European banks replace closed models with open-weight models to process billions of transactions inside their walls

Concurrent announcements at the GTC conference in Berlin reveal a structural shift in the global financial sector: major banks and payment firms are no longer satisfied with renting AI through closed APIs, but have moved to building sovereign AI inside their digital walls, relying on open-weight models and accelerated computing that ensure full control over data and costs.

The shift is led by Revolut Bank, which rebuilt its entire technology stack around a unified foundational model called PRAGMA. The model runs on Nvidia Blackwell Ultra processors, supported by the cuDF toolkit and Dynamo-Triton software, and currently serves eighty million customers with engineering scalability to three times that size. Bavel Nesterov, head of the bank’s artificial-intelligence division, explains that processing billions of transactions required proving the foundational models could operate at that massive scale, confirming that the partnership with Nvidia closed the gap between research innovations and live banking systems in a real-production environment.

Operational sovereignty has shifted from a regulatory requirement to a mandatory technical architecture for banks.In the United Kingdom, Starling Bank became the first bank to train a fully integrated model and agent on a locally sovereign cloud architecture via the British Sifvo cloud, developing a Starling Intelligence layer that relies exclusively on its own operational data. In the same vein, Lloyds Banking Group built the ATUM model to convert billions of banking transactions into operationally reusable knowledge, accelerated by RTX PRO 6000 Blackwell GPUs, while SPS launched a dedicated sovereign-AI platform that enables banks to run generative and agent models without any data leaving their regulatory perimeter.

The economic equation is the most sensitive driver behind this trend. A bunq digital-bank experiment running its agent assistant Fin showed that relying on open-weight models hosted on the Nebius cloud matches the performance quality of major closed commercial models, but at only sixty percent of the cost per response. Customisation extends across all daily-operation dimensions: NatWest Bank is developing a system to process British dialects and complex financial terminology using the Nemotron 3.5 ASR speech models; payments firm Modular adopts models for transfer operations based on the Nemotron-3-Ultra-550B model; and Santander selected the open-weight COMO model family to handle structured financial data within its core banking systems.

The speed of processing complex data streams directly impacts financial defence lines and investment research. In offline tests, the Stripe platform demonstrated that accelerating fraud-detection data on the Stripe Radar platform with Ray Data libraries and GPUs yields a forty-two-fold increase in data-processing throughput while cutting operational costs by ninety-two percent. In online settings, HSBC raised data-processing speed fivefold with the Starburst Enterprise engine to reduce fraud risk and deliver real-time insights to customers, while the London Stock Exchange accelerated historical tick-data analysis by roughly ten times, alongside Revn Back’s development on the Bigdata.com platform of automated pipelines that turn natural-language research hypotheses into ready-to-test datasets.

This shift reshapes spending priorities and skill requirements in financial institutions across the region.In Gulf markets, where central banks impose strict data-localisation and confidentiality regimes, the success of open European architectures gives technology leaders practical proof that the latest AI capabilities can run locally without breaching compliance frameworks or incurring escalating fees from closed-cloud model providers. In Egypt and the Levant, where startups and development teams aim to lower inference and compute costs, the open-compute model shows that building smart payment and financial-service platforms does not require massive budgets for foreign API calls, but rather a reallocation of engineering talent toward locally tuning models and accelerating structured-data processing on servers owned and operated by the developer.

Operational indicators from banks reveal that owning an open model and customizing it on internal infrastructure has become the cornerstone of sustainable financial competition, where sovereign-compliance requirements intersect with field-cost viability, free from cloud-dependency constraints.