Memo: Re: The Bond That Matures When Nobody Reading This Is Alive

To: Desk
Re: Hyperscaler debt issuance, long-end yields, and the quiet end of the "AI is self-funding" thesis
Status: Not for external distribution, obviously

Let's start with the number nobody wants to say out loud on a call: hyperscaler gross debt issuance is running at roughly $132 billion year to date, against $93 billion for all of 2025 and an average of $35 billion a year between 2020 and 2024. That's not a trend line, that's a hockey stick, and it includes one multitranche offering north of $53 billion — among the largest corporate bond sales ever priced — plus a genuine curiosity: a hyperscaler century bond, 100-year maturity, due in 2126. Somebody in that syndicate desk actually modeled cash flows to a date at which every person on this memo distribution list is dead, the company's current org chart is trivia, and — if the last three AI cycles are any guide — the underlying business model has been renamed twice. That's not hedging a capex cycle. That's a statement of theological confidence dressed up as a coupon.

The bull case on AI capex has always rested on one load-bearing fact: it's been mostly self-funded. The 28-name direct-AI universe still represents something like 50% of S&P 500 market cap against roughly 5% of S&P 500 net debt, and as of Q3 2025 the share of tech-sector capex and dividends financed by debt rather than operating cash flow was still historically low — nothing like the fiber-optic bubble of the late '90s, where telecoms borrowed themselves into oblivion building capacity nobody used for a decade. That comparison got repeated so often it became a comfort blanket. Repeat a reassuring number enough times and everyone forgets to ask whether the number is still true.

It isn't, anymore, at the margin that matters. Total AI-related debt issuance for 2026 — hyperscalers plus the wider chipmaker, data-center-developer, and utility ecosystem — is now estimated at $300 to $570 billion for the year. That range is wide enough to drive a data center through, which is itself the point: nobody actually knows how much of this buildout is now leveraged, only that the direction of travel is unmistakably up, and that it's happening in the same twelve months the long end of the Treasury curve decided to reprice like it's 2007 again.

Which brings us to this morning's tape. Marvell, Intel, and Lam Research were down 3% premarket. Nvidia off 1%. Oracle down 1.5% ahead of last night's print, which — for what it's worth — came in strong, cloud infrastructure revenue up 121% to $7.4 billion against a $7.19 billion Street estimate, and the stock still couldn't hold the move once the 10-year opened the session near 4.85%, its highest level since November 2023. Three weeks ago the 30-year touched 5.33%, a level last seen in 2007, back when the debt being priced against it was mortgages, not GPUs. Every basis point on that curve is now a direct tax on the terminal value of a data center that won't generate positive free cash flow for years, financed increasingly with paper that has to be rolled, refinanced, or serviced at whatever rate the market decides to charge next quarter — not the rate that looked reasonable when the term sheet was signed.

Here's the mechanism risk, spelled out plainly because nobody on the earnings calls is spelling it out plainly: self-funded capex is a bet on your own cash flow. Debt-funded capex at a repricing long end is a bet on someone else's continued willingness to lend you money at a rate that still makes the project's economics work — and that willingness is precisely the thing a hawkish, war-adjacent, 5.4%-PPI-printing macro backdrop is currently testing. Oracle already carries negative free cash flow near $23.7 billion off aggressive data-center buildout and reportedly needs to raise something like $40 billion more this fiscal year. Dell is out marketing roughly $4 billion in fresh investment-grade paper this week alone, ostensibly to refinance 2026 maturities, coincidentally right as AI server demand gives the credit story a better narrative than "we needed to roll debt." None of this is illegal, none of it is even unusual for a capital-intensive industry. It is, however, the exact mechanism that turns a valuation correction into a credit event, and credit events don't ask politely whether your GPUs are still the best ones on the market.

Nobody is calling this a bubble on the record, and to be fair, calling it a bubble is premature — the revenue growth underneath it is real, unlike 1999's. What's not premature is noting that the financing structure of the AI buildout quietly changed character sometime in the last two quarters, from equity-and-cash-flow story to credit story, at the exact moment credit got more expensive to issue. The century bond isn't the risk. The century bond is the tell — the moment an institution with access to the best information in the building decided the smart move was to lock in today's terms for a hundred years, rather than bet on cheaper money later. Somebody, somewhere, already knows which way this goes.