Stress testing the stress tests

The good news is, the U.S. banking system is sound. That is the broadest takeaway of the Federal Reserve's latest stress test, as our Kyle Campbell reported. The central bank ran a hypothetical: assume a severe recession hits the U.S. Could the banks withstand it? 32 of the nation's largest banks took part in the test, and the Fed concluded that all 32 of them could, indeed, withstand the stress.

Maybe I'm just getting old and cranky, but those sunny results really stick in my craw. All of them? Not one bank would fail? Not one bank would suffer serious losses and need a rescue? It's like a class full of students taking a test and everyone gets an A. It just seems off, like either the test was too easy, or some people were cheating. The face-value results are just too convenient. There has to be a flaw in there somewhere.

To be fair, every year the results are like this, and it's worked so far. Even when the pandemic hit in 2020, the banking sector was in fact able to handle the pressures. And the "adverse" scenario in the test against which the Fed subjects the banks to analysis is pretty adverse. The theoretical imagines the stock market down nearly 60%. Residential real estate prices tumbling 30%; commercial real estate prices down almost 40%. Credit spreads widening by 4.7% percentage points. 

Those numbers are worse than the numbers seen during the 2008 financial crisis. And while all 32 banks would in that hypothetical scenario still be able to operate and lend and keep the economy lubricated with credit, it wouldn't be painless. Seventeen of the 32 would report negative pretax income, the Fed found. So maybe everybody didn't get an A after all. Looks like there were a few Bs and Cs in there. 

Now, in the real world today U.S. banks are extremely profitable. Looking at the first-quarter results, I was struck by how many banks were not only reporting multi-billion profits, but also multi-billion stock buybacks. So we don't really need a stress test to tell us the banks are healthy. We know they are healthy. But the goal of the stress tests isn't necessarily to determine whether the banks are solvent. The goal is to show that the credit market could keep working through a severe economic crisis. That's a subtle but important difference.

And that, I think, is where you can start to see maybe where the limitations of the abstract stress tests are missing the risks to the system. Because the credit market is not only those 32 banks, or even of all the banks in the U.S. The plain reality is that more credit is coming out of the so-called nonbank sector than is coming from banks. In 2023, 50% of all global credit came from nonbank lenders, according to the International Monetary Fund.

Partly that is by design. The post-crisis reforms were intended to keep the banks from taking on too much risk, from getting themselves caught up in another subprime bubble like in the Aughts. But the demand for that kind of credit didn't go away. The nonbank sector has been servicing all that risk that regulators won't let regular banks take. But it's not for lack of trying. U.S. banks have more than $1 trillion in loans to nonbanks. Private credit firms use banks and insurance companies for funding. In other words, there are material connections between banks and nonbanks.

So, here's a stress test scenario you can try at home: The AI industry is going to borrow several trillion over the next couple of years to build out data centers and perfect the HAL 9000 and put people on Mars. Some will come from banks, but most likely will come from nonbanks. To support that much borrowing, the AI industry will need to gin up somewhere in the vicinity of $2 trillion in new revenue by 2030, according to consultancy Bain. So, like, what if it can't? What if the AI industry doesn't come up with the money? OpenAI, the most prominent of the companies in the space, had 2025 revenue of about $13 billion. How much higher does it really expect that to go? We'll find out later this summer when the company releases the prospectus for its planned IPO. No matter what they say, it may not be enough. Bain estimated the industry will be $800 billion short of revenue to profitably run its data centers. 

Those aren't just some theoretical, hypothetical, abstract numbers. That is a real issue that could really manifest, and nonbanks and banks are exposed to it. And the AI sector is not just some hived-off science project. AI contributed more to economic growth in 2025 than tech did during the dot-com boom, according to the Federal Reserve Bank of St. Louis. It's become central to the economy's growth, it's borrowing trillions, and it needs to drive trillions more in revenue growth to pay back the money. How bad will the damage to the economy be if the AI industry is the next railroad, dot-com, telecom bubble popping? How well will the credit markets weather that storm? How well protected are the banks really? How protected is their balance sheet from marks they don't control?

Now that, Mouseketeers, is what I call a stress test.