The Vendor’s Loan: How a Buildout Tells You It Is Ending

The salvage thesis is easy to agree with in retrospect and nearly impossible to act on in the present. Everyone can see now that dark fiber carried YouTube. Almost nobody was buying dark fiber in 2002.

The reason is not nerve. It is that the thesis, stated generally, gives you nothing to do. Buy the overbuild at salvage cost is not a position. It is a slogan with two holes in it. Which part of the overbuild, and when. Get the first wrong and you are the man who bought a warehouse of telecom switching gear in 2003. Get the second wrong and you are correct six years early, paying carry the whole way.

This issue fills the holes.

What actually made the fiber worth buying

Fiber was not salvageable because it was cheap. Plenty of things were cheap in 2002. It was salvageable because it held five properties at once, and the five are the sorting test for everything that follows.

It did not decay. Glass in a conduit has a twenty-five year physical life and no meaningful wear mechanism. The asset a buyer acquired in 2004 was the same asset the builder had installed in 1999.

It did not care. Fiber is indifferent to what runs over it. Voice, video, packets, protocols that had not been written yet. Nothing about the medium expressed an opinion about the application.

It cost nothing to hold. Unlit fiber has a marginal carrying cost near zero. You could own it for a decade while you waited for the demand curve to arrive underneath you, and the waiting cost you almost nothing.

It was severed from its cost basis. This is the part that requires a bankruptcy. Orderly declines do not produce ten cent assets, because an orderly seller holds out for something closer to what he paid. Global Crossing and Williams and 360networks did not hold out. The receiver sold, and the price had no relationship to the twenty billion that went into the ground.

And it was the binding constraint. Cheap bandwidth did not give Google its idea. It changed the arithmetic under a model that had been uneconomic. Serving the world for free and settling up through an auction on attention only works when transport is nearly free, and transport became nearly free because somebody else had already paid for it and gone under.

Hold those five. They are what separates an asset from a write-off.

The accelerator fails the test

The consensus answer to what the dark fiber of this cycle will be is the accelerator. Warehouses of stranded silicon, bought at auction from the wreckage, powering the next thing.

Run it through the test.

It decays. Economic life on an accelerator is three to five years against fiber’s twenty-five, and the argument currently running through hyperscaler accounting about whether useful life is five years or six is not a footnote dispute. It is the whole question of whether the fleet is an asset or an expense that has been deferred.

It costs to hold. This is the property everyone skips, and it inverts the entire trade. Unlit fiber costs nothing. A dark accelerator costs power, cooling, floor space, and failure replacement whether or not it computes. The waiting is not free. You cannot buy the stranded fleet and sit on it for eight years until the demand shows up, because the carry will kill you before the demand arrives.

And it half cares. The fleet is tuned to the numeric formats, memory bandwidth ratios, and attention shapes of the current architecture. A shift in that architecture does not make the fleet old. It makes it mismatched, which is worse, because old hardware still runs the same workload slowly and mismatched hardware runs the new workload badly at any speed.

The accelerator is not the conduit. It is the switching gear at the end of the conduit. The conduit made a decade of companies. The switching gear went to scrap, and the company that manufactured it never recovered its position.

What passes

Power, and specifically the paperwork. The United States interconnection queue holds roughly 2,600 gigawatts against a median wait approaching five years, and of everything that entered that queue between 2000 and 2019, nineteen percent had reached commercial operation by the end of 2024. Developers are routing around it rather than waiting in it, with something on the order of 101 gigawatts of behind-the-meter gas generation announced specifically to bypass the bottleneck.

Read what the scarce good actually is there. Not the turbine. The entitlement. The queue position, the powered land, the interconnection agreement, the transformer slot, the water rights, the local approval that took three years and a lawsuit. Thirty to forty year lives, complete indifference to workload, and a scarcity that deepens rather than erodes. A queue position acquired out of a distressed campus in 2031 is worth more than the same position is worth today, because the queue only gets longer and the study process only gets slower. That is a fiber-shaped asset with an appreciating scarcity attached, which fiber never had.

Shells and thermal capacity. Accelerator lifecycles run around seven years against facility lifespans of twenty to thirty. That mismatch is not a risk to be managed. It is the seam along which the salvage happens. The pad, the substation, the switchgear, the liquid cooling loop sized for hundred kilowatt racks. The building outlives four generations of what sits inside it.

The corpora of the labs that fail. When a well funded model company goes under, the headline will be the hardware auction, and the hardware will be dumped into a market already glutted with the same hardware. The asset is the thing nobody photographs. Preference datasets. Evaluation harnesses. Domain labeled corpora that cost nine figures and years of human hours to produce. Indefinite shelf life, no opinion about architecture, and fully reusable by whatever comes next. There will be an unresolved provenance question attached to how much of it was sourced, and that question is not an obstacle to the trade. It is the mechanism that produces the discount. Assets with clean title do not sell at ten cents.

People. The price of an engineer who has actually trained at frontier scale is a function of the cycle, not of the engineer. Google staffed its infrastructure on operators who had been expensive in 1999.

The tell

Now the harder question. When.

Valuation will not tell you. Valuation was wrong for four consecutive years in the last cycle and wrong in both directions, and anyone who traded on it either sold in 1997 or bought in 2000. What tells you is the financing structure, and inside the financing structure there is one pattern that has preceded every infrastructure overbuild worth salvaging.

Union Pacific’s construction was executed by a company its own directors controlled and paid substantially in the road’s own securities. The builder, the lender, and the shareholder were not three parties. The line got built. The line went into receivership.

Samuel Insull assembled the electrification of the American Midwest on holding companies stacked on holding companies, each layer capitalized by the layer above it, every dollar of that structure ultimately underwriting equipment purchases inside a sector Insull sat on both sides of. Sixty utilities at the peak. All of it gone by 1932. The generating plants kept running under new owners for another forty years.

Lucent and Nortel carried billions in receivables from competitive carriers they had lent the money to buy Lucent and Nortel equipment. The carriers took the loans and put fiber in the ground. The carriers went bankrupt. The fiber stayed. The switching gear did not, and neither did the balance sheets of the two companies that financed the whole arrangement.

Notice what survived each time. Never the vendor. Never the borrower. Always the physical asset with the long life and no opinion about who holds the title.

Now put today’s structure next to those. A neocloud takes a seven and a half billion dollar facility collateralized by its accelerators and its customer contracts, priced around eleven percent variable, with repayment beginning precisely as the collateral value turns down. The dominant chip supplier stands up a backstop program supporting multibillion dollar facilities for the companies that buy its chips. Aggregate capital spending at the largest five now runs above projected free cash flow, which means the marginal dollar of buildout is a borrowed dollar rather than an earned one. And the resulting paper is being packaged, with data center securitization projected at thirty to forty billion annually across 2026 and 2027, on the order of seven to ten percent of combined commercial mortgage and asset backed issuance.

Four facts, one shape. The supplier is underwriting the buyer. The collateral depreciates faster than the loan amortizes. And the exposure is being distributed into structured credit, which means the party that underwrote the risk will not be the party holding it when it resolves.

That is not a prediction of a bust. It is the mechanism by which a downturn, if one arrives, becomes disorderly enough to sever assets from their cost basis. A soft landing produces no salvage. Receiverships do.

Why the underwriter sees this before the allocator

Put the identical structure on a merchant application and watch how fast it gets declined.

A merchant reports three million in annual card volume. Pull the file. His largest customer shares an officer with him. His largest supplier is the same entity under a different registration. His working capital line comes from a party related to both. Every document in that file is authentic. Every number is real in the sense that the wires cleared. And the file does not cohere, because the revenue and the demand and the funding all resolve back to the same node.

An underwriter does not need a fraud finding to decline that, and looking for one is a misread of what is in front of him. The finding is structural. Volume that circulates inside a closed loop is not volume. It is one dollar, counted again at every turn.

That is what the current capital stack looks like from an underwriting seat rather than an allocating one. Not a fraud. A coherence failure, and a legible one.

The difference between the two seats is the question each is trained to ask. The allocator asks whether the demand is real, which is unanswerable in the present and will only be settled by the outcome. The underwriter asks whether the demand is independent of the financing. That one is answerable today, from documents, by anyone willing to trace the counterparties instead of reading the aggregate.

The position is not capital

The trap in the salvage thesis is that it reads like a capital play, and it is not one. Plenty of parties bought distressed fiber at ten cents and earned ordinary landlord returns on it. One of them built something else.

Google did not acquire fiber in order to own an asset. It acquired fiber because it already ran a business that consumed transport at enormous volume and was paying rent on every unit of it. The salvage was not the strategy. The salvage was an input price collapsing underneath a machine that was already turning.

Which means the pre-positioning question is not what will be cheap. It is this. What is uneconomic today only because inference costs what it costs?

The answer is a floor, and the floor is a price rather than a capability. Today you can justify a large model against a fifty thousand dollar credit file. You cannot justify one against a twelve dollar transaction, a single support ticket, a single claim line item, a single log entry, a single chargeback. Not because the model cannot do it. Because the arithmetic does not close.

Everything sitting under that floor is a business waiting for a number to move. And the party that captures it will not be whoever shows up at the auction with cash. It will be whoever already has the product built, validated, and starved of exactly the input that is about to become nearly free.

I have watched that floor from underneath for fifteen years. Underwriting is a cost per file problem before it is anything else, which is why the industry reviews at boarding and then stops looking. Continuous review is not a technical impossibility. It is a line item that does not currently pencil. The entire discipline is shaped around a constraint that most people in it have stopped recognizing as a constraint at all, because it has been true for the whole of their careers.

That is the same error the allocator makes about the buildout and the risk officer makes about the fraud. Treating the present shape of the thing as its permanent shape. The overbuild that looks like waste, the attack surface that looks closed, the price that looks fixed. One mistake, three rooms, and it has never once been earned.