Fintech, AI, and open source | an NVIDIA interview

Kevin Levitt, Director of Global Industry Business Development for Financial Services at NVIDIA, shares his thoughts on AI investments and the role of open-source tech in deploying AI-enabled apps.

Multiple recent fintech investments centred on AI. From what you've seen, in which AI applications was the most capital injected, and where is the ROI already materialised vs. experimental?

One of the key findings of our 2026 ‘State of AI in Financial Services’ survey is that 89% of respondents stated AI is reducing costs and generating new revenues for their firm. When financial institutions talk about ROI, it's in terms of these two aspects, and nearly 9 in 10 are seeing both benefits from their AI investments.

 When we further dove into the top use cases that are really driving ROI for the fintech community, it was actually in three areas:

• customer experience and engagement,
• document processing and helping with middle- and back-office functions,
• and risk management.

Across both fintechs and financial services firms, the common theme is extracting intelligence from data. There is arguably no industry whose success is more predicated on being able to do exactly that. Every day, banks are analysing mountains of data and extracting intelligence to better acquire, underwrite, and service customers, and to better predict how they will engage with products, and which new offerings they’re most likely to value. That's why we see so much investment going into what we call AI factories: horizontal platforms that let developers and data scientists build and deploy AI-based solutions across every function of the organisation.

On the more experimental side, we are seeing strong momentum around transaction foundation models, which are emerging as a key AI use case. These models leverage transformer-based architectures, which we’ve used for large language models (LLMs) that predict the next word, sentence, or paragraph. Here, we’re using transaction foundation models to predict the next transaction. We take structured data, such as tabular transaction data, tokenize it, feed it into a transformer model, and the transaction foundation model generates what the next transaction should look like. When that actual transaction comes through from the consumer or the merchant, we can analyse whether it looks as expected, falls within a normal distribution, or appears anomalous and potentially fraudulent. What’s particularly powerful is that a pre-trained transaction foundation model functions as a reusable intelligence layer within the institution — trained on the bank’s full event history and then efficiently fine-tuned for each downstream task, from fraud detection and credit scoring to personalisation and lifetime value prediction.

Stripe shared their success in leveraging transaction foundation models at Stripe sessions in May of last year, highlighting a dramatic improvement in their ability to fight card-testing fraud. Revolut has gone further still: working with us, they developed PRAGMA, a family of transaction foundation models pre-trained on a huge dataset – 26 million users, 24 billion events across 111 countries. What’s interesting is that they use a single core model that can scale from 10M to 1B parameters. And instead of building separate models for each task, this one model can manage everything simultaneously: credit scoring, fraud detection, lifetime value prediction, communication engagement, and product recommendation. There’s no need for hand-crafted features either. The models can be fine-tuned quickly for new use cases. In practice, it’s like having one central intelligence layer – trained on the bank’s own data – that powers the entire organisation.

Lightweight fine-tuning adapts the shared model to each new use case in a fraction of the time. This is the intelligence layer model in action: one sovereign model, trained on the bank’s own data, powering the entire organisation.

How can fintech players identify which AI architectures and technical approaches best fit their needs?

Fundamentally, there are a few certainties when it comes to AI in financial services. One is that AI requires accelerated computing. Traditional CPU-based environments simply can’t meet the demands of modern model training and inference. The second, as we already discussed, is that AI delivers a positive ROI. That's why nearly all respondents plan to maintain or increase their AI budgets in 2026, compared to last year. A large part of the reason is that fintechs and other financial enterprises today are building and deploying models that are yielding more accurate outcomes, which in turn drives the need for inference at scale. These AI-enabled applications produce better results and, naturally, more engagement – especially with new agentic AI and reasoning-based models that think more like we do, and require far more compute than traditional, one-shot inference models.

All of these combined requirements (inference at scale, agentic reasoning, and dramatically higher compute needs) are driving the move to accelerating computing and NVIDIA AI factories to support not just model development and training but also deployment into production. The last thing we want is for our customers or internal stakeholders to wait inordinate amounts of time for results; we want everyone to be as productive as possible and, certainly, to get the most accurate answers possible. That is where our accelerated computing platform and AI factories really play an important role.

The other important aspect is lowering the total cost of ownership. That's why we’ll keep seeing banks invest in building proprietary models using open-source software and models: they're going to build these proprietary LLMs, both generative and agentic AI models, using open-source models in combination with their proprietary data to yield more accurate results and lower their total cost of ownership (TCO).  And as we discussed, no industry is more dependent on extracting intelligence from data than financial services and fintech.

As you mentioned, we’re seeing more players building proprietary AI solutions. Which factors are most important when deciding whether to build vs. buy/outsource?

The key factors we see driving these decisions today are data sovereignty, data privacy, competition, and compliance. Last but not least are ROI and lowering the TCO. All of these considerations tie back to banks beginning to invest at scale in building their proprietary LLMs, generative AI models, and agentic AI models.

When ChatGPT first came on the scene in November 2022, there was a lot of pressure just to get to market with some GPT-based capability. Many financial institutions learned from that experience that accuracy and TCO are key to building AI-enabled applications that drive ROI.

The best way to improve accuracy is to use proprietary data combined with open source models to develop generative and agentic AI models. The way to lower TCO is to own the model. That’s why 84% of survey respondents in our report said open source software is moderately to extremely important to their organisation's AI strategy. The pendulum has swung. Financial services companies are looking at how to build proprietary AI solutions to drive more accurate outcomes and improve ROI, while investing in open source capabilities on top of our AI factories to generate more accurate, function-specific models. You might have a GPT-style model for wealth management, another fine-tuned for customer onboarding, one focused on cross-sell and customer engagement, and others dedicated to risk management and fraud detection – each tailored to specific data and AI requirements. By leveraging open source models, banks can build and deliver both the accuracy and the TCO needed to be in the market.

More on AI tomorrow.

About the author

Kevin Levitt is the Senior Director of Global Industry Business Development for Financial Services at NVIDIA, part of NVIDIA's global financial services team. The team is responsible for setting NVIDIA's strategy in the ecosystem, focusing primarily on trading, banking, and payments, understanding the priority workloads and use cases for NVIDIA's accelerated computing platform, and helping both end customers and our partner ecosystem leverage it to build and deploy AI-enabled applications at scale.