The news: Capital One is exploring alternatives to Amazon Web Services (AWS) for AI workloads, according to an internal Nvidia memo, as the bank’s needs may cause its AWS bills to “soon get out of hand.” Capital One reportedly discussed in-house data centers with Nvidia that could be used instead of or in addition to public-cloud providers to train and operate AI models.

Zoom out: Capital One’s move is a natural extension of its leading AI strategy as it adapts its infrastructure to the rising complexity and cost of new models. The bank was an early mover among financial institutions (FIs) in the space. It migrated all eight of its data centers to the public cloud by 2020. And this year it came in second to only JPMorgan Chase in Evident’s AI Index measuring the AI maturity of global FIs.

Our take: The race among large FIs to adopt genAI is top of mind in the industry. Forty-seven percent of US banking decision-makers at banks with assets above $20 billion say their institutions have already rolled it out, according to a 2025 EY-Parthenon survey. And the biggest report spending huge sums: JPMorgan said recently that it spends $2 billion per year on AI alone out of its annual technology spend of about $18 billion.

But focusing on AI adoption alone distracts from the equally consequential evolution of banking infrastructure. Cloud banking—moving a bank’s workloads to a service like AWS—was news five years ago. Now it’s AI cloud banking, a separate but related discipline that’s more complex and expensive while also requiring a level of technical expertise that few banks have on staff. FIs may struggle to grapple with the additional challenge of managing those costs.