Dark Fiber With a Power Bill

Run this framework’s only test and the answer arrives immediately. Does the thing show up when you offer money for it?

Offer money for chips and they arrive, in quantity, on a schedule. Offer money for money and it arrives in almost unlimited size: about $630 billion of capital spending this year against $388 billion last year, a jump of 62 percent, with an estimated $4.1 trillion of debt standing behind the wider build. Offer money for electricity and nothing happens at all.

Power is now the thing holding up construction, ahead of chips and far ahead of money. At the end of 2025 roughly 8,200 projects, representing about 1,312 gigawatts of generation, were sitting in line waiting to connect to the American grid. In some regions a project that depends on the grid faces up to ten years before it can be switched on.

That is what a real shortage looks like. Unlimited money meets a limited thing and what comes out the other end is a queue.

What has actually run out

Underneath the power problem sit four physical bottlenecks. Not one of them is about money, and not one of them gets shorter because more money shows up.

That last line is the one that connects this to everything else here. The president of Microsoft has said the single biggest thing slowing data center construction in America is a shortage of electrical talent. Not capital. Not permits. People who can pull cable and wire up switchgear, whose numbers are set by a training pipeline measured in years.

That is the hours ceiling, arriving in the richest industry on earth. A trillion dollars does not produce a journeyman.

And there is a detail here that ought to make everyone involved uncomfortable. The transformer bottleneck runs on a specialised electrical steel, and China holds roughly 60 percent of the world’s transformer manufacturing. So the build-out we describe as a race against China depends on Chinese equipment to switch it on, and the tariffs tax that equipment on the way in.

The scissors closing

Now hold the cost of building one of these next to the price of what it makes. This is where the question of whether any of it can last actually gets settled.

Something paid off over five and a half years, built with gear on multi-year waiting lists, is producing a product whose price has fallen 95 percent in two. That is not a demand problem. Demand for this stuff is enormous and growing. It is a price problem, and the two are easy to confuse, because the usage charts look wonderful right up until you notice the revenue per unit collapsing underneath them.

The accounting has already started bending. Meta stretched the assumed lifespan of its servers from four years to five and a half, which cut its 2025 depreciation charge by $2.9 billion, and every big player made a similar move between 2020 and 2025. The bear case, argued loudest by Michael Burry, is that a high-end chip really lasts two or three years and that these extensions understate the cost by something like $176 billion across 2026 to 2028.

This framework has no view on who is right. It has a view on what the argument is about, which is that the lifespan of the asset is being estimated by the people who need it to be long.

China is not trying to win

China’s build-out usually gets described as a race, and on the surface it looks like one. A drafted plan worth about $295 billion over five years for a national computing grid, running on roughly 80 percent home-made chips, shutting out Nvidia and AMD by design. National computing capacity was near 215 exaflops in 2025 and is heading past 330 by the end of 2026, growth of about 40 percent a year. Huawei shipped some 812,000 of its own accelerators in 2025.

Through this framework, that is the less interesting half. Capacity is the physical layer. What China is actually going after is the money layer, and the treatise says the money layer is the one that comes first.

The weapon is price. The gap between free models and paid ones narrowed from 8 percent to 1.7 percent in a single year. A Huawei-backed model now sells its output at eleven cents per million words. DeepSeek is at twenty-seven cents, Google’s cheapest tier at twenty-five, and OpenAI has cut its own prices by as much as 80 percent in response.

A competitor funded by a government does not need to earn a return. And a competitor that does not need to earn a return sets the price for every competitor that does.

It is the same move this framework describes everywhere else, run backwards and on purpose. The American build-out is an enormous bet on physical stuff, paid for with borrowed money. The Chinese answer does not fight the physical stuff at all. It devalues the borrowed money, by collapsing the price of what the physical stuff makes.

Who is left standing

It will overbuild. The people building it say so out loud. Satya Nadella has said plainly that there will be an overbuild, and that prices will fall further when the correction comes.

Meanwhile every one of these companies reports being short of capacity rather than short of customers. And three of the four lost market value after their last earnings calls. Both things are true at once. That is the ordinary condition of a building boom near its end: physically short today, financially long tomorrow.

This framework’s answer to who survives is not the usual vague thing about strong companies. It is specific, and it falls out of the ordering. Whoever owns the thing that ran out survives. And the thing that ran out is not computing.

A company that raised equity and holds signed power has an option. A company that funded itself through the circular arrangements now estimated at more than $800 billion, where the seller lends its own customer the money, has an obligation. When the price of the product falls to the cost of making it, the first can wait and the second cannot. And that difference has nothing to do with which of them has the better model.

What is left when the money stops

Everybody reaches for the fiber-optic comparison. It is half right. Which makes it worse than useless if you take the wrong half.

Here is the right half. When the telecom boom ended, the physical stuff outlived the money completely. Shareholders in the companies that laid the cable were wiped out. The glass stayed in the ground. Years later somebody who had not paid a cent for it bought the capacity for pennies, switched it on, and made a fortune. The fake part was in the financing, and the financing is what got destroyed. The real layer was fine.

That is the treatise’s distinction made concrete, and it is the most likely shape of an AI correction too. A transfer of wealth from the people who paid for the capacity to whoever owns it second.

Now the wrong half, which is where the title comes from. Unlit fiber cost almost nothing to hold. It sat in the ground for a decade with no meter running, and a decade later it was worth a great deal, because glass does not age.

A data center cannot do that. It has a power contract running whether or not the machines are earning. It has cooling. It has maintenance. And it has silicon aging on a clock measured in a few years rather than in decades.

Idle fiber waits. Idle computing rots.

So the thing that survives is not the computer. It is the hole in the ground with a grid connection attached. And in a market where getting connected can take ten years, the connection is worth more than everything plugged into it.

Which gives a testable prediction. In a correction, the chips get written down hard and fast, because they do not last long and the next ones are better. The building, the substation, the grid connection and the water rights do not get written down at all in real terms, and may well go up, because they are the genuinely scarce thing and nobody can make them faster by spending money. Expect distressed sales priced by the megawatt rather than by the rack.

Who is paying for it right now

The investors are not the only people paying for this. Here is where it stops being a story about markets, and becomes a story about the Toilet Paper Theory.

In the PJM power market, the price of having capacity available went from $28.92 per megawatt-day for 2024 to 2025, to $329.17 for 2026 to 2027. Data centers were responsible for about 63 percent of the increase in one auction, some $9.3 billion of it. That flows straight into ordinary bills: roughly $21 a month more for Pepco customers, about $18 in western Maryland, about $16 in Ohio, and a projection near $70 a month for an average family in the region by 2028.

Electricity is the cleanest example of a necessity you cannot buy less of. A family cannot decide to use meaningfully less without giving up heat or cooling. So the price does not destroy demand the way a tariff on shirts does. It simply rises until it has taken whatever the family had. The poorest tenth pays the charge and owns none of the shares, and the transfer runs through a regulated rate sheet nobody voted on.

Three real ones in a row, which should worry me

This framework spent its first year finding fake shortages nearly everywhere it looked. It has now found three genuine ones back to back: the Strait, the munitions, and the grid. Either the world changed or the person reading it did, and honesty means naming the second possibility as well as the first.

The defense of the first reading is that all three share something the earlier cases did not. Each one is unlimited money meeting a physical process with a long, stubborn cycle time: a shipping lane, an explosives plant, a transformer works. That is a specific and checkable condition rather than a mood. But the framework should be uncomfortable about it, and the discomfort belongs in the scoring table rather than in a footnote.

One more thing, because it limits what this piece is allowed to say. The gauge’s forward reading found that a serious data center credit blow-up, on its own, produces a reading of about plus 0.90 against a threshold of plus 1.0. It falls short. An AI bust by itself is not a credit crisis by this framework’s own instrument. It needs company, and nothing here should be read as predicting any.

What would prove this wrong

Two warnings. The 2026 spending figures are guidance, not results. Published estimates run from about $630 billion for the biggest four firms up to roughly $750 billion for the fourteen largest operators. Treat the exact number as a range.

The second warning is bigger. The claim that idle computing rots faster than idle fiber rests entirely on how fast chips go obsolete. That is the most contested number in this whole subject, and it is the one the accounting argument turns on. If a high-end chip holds its value for eight years rather than three, most of the urgency here dissolves.

What does not depend on that number is the order of things. The money is everywhere. The power is not. And no amount of the first produces the second inside the time the financing assumes. The build-out will get further than its critics expect, because a bid this size does move a great deal of the physical world. And it will arrive later and cost more than its promoters expect, because the last mile of it runs through a transformer factory and an apprenticeship. When it corrects, the money will be destroyed and the concrete will not, and the people who end up owning the concrete will mostly not be the people who paid for it. That is the oldest story in this framework. The only unusual thing here is that everyone can see it coming and is building anyway.