JPMorgan says its AI financing exposure is modest but not immune to a reversal

JPMorgan Chase is weighing how much more of the artificial intelligence buildout it wants to finance, Chief Financial Officer Jeremy Barnum told analysts Tuesday on the bank’s third-quarter conference call.

Even the nation’s largest bank would not be immune to a major shift in AI sentiment, Barnum said.

JPMorgan’s current exposure to the AI ecosystem is “quite modest” relative to its overall wholesale exposure, he said.

The bigger question, Barnum said, is how much more exposure JPMorgan wants as companies pursue their announced capital spending plans and seek financing for them.

Barnum said JPMorgan will manage its exposure to any single borrower. The deals require close analysis of direct and indirect credit exposure, where a lender stands in the capital structure and what the collateral is worth.

Borrowers have tapped more parts of the capital markets over the past year to finance the buildout, Barnum said. Some market participants have raised concerns that the demand could crowd out other borrowers.

AI financing now extends beyond construction loans into working capital credit lines and equity markets, Barnum said.

JPMorgan stress tests how the firm would fare if the AI outlook changed sharply, Barnum said. Barnum said he would be surprised if any institution escaped a major shift in sentiment.

The largest technology companies are spending on data centers at a pace with few precedents. Amazon, Microsoft, Alphabet and Meta are on track to spend roughly $600 billion to more than $700 billion this year on capital projects, depending on which companies analysts count. Most of that money goes to data centers, chips and related equipment.

Amazon has said it expects to spend about $200 billion, Alphabet up to $185 billion and Meta $125 billion to $145 billion.

Companies are borrowing a growing share of the money. Morgan Stanley expects AI-related debt sales worldwide to reach nearly $570 billion this year. Borrowers had raised about $236 billion by the end of May.

Meta sold $25 billion in bonds this spring, which was then the largest single data center financing.

Higher-risk borrowing is also rising. U.S. AI infrastructure projects issued $26.6 billion in high-yield, or junk, bonds in the first four months of the year, more than double the total for all of 2025, according to PitchBook.

Private credit funds, which lend outside the traditional banking system, are taking a larger role. Morgan Stanley has forecast that they could supply more than half of the estimated $1.5 trillion needed for data center expansion through 2028.

In a Bank of America survey this summer, 48 percent of fund managers named hyperscaler spending as the most likely source of a systemic credit event.

Barnum said the wave of AI-driven innovation and the many ways companies are raising money require JPMorgan to watch its credit exposure closely.

In response to a separate question, CEO Jamie Dimon said the bank is tightening credit standards slightly, citing looser loan terms such as weaker covenants.

Barnum said a previously cited figure of about $2 billion in annual AI spending was not exactly right. AI expense at the time was about $700 million a year, he said.

AI spending now runs across the bank, making it difficult to tally separately, Barnum said. He is also moving away from putting a number on its benefits.

Those benefits have yet to show up in quarterly earnings, Barnum said. He cited software engineering as a promising area. In his own finance technology group, the bank tried using AI to rewrite minor legacy tools in a couple of weeks. The results were imperfect, but suggested ways to speed development.

Token costs will offset some of the productivity gains, he said.

JPMorgan aims to control those costs by avoiding dependence on any one AI provider, Barnum said. The bank can run proprietary, open-weight and open-source models, including some in its own data center.

Dimon said model-routing tools are coming that will send each query to the best, fastest and cheapest model.

JPMorgan uses AI agents internally and plans to use them for customers, Dimon said. But the bank won’t let just any outside agent into the company because of the risk to clients, he said.

He warned of rogue agents, saying “there could be thousands of them.” Customers who give agents access to their accounts need to understand they are accepting a liability transfer, Dimon said.

Barnum said JPMorgan blocks agents that reach customer accounts through shared credentials where it can, a practice the bank has long opposed. Personal AI agents released this quarter moved bank stocks, Barnum said. Barnum said the competition combines two pressures banks already face: depositors seeking higher yields and fintech rivals.

Schorr asked whether agents moving money for customers could accelerate deposit runs. Barnum said he is skeptical that the technology would cause a sudden rush of withdrawals.

Dimon said the bank holds $1.5 trillion in cash and marketable securities and that regulators should prepare for risks posed by bots. AI also adds uncertainty to JPMorgan’s 2027 expense planning, Barnum said. The bank’s preliminary view is close to the analyst consensus of about $114 billion, he said.

JPMorgan reported third-quarter net income of $15.3 billion, or $5.57 per share, on revenue of $50.9 billion, up 8 percent from a year earlier. Analysts had expected $5.90 per share, according to Forbes.