Leadership1 publisher3 min readPublished
JPMorgan put a $2,000 monthly ceiling on some engineers' Claude Code spending
Some of the bank's engineers found the limit by hitting it. The number arrived alongside a new sandboxed coding environment hosted in AWS, and JPMorgan would not say what its token limits or cost strategy are.
The Board Room · Leadership desk
What happened
- JPMorgan has set a $2,000 monthly spending limit for some employees using Claude Code, according to internal messages seen by Business Insider.
- Engineers learned of it from a new error code that surfaced in a Teams group last month, reading "ExceededBudget" and "Budget=2000.0".
- Employees in the group said the budget resets monthly and that users can request an increase to their limit.
- Business Insider reported that the bank has tracked engineers' AI usage for months, and that this is the first indication of AI spending or token caps.
- On the July second-quarter call, CFO Jeremy Barnum called token expenses "trivial" at that point while forecasting "meaningful acceleration in that number" in the second half.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision Because the ceiling can be raised on request, an engineer who needs more capacity now has to ask someone who owns the number, and that person has to approve or refuse it in writing.
- constraint Enforcement sits inside the environment the bank operates, so the cap reaches only as far as engineers have moved off desktop Claude.
- contradiction In June the bank's payments data chief described staff outspending their salaries on tokens with no firmwide pullback underway. The limit now in force covers certain engineers, and the metering is selective.
- precedent A named budget error inside the coding tool makes per-engineer, per-month AI allowances a normal provisioning unit.
A spending cap has to be enforced somewhere. Until recently, some of the bank's developers piloted Claude directly on their desktops [8]. The limit was mentioned in a Teams group about a new coding architecture called Devspace [11], which an internal technical blog post describes as a "containerised," sandbox-like server hosted in Amazon Web Services [9]. A JPMorgan spokesperson said Devspace is a longer-term plan that restricts Claude's access to employees' credentials and its ability to interact with internal systems [10].
Two thousand dollars a month is $24,000 a year for one engineer [24]. Extend it across the whole Global Technology division and the ceiling is $130 million a month [19], or $1.56 billion a year [20] against a technology budget of nearly $20 billion [5], about 7.8 percent of it [21]. Business Insider reported that it is unclear when the firm began introducing Devspace or how many engineers have access [13].
A technology budget of nearly $20 billion now meters coding assistance per head [5][1]. Per-seat AI consumption joins cloud spend among the things an engineering manager defends at review. Zachery Anderson, JPMorgan's chief data and analytics officer in payments, said at a tech event in June that some employees are "spending more on tokens than their salary." There wasn't a firmwide effort to scale back usage, he said, according to Semafor [7]. Certain engineers can ask to have the ceiling raised [3].
The security case for the container was made in public months before the cap. Pat Opet, JPMorgan's global chief information security officer, spoke at a cybersecurity conference in April. Putting powerful AI tools on the same desktops that house employees' emails, sensitive files and personal information can be risky, he said [17]. "Ideally, we would want these agents to run in an ecosystem where they have no entitlements. They have an identity, but no entitlements," he said [16]. If an agent figured out how to break out of the environment, Opet said, it would not have the credentials to reach internal systems or "abuse the enterprise" [18].
An agent with an identity and no standing entitlements is also an agent whose consumption belongs to a named person on a server the bank runs. A per-engineer budget needs that. This quarter the cap tells engineers that a number exists. Next quarter the record a budget owner carries into planning is the queue of requests to raise it, and what the requesters shipped with the extra headroom. JPMorgan declined to comment on the specifics of token limits or its cost strategy [14]. A spokesperson said the bank is "approaching AI with discipline, tying costs to measurable business value" and considers a range of metrics including adoption, outcomes, productivity improvements, quality, speed, capacity creation, risk reduction and business impact [15].
What to watch
- Whether the $2,000 ceiling extends from certain engineers to the wider Global Technology division as Devspace rolls out.
- Whether the token acceleration Barnum forecast for the second half shows up as a named cost line on a later earnings call.
- How many requests to raise the limit are granted, and against which of the outcome metrics the spokesperson listed.