UK digital bank Starling is building a transactional foundation AI model with support from Nvidia. The model is being trained and hosted on the cloud of UK provider Civo and will draw on more than 15 billion anonymised transactions. It is meant to become a single intelligence layer for all of the bank's AI services. Starling links the choice to the need to reduce the financial sector's reliance on a handful of global cloud providers.

What it means: Starling's move shows how a bank can turn its transaction history into a competitive edge. A model that has learned from billions of real payments can spot fraud, anticipate liquidity needs and give more personal guidance than a general language model. The infrastructure choice is just as interesting. Training happens on a local cloud provider rather than one of the three big US platforms, precisely to reduce concentration risk. European supervisors share that concern through DORA, which asks banks to manage their dependence on critical third-party providers. For Greek banks, the example points to two paths. One is investing in their own models built on their own data, which takes a sizeable team and budget. The other is working with European infrastructure providers to keep data inside the EU. Either way, anonymisation and the legal basis for processing under GDPR must be documented from day one. For Greek fintechs, well-organised transaction data is only becoming more valuable. Data teams should start by cataloguing what transaction data they hold, how clean it is and under what consent it was collected, because that inventory decides what any future model can legally learn from.

Read the original on Starling Bank

Related service: myDATA integrations

Related stories