside the firm’s security perimeter. The common thread is the desire to create proprietary intelligence that remains under direct organizational control.
These deployments are not experimental pilots. The organizations presenting at the event described production workloads already handling real customer data and internal processes. The choice of open-weight models allows each firm to inspect and modify the underlying weights rather than treating the model as a black box supplied by a third party. Running the models on owned NVIDIA hardware further removes any requirement to send data outside the organization’s network boundaries.
Technical and operational patterns
The pattern that emerged across the presentations centers on two technical decisions made together. First, the institutions selected open-weight models so they could retain full visibility into model behavior and apply domain-specific fine-tuning. Second, they placed the entire inference and training stack on NVIDIA-accelerated systems located in their own data centers. NVIDIA cuDF appears in at least one described pipeline as the component responsible for preparing large datasets before they reach the model.
This combination addresses a set of constraints that are especially acute in European financial services. Data-residency regulations require that certain categories of customer information never leave designated jurisdictions. Closed models accessed through public APIs make compliance verification difficult because the hosting location and data-handling practices sit outside the bank’s direct control. By contrast, an open-weight model running on local NVIDIA hardware lets the institution document exactly where every training token and every inference request is processed.
Reactions and remaining questions
The accounts shared at GTC Berlin came primarily from the organizations themselves rather than from independent auditors. No detailed performance benchmarks or cost comparisons were released during the sessions. Several presenters acknowledged that moving from vendor-hosted models to self-managed infrastructure requires new operational skills inside their engineering teams. They did not, however, present evidence that the open-model route is cheaper or faster than the previous approach; the stated advantage was control rather than speed or cost.
Why it matters
Financial services firms handle regulated data and face strict residency rules. Running open models on owned infrastructure gives them a concrete way to meet those constraints while still using modern foundation-model capabilities. The approach also reduces dependence on any single external provider. For teams that must audit or modify model behavior, open weights plus local hardware supply a practical path that closed APIs cannot match. The pattern shown at GTC Berlin indicates that the necessary software and silicon stack is now mature enough for day-to-day use in production financial systems.
Over time, this model of deployment could change how regulators evaluate AI risk in banking. When a model’s weights and its execution environment are both under the institution’s direct control, examiners can request full access to training data, fine-tuning steps, and inference logs without negotiating with an external vendor. That level of transparency is difficult to achieve when the model lives behind a third-party API. Whether other regulated industries follow the same route will depend on how quickly the operational overhead of self-hosted open models decreases and how clearly regulators signal that on-premises deployments satisfy their oversight requirements.
The European financial sector’s early moves therefore serve as a test case. If the institutions that spoke at GTC Berlin can demonstrate sustained reliability and regulatory compliance, the same combination of open weights and owned NVIDIA hardware may become a standard template for any organization that must keep both data and model logic inside its own perimeter.
---
Sources:
{
"publisher": "NVIDIA Blog",
"title": "Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence",
"url": "https://blogs.nvidia.com/blog/european-financial-services-organizations-open-models-proprietary-intelligence/",
"published_at": "2026-10-22T05:00:56.000Z",
"summary": "At GTC Berlin, banks, payment companies and financial data providers are showing what production-scale intelligence looks like — built on NVIDIA accelerated computing, powered by open weight models and running on their own infrastructure. Revolut is rebuilding its technology stack on a single foundation model PRAGMA, powered by NVIDIA accelerated computing. Using the NVIDIA cuDF […]"
}
No comments yet