The AI Boom Has Quietly Become a Credit Story, and the Bond Market Is Blinking First

We have been telling ourselves the wrong story about artificial intelligence. For three years the boom has been framed as an equity narrative, a wager on productivity so vast that valuation multiples became almost beside the point. That framing is now obsolete. The buildout has grown too large to be financed out of profits, and so it is being financed out of debt. Artificial intelligence has quietly become a credit story, and in the second week of July the credit market began, for the first time, to price the possibility that the story does not end well.

Consider the plumbing. According to Morningstar, investment grade bonds issued by hyperscalers, data center developers and adjacent AI concerns reached 218 billion dollars through July 8 of this year, a figure that dwarfs the 80.5 billion dollars raised across all of 2025. That is not a rounding error in the capital markets. It is a new and dominant supply of paper, and the buy side is beginning to choke on it.

Spreads Speak Before Stock Prices Do

The tell is in the spreads. Morningstar notes that new SpaceX 6.65 percent thirty-year bonds traded above 200 basis points over Treasuries this week, having priced at 175. Meta's 6.3 percent 2056 bonds, part of a 25 billion dollar package inked on April 30, drifted to their widest level yet at 145 over, roughly 25 basis points cheaper than where they changed hands a month ago. These are small numbers to a casual reader and enormous ones to a credit desk. Widening spreads are the sound of lenders demanding more to hold the same risk, and they tend to precede trouble in equities rather than follow it.

This matters because the equity market has convinced itself of the opposite. J.P. Morgan Asset Management observes that the ten largest stocks now account for roughly 35 percent of the S&P 500, against about 25 percent at the peak of the dot-com era. An index that concentrated is no longer a diversified holding. It is a leveraged directional bet on uninterrupted AI momentum, and the leverage is increasingly literal.

Cash Flow Was the Old Discipline; Borrowing Is the New Permission

For years the bulls had a genuinely strong argument. The hyperscalers were spending prodigiously, but they were spending their own money, throwing off cash faster than they could consume it. That discipline is dissolving. Reporting compiled by Epoch AI and others suggests operating cash flow across the group is growing at roughly 23 percent a year while cash capital expenditure grows near 70 percent, two lines that cross around the third quarter of this year, at which point aggregate free cash flow approaches zero.

The individual figures are stark. Amazon's trailing free cash flow has collapsed from about 26 billion dollars a year ago to a little over 1 billion; Alphabet's is down roughly 38 percent; Microsoft's has fallen more than a fifth. To keep the buildout going, the four largest spenders are expected to issue more than 400 billion dollars in new debt this year, more than double the 165 billion of 2025. Morgan Stanley and JPMorgan analysts have floated the possibility that the sector may need 1.5 trillion dollars of fresh debt over the next several years. Once a company stops funding its ambitions from earnings and starts funding them from bond investors, the relevant question changes. It is no longer whether the technology is transformative. It is whether the cash arrives before the coupon does.

Accounting Can Flatter Earnings but Cannot Repay a Bond

Here lies the uncomfortable part. Even the profits underwriting all this borrowing may be softer than they appear. Michael Burry has argued that the hyperscalers are stretching the assumed useful life of their AI hardware, depreciating chips over five and six years when the working life is closer to two or three. Microsoft extended its GPU depreciation schedule from four years to six; Meta moved to five and a half. By Burry's estimate this understates depreciation by around 176 billion dollars between 2026 and 2028 and inflates reported profits by more than 20 percent, with Oracle and Meta among the most affected. One need not accept every number to grasp the point.

Understating depreciation by extending the useful life of assets artificially boosts earnings, and it is one of the more common frauds of the modern era, Burry has written.

A bond investor cannot be paid in flattering accounting. The coupon is settled in cash, and cash is exactly the thing the depreciation debate calls into doubt. If the profits financing the debt are partly a product of generous schedules, the margin of safety is thinner than the ratings imply.

Warsh Removes the Cushion the Boom Was Counting On

All of this would be manageable in a world of cheap money. That world is closing. Federal Reserve Chair Kevin Warsh, speaking at the ECB forum in Sintra on July 1, delivered what CNBC described as an unambiguously hawkish message, declaring that anyone expecting the central bank to tolerate inflation above 2 percent would be disappointed. He was speaking against May PCE inflation of 4.1 percent, the highest since April 2023, with core running at 3.4 percent. Markets have moved accordingly, pricing meaningful odds of a rate increase later this summer.

A hawkish Fed is the specific poison for a debt-financed capital cycle. Higher rates raise the coupon on every new deal, widen the spread demanded on every refinancing, and shorten the runway for projects whose payback was already being pushed out each quarter. Sequoia's David Cahn has estimated a revenue gap of roughly 600 billion dollars between AI infrastructure spending and the ecosystem revenue meant to justify it. Financing that gap was tolerable at zero. It becomes a genuine strain when the risk-free rate is climbing and the central bank has promised not to blink.

Prudence, Not Prophecy, Is the Discipline This Moment Rewards

None of this is a prediction of collapse. The technology is real, the demand for compute is real, and some of these bets will pay handsomely. Our argument is narrower and, we think, harder to dodge. The AI trade has migrated from the equity market, where optimism is cheap and self-correcting, into the credit market, where optimism is expensive and the reckoning arrives on a fixed schedule. When investment grade spreads widen on the very names the equity market is bidding to records, the sensible investor listens to the bond desk, not the earnings call.

For a generation raised on the idea that technology and leverage are separate conversations, this is the lesson of July. They are now the same conversation. The buildout that was supposed to prove itself through productivity will instead be judged, first, by whether it can service its debt. That judgment has begun, and it is being rendered one basis point at a time.