01The demand is real and nobody serious disputes it. The way it is being financed is a different question entirely.
Compute is scarce, the demand is durable, and the buyers are the richest companies on earth. All of that is correct. The revenue is real money from real customers, and anyone arguing otherwise is not worth listening to.
Which is exactly why there is no money in that debate. Being early on a view the entire market already holds pays nothing at all.
The real exposure sits in the gap between when the money goes out and when it comes back. A data centre is concrete, power, land and hardware, paid for years ahead of the revenue it supports. Something has to bridge that gap, and the bridge sits in three places the income statement does not touch.
First, the commitment book. Multi year agreements to buy capacity, chips and power are fixed claims on future cash whether or not the revenue appears. On several of the largest names that total runs ahead of the capital expenditure figure the street quotes.
Second, the identity of the lender. Prepayments, vendor financing, equity stakes taken in customers and supply agreements signed with those same customers build a loop where one company's revenue is another company's obligation. None of that is improper. It is correlated, and the revenue and the credit risk now move together in a way most models still treat as independent.
Third, the calendar. Capacity is built with debt that has a date on it. Refinancing is routine until the week it is not, and what makes it hard is almost never anything to do with whether the technology works.
None of this is improper and none of it is hidden. It is drawn because the same dollar is doing several jobs at once, and most models treat these four as independent counterparties.
The received view is that the model companies carry no leverage. No bonds, no bank syndicate, no covenants. On a conventional screen they look like the least geared participants in the entire complex.
That is true and it is beside the point. Leverage is a fixed claim on future cash, and a multi year commitment to buy compute is exactly that. Sign an agreement to take capacity for five years and you have taken on an obligation that behaves like debt in every way that matters: it is large, it is contractual, it does not care whether revenue arrives on schedule, and it has to be funded.
Put that next to the burn. These are businesses spending far more than they earn, by design, in a race where slowing down is the only unrecoverable mistake. Fixed forward obligations plus structural cash burn is a leverage profile. It simply is not called one, because no instrument was issued.
And it is invisible. A private company files nothing. The commitments are known only through announcements, which means the scale is reported rather than disclosed and the funding behind it is not public at all. The chart below gives every company that files a bar. The two that do not file get a void, and the void is the finding.
The control case. Commitments are enormous in absolute terms and still comfortably inside a year of cash generation.
Same shape as Microsoft. The build is being funded out of the business rather than out of the credit market.
Promised several times what the business earns in a year. The distance between those two bars has to be funded, and that funding is debt.
The pure expression of the trade. Contracted capacity, contracted debt, and a customer list short enough to name.
Multi year compute agreements have been announced at a scale that would dominate this chart. None of it is filed, and the cash generation behind it is not public.
Same structure, same invisibility. Large announced compute arrangements, no periodic reporting, no public view of the funding behind them.
Figures as last reported. Verify against the current filing before relying on any of it.
The build cannot be funded from the equity market alone, so credit has arrived, and it has arrived in a form the market has not priced because most of it is not public.
Four channels matter. Syndicated bank facilities secured against data centre assets and the contracts attached to them. Lending collateralised on the accelerators themselves. Private credit written directly to operators and project vehicles. And vendor financing, where the supplier funds the customer who is buying from the supplier.
The second of those deserves the most attention, because the collateral and the risk are the same object. A loan secured on hardware assumes a resale market for that hardware, and the useful life of an accelerator is the open question in theme seven on this board. If the replacement cycle turns out shorter than assumed, the asset backing the loan and the asset being written down are the same asset.
Pricing on almost all of this is privately negotiated and does not appear in any public document. Anyone quoting a precise rate on it is guessing. What is knowable, and what matters more, is what each structure is secured against and who is left holding it.
The pricing column is mostly empty on purpose. This paper is privately placed and the terms are negotiated, so an honest cell is a blank one. What the structure is secured against is public enough, and it matters more.
Large banks have moved into lending against compute infrastructure and the contracts attached to it. Terms are privately negotiated and the pricing does not appear in any public document.
The structure that deserves the most scrutiny. It assumes a resale market for hardware whose useful life is the open question in theme seven, and the collateral and the obsolescence risk are the same asset.
Where the first theme meets the third. Paper written at today's spreads, marked by the manager, and not yet tested by a realisation.
The least visible of the four and the one that closes the loop, because the lender and the counterparty are the same company.
Because none of it is hidden. The commitments are disclosed, the maturities are scheduled and the related party arrangements are set out in plain language. Disclosure is not attention, and none of it sits in the headline number.
This reprices on an ordinary quarter, not a scandal. A capital expenditure guide that rises while the free cash flow guide falls. A financing round clearing at a visibly worse price than the last. One large customer reopening a term everyone had modelled as fixed.
A crash is not required for the trade to work. All that is required is for the market to move from pricing the demand to pricing the financing, and those are different multiples on identical revenue.
Composites of the conversation around this trade, including the parts that argue against it. Nothing here is a quotation and nothing is attributed to a person or a firm.
The case against this theme“Nobody is modelling what happens when the labs decide to charge what the product is worth.”
The bull answer to the burn. Enterprise pricing has been set to win accounts rather than to cover cost, and there is a large amount of unexercised pricing power sitting in these businesses. A single pricing change closes a lot of a funding gap.
The case against this theme“The forward deployed engineers are the moat. Once they are in, the account never leaves.”
The most substantive bull point. Labs are putting their own engineers inside customers to build the first real workflows, which converts a model subscription into an operational dependency. Switching cost stops being about the model and becomes about the integration, and that is the kind of stickiness that supports both the price and the term of a contract.
The case for it“Every one of these contracts has an out, and the outs have not been read.”
The bear reply. Multi year commitments carry termination, delay and ramp provisions that are rarely disclosed in detail, which cuts both ways: it limits the downside for the party that signed, and it weakens the backlog the other party is being valued on.
The case for it“It is not a bubble until the lender blinks, and the lender is not even in the model yet.”
Where the theme actually sits. The argument is not about demand and never was. It is that the market is pricing an income statement while the risk has moved onto balance sheets, some of which are not public.
Purchase obligations represent agreements to purchase goods or services that are enforceable and legally binding, specifying all significant terms including quantity and price.
Boilerplate that appears across the whole complex, and the sentence is not the point. The total underneath it is: how much of the next three years is already committed before a single new customer signs anything.
We expect to fund our capital expenditures through a combination of cash on hand, cash generated from operations, and additional financing.
Additional financing is the clause the whole short book is built on. It is a company stating that the plan needs money it does not yet have, on terms that do not yet exist. Harmless in a friendly market, and the entire position in an unfriendly one.
Its own balance sheet is not in question. What matters is how much of its demand is funded by customers whose solvency runs back through it.
Financing is not a footnote here, it is the business model. Obligations, maturities and customer concentration are the investment case.
Buying share in a capital hungry market on a balance sheet already working hard. The number that matters is the gap between commitment and cash generation.
Carried as the name that should break the thesis. Cash generation at this scale absorbs an enormous amount of forward commitment.
Asked of every desk before the theme was published, not after it failed.
