JPMorgan, Citi and CIBC tighten controls on AI spending as computing costs rise

JPMorgan Chase, Citigroup and Canadian Imperial Bank of Commerce have tightened controls on AI spending as computing costs rise and get harder to forecast, according to a Bloomberg report from Oct. 21-22, 2026 as relayed by Crypto Briefing.1

The cause is a shift in billing: early AI deals often charged a fixed fee per user each month, while more of the market now bills by usage, measured in compute and inference, according to Crypto Briefing's account.1 Usage billing is manageable in a small pilot but harder in full production, where thousands of employees and automated systems query models all day, and agentic AI tasks vary widely in compute needs, Crypto Briefing reported.1

JPMorgan CFO Jeremy Barnum said the bank spends approximately $700 million a year on AI-related costs, per the Bloomberg report as relayed by Crypto Briefing; the primary Bloomberg report was not among the sources provided.1

JPMorgan set monthly 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.1

CIBC found premium AI models can cost as much as C$3,000 per month per employee, restricted who gets access, and is looking at limits inside internal workspaces.1

Spain's BBVA cut the cost of a loan approval process from €5 to €0.03 by optimizing how it uses AI, according to Crypto Briefing; the sources do not say how Citi's specific controls work.1