Alphabet's 100-year sterling bond fell below 90 pence on July 23 for the first time since its February issuance, a quiet but significant signal that bond investors are starting to question the AI debt buildout before it has proved itself.
When Alphabet sold a £1 billion sterling bond maturing in 2126 back in February, investors couldn't get enough of it. Bloomberg reported at the time that the order book reached £9.5 billion, close to ten times the amount on offer. Five months later, Bloomberg data cited in market reports showed the same bond quoted at 89.978 pence on the pound. Anyone who bought at par has already watched more than 10% of the price disappear. The century bond, once a badge of investor confidence in the AI infrastructure story, has become a proxy for a much harder question: will any of this actually pay off?
The market's answer is not encouraging.
As Fortune reported on July 17, orders for hyperscaler bonds covered nearly five times the amount offered in February. By July, that coverage had slipped below two times. That collapse in demand matters because it happened while issuance was still accelerating, not slowing. Reuters reported in June that Morgan Stanley estimated AI-related issuers had already sold nearly $236 billion in debt globally by the end of May, about four times the pace of a year earlier, and the bank projected the 2026 total could reach nearly $570 billion. Nvidia raised $25 billion in June, its first corporate bond sale since 2021, according to Reuters. SpaceX priced its own $25 billion inaugural bond offering later that month, the company said. Amazon then had to sweeten a $25 billion July sale by offering 18 to 21 basis points of extra yield on its longest-dated paper, Fortune reported, while orders fell to 2.5 times the bonds on offer from 3.2 times in March.
More supply, weaker demand, and rising yields on new issuance: that's a credit market telling you something has changed.
The latest shove came from Moonshot AI's release of Kimi K3 on July 16, a 2.8-trillion-parameter open-weight model that traders quickly labeled another DeepSeek moment. The Financial Times reported that the Philadelphia Semiconductor Index fell nearly 10% that week and moved more than 20% below its June peak, putting chip stocks in bear market territory. The logic was blunt. If a Chinese lab can release a frontier-class open model at a much lower price than top U.S. systems, the premise that hyperscalers need to borrow hundreds of billions to build their way to AI dominance starts to look shakier. And if that premise is shaky, so is the debt that rests on it.
Startups feel the repricing next #
Frankly, the bond market is a more honest judge of this than equity markets are, because bond investors don't get upside. They lend money. They expect it back. When they start demanding wider spreads and coverage ratios halve in five months, they're not making a bet on AI's long-term potential. They're doing a narrower calculation: will this company generate enough cash to service this debt before the paper matures? For a 100-year bond on a company whose core business depends on an AI transition that has not yet produced revenue at the scale the capex implies, that calculation is genuinely hard.
The startup effect is real. Quartz's account of the Reuters report cited UBS data showing hyperscaler capital expenditure in 2026 running close to 100% of operating cash flows, versus a 10-year average of roughly 40%. If bond investors keep demanding higher yields to absorb new supply, and if secondary market prices keep falling, the appetite to write more checks into the AI infrastructure chain tightens. Cloud credits and compute access, the oxygen of the early-stage AI startup, come from the same balance sheets now under scrutiny. A repricing of hyperscaler debt doesn't stop the music immediately, but it does raise the cost of keeping it playing.
The hidden debt is harder to price #
The private-credit exposure is murkier. IESE Insight put recent off-balance-sheet hyperscaler financing at roughly $65 billion, including Meta's $27 billion private-credit joint venture with Blue Owl for its Hyperion data center campus in Louisiana. Separate Morgan Stanley projections cited by The Information put private-credit funds' data-center debt opportunity at about $800 billion by 2028. That is not a rounding error. It means the visible bond market is only part of the exposure, and some of the risk now sits in structures investors can't track as cleanly as public corporate bonds.
None of this means the AI buildout stops. These are companies with enormous operating businesses and investment-grade credit ratings. But the century bond was never just a financing instrument. It was a statement of confidence: lend us money for a hundred years and you'll be fine. At 89.978 pence, the market is quietly saying it's not so sure. You should pay attention.
Also read: SpaceX's record Nasdaq debut anchors a $106 billion IPO quarter and signals the widest public exit window in years • The EU cleared the Saudi-Kushner buyout of Electronic Arts and now only Washington stands in the way • Etched is raising at $10 billion and $20 billion simultaneously and investors are lined up