1. Why are banks tightening bank AI spending?
Major banks including JPMorgan Chase, Citigroup and Canadian Imperial Bank of Commerce (CIBC) are tightening controls on bank AI spending as the computing costs of artificial intelligence rise and become harder to predict. Reports on October 21 and 22, 2026 describe finance and technology leaders at the lenders trying to keep the bills in check while the technology spreads through their operations.
For several years, Wall Street has pushed AI into more of its daily work, from software development to customer service and internal research. The invoices for that expansion are now arriving, and the people responsible for budgets are scrutinizing them. The shift marks a change in tone from an era of experimentation, when the main question was what the tools could do, to one in which the question is also what they cost to run.
The measures being adopted are practical rather than dramatic. They include monthly spending ceilings, limits on who may use the most expensive models, and routing work to cheaper alternatives where the quality is adequate. None of them signals a retreat from AI. They signal that the largest banks now treat AI compute as a line item that needs the same discipline as any other large expense.
Understanding why this matters requires separating the headline spending figures from the way the money is actually consumed. Much of the cost is not a purchase of software but a charge for processing power, which rises with every question asked of a model. That makes the expense behave more like a utility bill than a license fee, and utility-style costs call for monitoring in real time.
2. How much does JPMorgan spend on AI?
JPMorgan Chase Chief Financial Officer Jeremy Barnum has said the bank spends approximately $700 million a year on AI-related costs. The figure gives investors and analysts a concrete benchmark for what a global bank with a very large technology organization pays for these tools.
The bank has also set monthly spending caps on its most expensive models, including Claude Code, the coding tool from Anthropic. Access to it is limited to selected engineers, and each of them faces a ceiling of $2,000 per month. Engineers who reach the limit can ask for it to be raised, which suggests the cap works as a control point and a prompt for review rather than as a hard prohibition.
A cap of that kind addresses a specific risk. A coding assistant can run for long stretches and consume a large amount of computing power in a single session, particularly when it is allowed to act on its own across many steps. Without a ceiling, one engineer's heavy use could turn into a large and unexpected charge. With a ceiling, the bank knows the maximum exposure per person in advance.
The $700 million number does not come with a comparison point in the reporting, so its size relative to revenue, to the bank's overall technology budget or to its peers is not clear from the figures alone. What it does show is that AI spending at the largest US bank is large enough to be tracked and discussed at the level of the chief financial officer.
3. Why are AI bills so hard to forecast?
The central difficulty is how AI services are billed. Early enterprise deals often looked like ordinary software subscriptions, with a fixed monthly fee per user. A finance team could multiply the fee by the number of seats and know the annual cost with reasonable confidence.
A growing number of vendors now use usage-based billing instead. Under that model the customer pays for the computing resources consumed, in particular inference, which is the processing a model performs each time it answers a request. The more a tool is used, and the more complicated each request is, the larger the bill becomes.
Usage-based pricing is manageable in a small pilot, where a limited group of employees tries a tool and the total stays modest. It becomes far more complicated when the same tools reach full production, with thousands of employees and automated systems querying models every day. Costs then depend on behavior that is difficult to estimate in advance.
Agentic AI adds a further layer of uncertainty. These systems carry out multi-step tasks on their own, and the amount of computing a task needs can vary widely depending on how many steps are required and how much information the system has to process. Two tasks that look similar to a user can produce very different charges.
AI costs also compete for money with other priorities. Modernizing legacy systems and strengthening cybersecurity were already demanding large budgets at banks, so a new and variable expense puts additional pressure on planning.
The practical consequence is that a budget set at the start of the year can be overtaken by changes in how employees work. A tool that becomes popular, or a workflow that is automated, can multiply consumption without any new contract being signed.
4. What is the case for continuing to spend?
Executives on third-quarter earnings calls have tried to hold two ideas together: costs need to be controlled, and the technology is delivering gains. Citigroup is one example of that balance. Its leadership has linked AI tools to faster product launches, and its chief executive said the bank has reduced its technology vendor spending by 30% year over year as a consequence.
That vendor-spending figure is a separate measure from the cost of running AI queries, so it should not be read as a direct offset to the compute bill. It does, however, illustrate the argument banks make when defending the investment: that AI can lower other costs or speed up work, and that the savings can outweigh what the tools consume.
The strength of that argument depends on measurement. A productivity gain is only valuable if it is larger than the full expense of producing it, including compute. As the bills grow, banks have more reason to track results closely and to match the power of a model to the importance of the task it is performing.
5. How are banks controlling the costs?
Banks are relying on a handful of tactics. Usage caps are the bluntest of them. JPMorgan's $2,000 monthly ceiling on Claude Code for selected engineers is an example: it fixes a maximum for each user and forces a conversation when someone reaches it.
A second approach is model-agnostic routing. Instead of sending every request to the most powerful and most expensive model, a bank can send simpler tasks to cheaper models and reserve the premium option for work that needs it. JPMorgan is pursuing a model-agnostic strategy, and the same logic is behind the wider interest in smaller models and in alternatives with lower prices per token.
A third lever is restricting access. CIBC has taken that route, limiting who can use its premium AI models. Under this approach, expensive tools go to the employees whose work justifies them, while others use lighter tools that cost less.
Each tactic involves a tradeoff. Caps can slow down a productive employee, routing requires careful testing so that a cheaper model does not degrade quality, and restricted access can leave staff without a tool they would have used well. The banks are weighing those tradeoffs against the cost of leaving usage unchecked.
6. What does this mean for investors and AI vendors?
For investors in banks, rising and unpredictable AI costs could put pressure on profit margins. The productivity gains that executives describe only improve results if they exceed what it costs to generate them, so the compute line is becoming a figure worth watching alongside the more familiar expense categories.
JPMorgan's disclosed annual figure sets a reference point. Other banks may now face questions about how much they spend on AI and whether their claims of efficiency still hold once computing costs are counted in full.
For AI vendors, the behavior of their largest customers carries a message. Spending caps, limits on access and routing to cheaper models suggest that buyers will increasingly compare prices rather than default to the premium option. Vendors that rely on usage-based billing may find that large clients respond by constraining how much of the product their staff can use.
The result is a market in which cost discipline becomes part of how AI is purchased, even at institutions with very large technology budgets. That is a change from the early period of adoption, when enthusiasm often outran the accounting.
Over time, such pressures may reward vendors that offer predictable pricing, and they may reward banks that build the internal tooling to measure what each team spends. Visibility into consumption is the precondition for every control described above.
7. What to watch next
The reporting does not set out a specific upcoming date or decision, but the direction is clear enough to watch. Banks will continue to report earnings and to field questions from investors about technology spending, and AI costs are likely to come up in those discussions.
Three developments would show how the situation is evolving: whether more banks disclose an annual AI cost figure similar to JPMorgan's, whether vendors adjust their pricing in response to cost-conscious customers, and whether banks keep expanding their use of agentic systems, whose costs are the hardest to predict.
For now, the largest lenders are signaling that AI remains a priority but that it will be managed like any other major expense, with limits set in advance and spending reviewed against the results it delivers.