Two weeks ago we said the design was solved but the factory couldn't keep up. Last week, it was the chip's memory. This week, follow the money — because the biggest earnings stretch of the year is under way, and the interesting story isn't the number everyone reacted to.

When the seller backs the buyer.

On Wednesday 29 July (US time) Microsoft and Meta reported. Microsoft did $90.0 billion in a single quarter and $331.8 billion for the full year, and its cloud business, Azure, didn't just grow — it sped up, to 43%. Azure brought in more than $100 billion of revenue over the 2026 financial year. That's real money from real customers, and it says cloud and AI-platform demand is strong. On the surface, the AI trade is working.

Now look at what it costs to keep that going.

Microsoft reported about $41 billion of capital spending in that one quarter — though that figure includes finance leases; the cash it actually paid out was closer to $36 billion, and not all of it is "AI gear" (a big chunk is short-lived chips, the rest longer-lived buildings and kit). Either way, it's enormous — for a single quarter. And across the five large US hyperscalers the Bank for International Settlements is watching — Alphabet, Amazon, Meta, Microsoft and Oracle — it says they're on track to spend more than $1 trillion across 2025 and 2026: more than those companies earn, and more than the spare cash their businesses generate.

Spend more than you make, and the gap gets filled from somewhere. Increasingly, that somewhere is borrowing — bonds, leases and private credit.

Now the part that made me stop.

On the weekend of 26–27 July the Wall Street Journal reported that Nvidia — the company that sells the chips — is in talks to guarantee around $250 billion of financing to help build a giant OpenAI data centre in Ohio. The detail matters: the $250 billion is credit support for the data centre's lease and construction, not for the chips. OpenAI doesn't hold a top-tier credit rating, so a Nvidia backstop would make lenders more comfortable and lower the cost of the loan.

None of it is signed, and the talks could fall over. But sit with the shape: the company selling the chips is offering to stand behind the borrowing of the company buying them. When the seller backs the buyer like that, it raises what's called circular-financing concern — because demand can start to look bigger and healthier than it really is when the vendor is helping fund it.

Watch who reacted how.

The market's response to the earnings was split: Microsoft jumped about 9%, while Meta fell about 8%. And the people who watch debt are getting warier. The same BIS report likened the scale and pace of today's AI spending to past investment booms — the dot-com era among them. Bank of America's July survey of large fund managers ranked AI spending as the leading perceived trigger for the next big credit event, ahead of everything else. And it's showing in prices: recently issued bonds from Nvidia and SpaceX traded weakly, and even Amazon paid more than its usual rate to borrow. Credit investors aren't fleeing — but they're getting pickier, and charging more for the risk.

Be fair about the other side.

The strongest argument against all of this is the one Microsoft just made: unlike the speculative, revenue-light names of the dot-com era, today's big hyperscalers already generate huge revenue and profit. Azure didn't stall, it accelerated, and management guided it higher. The hyperscalers are large and profitable — that's real, and it's the part that could prove us wrong. (Worth noting the other end of the chain is not: OpenAI itself is still lossmaking.)

So what?

The point isn't "it's a bubble about to pop." It's quieter and more useful: the AI buildout has changed how it's paid for. These companies historically funded most of it from their own cash flow. Now they increasingly lean on borrowing, and on deals where the seller helps stand behind the buyer. That doesn't make it doomed — it makes it more fragile, more dependent on the returns actually showing up and on lenders staying comfortable enough to keep the money flowing. If AI revenue keeps accelerating the way Azure just did, none of this bites. If it merely slows, the backstop-and-borrow structures are the first thing to wobble.

Same lesson as our first two issues, moved one more step. First the question was whether we could build the thing. Then whether we could make the chips. Now it's who's holding the debt when the bill comes due. The bottleneck keeps moving — and the market keeps pricing the worry that just passed, not the one arriving. Where the exposure sits, as a map and not a recommendation: not one company, but the whole AI-infrastructure complex — the hyperscalers spending it; Nvidia, if the reported backstop proceeds; and the private-credit funds quietly financing the rest.

OUR CONVICTION: 58 / 100  ·  as of 30 Jul 2026, strengthening

Forwarded this? Subscribe free — one structural read a week.

Structural intelligence, not financial advice. Conviction scores reflect our own structural assessment — not price targets — and are not a recommendation to buy or sell any security. Companies named are described to illustrate a structural shift, not ranked as picks. The author may hold positions in securities or sectors discussed.