Chasing the Singularity — on Credit

"Whether I shall turn out to be the hero of my own life, or whether that station will be held by anybody else, these pages must show."

Charles Dickens, David Copperfield

One assumption investors stop questioning, stress-tested.

01The assumption

The AI infrastructure build-out is self-funding. It is paid for out of the largest cash flows in corporate history — so whatever else it may be, it cannot be a debt bubble.

And beneath it, the assumption that really matters: this is the single most-used argument for why this cycle differs from the last great infrastructure mania. The telecom fibre build of the late 1990s was speculative and debt-financed; this one is financed by profits.

We'll leave the payoff question for another entry. This one is about the bill.

02Why it's believed

Because it was true.

In 2023 and 2024 the argument was simply correct. The largest technology companies generated combined net income approaching half a trillion dollars a year, and their AI spending sat comfortably inside operating cash flow. Balance sheets were among the strongest in corporate history. Investors reaching for the dot-com comparison were fairly answered: show me the debt. There wasn't much to show.

The assumption earned its place honestly. The trouble is what assumptions do once earned — they stop being checked.

03The stress test — run the funding identity

So run the check today: does the money still come from inside?

Hyperscaler capital expenditure in 2026 is a competitive race to the top. More is better. Until it isn't.

Current guidance runs around $700 billion across the five largest spenders — the precise figure depends on the day and the estimate — up by roughly two-thirds in a year. But the sharper measure is intensity: somewhere between 45 and 57 cents of every revenue dollar is now going into the build. Ratios like that belong to utilities and telecoms, which need infrastructure. AI needs it too. But the depreciation cycle on a warehouse full of chips? Oh boy.

Internal funding no longer covers the bill. Alphabet is guiding $175–185 billion of capex against roughly $73 billion of trailing free cash flow — a gap only external money can close. Across the group, current plans imply negative free cash flow. And the external money is arriving: over $100 billion of debt raised in 2025, projections of some $1.5 trillion of issuance over three years, and the four largest names' weight in the investment-grade bond index roughly doubling inside a year.

Below the headline debt sits a quieter layer — leases in place of owned data centres, joint-venture vehicles, asset-backed structures, private credit, financing secured on the chips themselves. Capital that funds the build without announcing itself as leverage. Complexity in financial structure is a signal in its own right — we have written on that elsewhere.

And there is a stranger feature still, observed by Jeremy Grantham among others: the great monopolies are spending their monopoly profits to fund a competition — with each other — in a market that may crown only one or two winners. Sharp-elbowed rivals outspending one another toward AI glory. Competitive spending has no natural brake; no one can stop first. Which is exactly why the funding source matters: spending that cannot self-moderate runs until the funding stops it.

Then there is the circularity. Vendors taking equity stakes in their customers. Compute commitments between counterparties who are simultaneously each other's suppliers, customers and investors — Oracle's contracted future revenue now runs at nine times trailing revenue, concentrated on a single counterparty, while the market's price for insuring its debt has trebled in months. Students of the last cycle will recognise the shape: in 1999 and 2000, the defining tell was equipment makers financing their own customers' purchases.

Demand that a cycle funds for itself is not demand. It is the cycle, counted twice.

Cognitive dissonance is a standing feature of financial markets. It can persist for years. Counting on its persistence, though, is one more assumption going unexamined.

The mechanism is what matters, and it is simple. Self-funded spending answers to the business: it continues as long as management believes in it. Externally funded spending answers to the credit market: it continues as long as capital markets remain willing — and that willingness is pro-cyclical, most abundant at the top and scarcest exactly when the spending most needs continuity. The moment the marginal dollar of capex is borrowed, the build-out's continuation is conditional on credit conditions. The assumption doesn't require AI to fail in order to fail itself. It only requires the funding source to have changed while the story didn't.

It has changed. The story hasn't.

Or perhaps we are being too curmudgeonly about all this. Micawber was in misery over a trifling sum. That is not the modern way.

04The threshold condition

The test is the funding identity, run company by company: operating cash flow, minus capital expenditure including leases and off-balance-sheet commitments. While that number is positive, "self-funding" holds and the dot-com comparison genuinely fails. When it crosses zero, the spending has become market-dependent — whatever the press release says. For the group in aggregate, that line is being crossed now. The assumption is living on the memory of a funding regime that has already ended.

What no one can specify in advance is how much external funding the credit market will extend, or for how long — that depends on the revenue the spending eventually produces, and the multiple of revenue-to-capital this build-out requires has not yet been demonstrated at scale.

The old adage: refinance when the market lets you, not when you need to. There is a corollary assumption that these companies will always have optionality. And they do — until the regime changes. Regimes change.

05What to watch

Free cash flow after full capex, per company, each quarter. Sign and direction. This is the assumption's health reading, published four times a year.

The credit market's price for the build-out. Spreads and default insurance on the most levered participants — the market's most honest opinion of the funding structure, and historically an earlier signal than equities. Watch also the hyperscalers' rising concentration in the bond index: the build-out is becoming a fixed-income event.

Circular financing announcements. Vendor stakes in customers, compute deals between mutually dependent counterparties, financing secured on the equipment being sold. Each one shifts demand from external to self-generated — the tell that marked the last cycle's top.

06One more thing

None of this is a claim that AI's economics fail. The technology can succeed completely and the financing structure still set the cycle's timing — in 2001 the internet was real, and the fibre-funding structure collapsed anyway. The asset and the funding of the asset are different things, and the assumption under stress here concerns only the second. The build-out was self-funding. The tense of that sentence is the entire point.

As for who turns out to be the hero of this story — the pages must show. They are published quarterly.