In August, Alphabet, one of the most valuable companies on the planet, went to raise money in the bond market. To get the deal done, it had to sweeten the offer. Investors weren’t rushing in.
Around the same time, an insurance broker most people have never heard of asked for $13.5 billion to fund an acquisition. Investors sent in $65 billion worth of orders. The pricing on part of that deal tightened by 35 basis points because so many buyers wanted in.
Same bond market. Same week, roughly. Two completely different receptions.
That contrast isn’t a fluke. It’s the clearest sign yet of a shift happening in the AI debt market right now. Portfolio managers aren’t worried these companies will default. Hyperscalers are still enormously profitable, and nobody credible is questioning their ability to pay their bills. What’s changed is how much unchecked AI debt the market is willing to absorb at once, and at what price.
The AI boom isn’t losing its believers. It’s just started making them work harder to say yes.
The Numbers Behind the AI Debt Caution
The scale of AI borrowing is what’s driving this shift. Gross debt issuance from hyperscalers is expected to hit a record $420 billion next year, a 60% jump from 2026 estimates, according to Goldman Sachs data. That’s not gradual growth. That’s a sprint to fund chips, power infrastructure, and data center capacity fast enough to keep up with demand.
The pricing gap makes the divide official. Spreads on AI-related bonds have widened to around 115 basis points, compared with 78 basis points for the broader investment-grade market, based on the latest Goldman and ICE BofA data. A wider spread simply means investors are charging AI borrowers more to lend them money, even when those borrowers carry strong credit ratings. Alphabet’s August concession and Aon’s oversubscribed order book are two snapshots of that same widening gap playing out in real time.
How Big Has the Borrowing Already Gotten
To understand why investors are getting cautious, it helps to see how fast the borrowing itself has scaled.
Big Tech raised a record $108 billion in debt in 2025, more than three times the average of the previous nine years, according to Nomura. And that was before this year’s acceleration. AI-related debt issuance has already topped $220 billion in 2026, with Morgan Stanley estimating the total could reach roughly $570 billion by year-end.
Some of that borrowing is happening at eye-catching size. Broadcom was reportedly in talks over the summer to raise more than $60 billion in debt to help fund chip and computing capacity for Anthropic and other AI companies, a single deal that underscored a bigger concern hanging over the entire AI trade: the sheer cost of funding all of it. Whichever year-end estimate proves closer to accurate, the direction is unmistakable. Borrowing is compounding faster than most credit desks expected even a year ago.
Why This Isn't the Same as a Default Scare
It’s worth being precise about what’s actually happening here. This isn’t a credit-quality panic. It’s a capacity problem.
One market analysis described it well: the bond market isn’t voting on whether AI companies can pay their debts. It’s judging how much AI-linked supply the market can comfortably absorb at once. Investors aren’t fleeing these borrowers because they expect defaults. They’re hesitating because the pace of issuance is testing how much concentration risk they want to carry in one basket.
That distinction matters for how you read this trend. It’s less “AI debt is dangerous” and more “AI debt is arriving faster than the market can digest it comfortably.” Companies are still spending aggressively on infrastructure the way they’ve been reworking technology operating models to support that growth, but the financing side hasn’t scaled as smoothly as the spending side.
Underwriters Are Already Adapting
The shift shows up in how deals get structured now, not just how they get priced.
When BlackRock arranged billions in debt financing for a Meta data center earlier this year, it deliberately steered away from short-term traders. The underwriters favored real-money accounts like pension and insurance funds instead, institutions that typically buy and hold rather than trade in and out, specifically to insulate the deal from the volatility hitting AI-linked debt elsewhere.
That’s a meaningful change in dealmaking behavior. It suggests banks and asset managers now assume some investors will treat AI bonds cautiously, so they’re building deals around buyers who won’t panic-sell if sentiment sours.
The Cash Flow Squeeze Behind the AI Buildout
Here’s the part that deserves more attention than it gets: the hyperscalers funding this buildout aren’t sitting on the spare cash they used to be.
Aggregate free cash flow across several of the largest hyperscalers peaked near $400 billion in late 2024. By the end of 2026, estimates put that figure at roughly $21 billion, as capital spending absorbs almost everything coming in the door. These companies aren’t earning less. They’re spending so heavily on AI infrastructure that there’s barely anything left over once the bills are paid.
Alphabet’s own numbers illustrate the squeeze. Last quarter, the company spent $44.9 billion on capital expenditures while generating $39.1 billion in operating cash flow, pushing its quarterly free cash flow negative for the first time since it went public. That’s a striking shift for a company that has historically thrown off cash by the tens of billions.
It’s a good reminder that even well-run, cash-rich organizations aren’t immune to the kind of overspending traps that derail digital transformation initiatives when capital outpaces the ability to absorb it responsibly.
Investors Are Watching Concentration, Not Just Credit Ratings
One detail from the reporting stands out: investors aren’t just tracking how much debt a single company issues directly. They’re adding up exposure across every financing vehicle tied to that company, including the special-purpose entities built specifically to fund data centers.
Some institutional investors are approaching single-company exposure limits once they count debt issued through data-center financing vehicles alongside the bonds issued directly by the corporate parent, according to Wellington Management portfolio manager Loren Moran. One Thornburg Investment Management strategist summed up the underlying issue simply: investors can only digest so much debt, so fast.
That’s a structural constraint, not a mood swing. It also explains why companies are getting creative, increasingly turning to special-purpose financing vehicles and asset-backed securities to fund data center construction outside their traditional borrowing channels. The AI buildout is starting to resemble the semiconductor world’s supply crunch a few years back, where demand outran the infrastructure needed to support it, a dynamic that shares more in common with the 2021 chip shortage than most people realize. In both cases, the bottleneck wasn’t appetite. It was capacity to absorb growth safely.
What This Means Going Forward
None of this means the AI boom is collapsing or that hyperscalers are in financial trouble. What’s changed is the market’s patience for absorbing unlimited AI-linked supply at the same easy pricing it once offered.
For companies building AI products and infrastructure, the lesson travels well beyond Wall Street. The gap between ambitious AI investment and the operational discipline needed to sustain it is exactly the kind of gap that shows up later in production, not just in financing. It’s the same gap that separates AI pilots from systems that actually hold up once deployed, something worth understanding through how AI agents perform in real production environments versus in demos. Financing discipline and deployment discipline tend to rise or fall together.
If Wall Street’s growing selectiveness sticks, it could slow the pace of the AI infrastructure race, or at least force a more disciplined version of it. Either way, the message from credit markets is clear: enthusiasm for AI hasn’t disappeared. It’s just no longer unconditional.
