The AI Stack Has Four Balance Sheets. One of Them Is Borrowing.

The neoclouds have enormous order books and the market keeps asking who is paying for all that compute. The answer is in their own disclosures, and it loops straight back to the companies that were supposed to be funding this themselves.

Every conversation about the artificial-intelligence buildout eventually reaches the same question: who is actually paying for it? The order books at the companies that rent out computing capacity are enormous, and a rented data centre is only as good as the tenant's cheque. So it is worth sorting the industry by the thing that actually determines credit quality — who bears the capital spending, and who owns the demand — rather than by what each company says it does. That sort produces four layers, not one. Three of them are familiar. The fourth is the one carrying the debt, and its largest customer turns out to be the layer above it. Once you see that, the whole structure stops looking like a new industry and starts looking like a financing arrangement.

Four layers, sorted by who carries the capex

Layer one: owner-operators with a cash engine. Microsoft, Amazon Web Services, Google. They fund capital spending out of operating cash flow, they own the buildings, and the demand is part internal, part rented out. Their risk is that depreciation outruns the revenue it is meant to buy.

Layer two: owner-operators with a cash engine elsewhere. Meta, Tesla, xAI, SpaceX. Same funding source, different motive — the compute is largely for themselves. Meta belongs here rather than with the pure labs because the advertising business pays for the build.

Layer three: the labs. Anthropic, OpenAI. No balance sheet worth the name, no data centres of their own, funded by equity and by prepaid compute. Their product is the model.

Layer four: the landlords. CoreWeave, Nscale, Lambda, Crusoe. Levered, asset-heavy, contracted. This layer exists for one reason: renting capacity is faster than building it, and the debt lands on somebody else's balance sheet.

The sort is not cosmetic. Each layer has a different cost of money, and the cost of money is what decides whether the buildout survives a higher-for-longer curve.

Who is actually in the order book

CoreWeave is the cleanest read because it is public. As reported for the first quarter of 2026, Microsoft accounted for roughly 62% of its revenue, down from about 68% a year earlier. OpenAI accounts for something close to a third of its contracted revenue over the next seven years, across three agreements that sum to roughly $22bn. The diversification list around them includes Meta, Anthropic, Mistral, Cohere, IBM and Jane Street.

Those figures come from company reporting and press coverage rather than from a filing we have read line by line, and they should be treated as reported rather than audited. But the shape is not in doubt, and the shape is the point.

The largest customer of the independent compute layer is a hyperscaler. Microsoft is not competing with CoreWeave; it is renting from it, the way it rents from anyone with a shell, a substation and a grid connection. The 'independent' layer is substantially a financing structure for the same four balance sheets that were supposed to be paying for this directly.

That is not a scandal and it is not a fraud. It is how capital-intensive buildouts have always been financed: the entity that can borrow cheapest builds, and the entity that needs capacity now signs a lease. But it does mean the phrase 'independent AI infrastructure' is doing more marketing work than analytical work.

What each layer pays for money

This is where the four-layer sort earns its keep. On 21 October the investment-grade index yielded 5.92% and the high-yield index 8.17%, with a high-yield spread of 327 basis points over Treasuries. Layer one borrows near the first number. Layer four is being described in the market at something like 15% on its paper — a figure reported in a 22 October interview rather than an index print, so treat it as directional, but the direction is not in dispute.

That is the entire investment case for the stack in one picture. The identical dollar of capital spending costs roughly 5.9% at the top, 8.2% one rung down, and around 15% at the levered layer that actually builds the capacity. Nothing about the demand curve changes that arithmetic.

Why the landlord pays 15%

It is not because the borrower is new, and it is not because the paper is private. It is the collateral. A graphics processor has a useful economic life of three to six years, and the generation behind it arrives faster than the depreciation schedule assumes. The loan against it runs longer than the asset does — a five-year asset against a ten-year liability is a maturity mismatch with a countdown timer on it.

Add concentration: when one customer is worth roughly 62% of revenue, the credit is really a bet on that customer's willingness to keep paying after the initial term. Add the newness of the contracts: nobody has been through a renewal cycle in this industry, so there is no history to price against. The 15% is what the market charges for being early on all three.

Which makes the contract terms the only thing that matters. If the labs prepay, the landlord is a utility earning a spread on somebody else's demand. If they pay monthly and can walk, the landlord is a levered bet on a tenant's balance sheet, holding a depreciating asset against a fixed obligation. That distinction is worth more than any rating.

Why this is a macro story, not a tech story

Three things follow, and none of them are about software.

First, the marginal AI dollar is expensive. An equity that has to clear a 15% financing hurdle is a different proposition from one that clears 5.9%, and the difference is not in the narrative — it is in the discount rate.

Second, the risk is concentrated where the capital is thinnest. The lab carries the cash burn, the landlord carries the debt, and the hyperscaler keeps the option to walk at renewal. In a downturn, the loss lands on layer four first, and layer four is the one whose paper is held by insurance companies and credit funds.

Third, this competes with the sovereign for the same buyers. If a levered data-centre borrower is paying 15% while the ten-year Treasury pays 5.15%, then every dollar of credit-fund capital that goes into that lease is a dollar not buying duration. That is crowding out, and it is measurable.

What would falsify this

The claim is not that the AI buildout is a bubble. It is narrower: that the stack is four balance sheets with wildly different funding costs, that the levered layer's customers are substantially the layers above it, and that the credit risk sits where the capital is thinnest.

Falsifier one: the neoclouds sign genuinely take-or-pay contracts with prepayment, and their spreads compress toward high yield while their customer concentration falls. That would mean the layer is maturing into a utility, and the 15% was an early-stage premium rather than a structural one.

Falsifier two: Microsoft's share of CoreWeave revenue keeps falling below roughly 40% without any widening in CoreWeave spreads. That would prove the independent layer has genuine third-party demand, rather than being a financing arm.

Until one of those happens, treat the order book as a promise and the spread as the price of believing it.