Wall Street spent the last couple of years racing to put artificial intelligence into every corner of the business. Now the invoices are arriving, and bank executives are reading them closely.
Finance chiefs and technology leaders at major banks are dealing with AI computing costs that are rising and getting harder to forecast, according to a Bloomberg report from October 21-22, 2026. JPMorgan Chase, Citigroup and Canadian Imperial Bank of Commerce have all tightened controls on AI spending as a result.
The price of a smarter workforce
JPMorgan Chase CFO Jeremy Barnum said the bank spends approximately $700 million a year on AI-related costs.
The bank has set monthly spending caps on its priciest models, including Anthropic’s Claude Code, which is available only to selected engineers with a ceiling of $2,000 per month each.
CIBC found that premium AI models can cost as much as C$3,000 per month per employee. Its response was to restrict who gets access, and it is also looking at placing limits inside internal workspaces.
Why the bills are so hard to predict
The core issue is a shift in how AI gets billed. Early deals often looked like software subscriptions: a fixed fee per user, paid every month, easy to budget.
More of the market is moving to usage-based billing instead. You pay for what the model actually does, measured in compute and inference, which is the processing a model runs every time it answers a request.
That works fine during a small pilot. It gets complicated when a bank moves those tools into full production, where thousands of employees and automated systems hit the models all day.
Agentic AI makes the math even messier. These are systems that carry out multi-step tasks on their own rather than simply answering a question. Because some tasks need far more compute than others, the cost of any given job can swing widely.
Adoption of these systems is accelerating. In the first quarter of 2026, banks announced 19 new agentic AI use cases, up from eight the quarter before.
All of this hits technology budgets that were already stretched. Banks still need to modernize aging legacy systems and keep up with cybersecurity demands. AI now competes with both for the same pool of money.
The productivity case is still strong
Third-quarter 2026 earnings calls showed executives trying to rein in costs while also pointing to real gains.
Citi has tied AI tools to a 30-40% improvement in how quickly it launches products. It also reported a 30% year-over-year drop in spending on technology vendors, which it linked to AI implementation.
Spain’s BBVA offers a case study in what efficient use can look like. By optimizing how it uses AI, the bank cut the cost of a loan approval process from €5 to €0.03.
How banks are fighting back
The playbook emerging across the industry has a few common elements. Usage caps, like JPMorgan’s $2,000 monthly ceiling, are the bluntest tool.
Another is model-agnostic routing. Instead of sending every request to the most powerful and expensive model, banks direct simpler tasks to cheaper alternatives.
Restricting access, as CIBC has done, is the third lever. Premium models go to the people whose work justifies the expense, while everyone else uses lighter tools.
What this means for banks and AI vendors
For bank investors, the story is about margins. If AI expenses keep rising unpredictably, they could pressure profitability. The productivity gains are real, but they only matter if they outpace the cost of generating them.
JPMorgan’s willingness to put a figure on its annual AI spending gives investors a benchmark. Other banks may face questions about their own numbers, and about whether their efficiency claims hold up once compute costs are counted.
For AI model providers, spending caps, access restrictions and routing to cheaper models all point toward customers who will shop on price rather than default to the premium option.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.