The Salvage Thesis: Why the AI Boom Pays Its Builders Last

The trade everyone agrees on is selling shovels. In this buildout that means the compute layer: the accelerators, the hyperscale campuses that house them, the power and cooling being poured in behind them. And I want to give the logic its due before I break it, because the logic is sound. In a rush driven by one scarce input, the surest money is not the prospector who might hit the vein. It is the outfit selling every prospector the same tool at a margin, strike or no strike. The tool seller gets paid on activity, and activity is the one thing a mania guarantees.

The name people attach to that logic is Levi Strauss. The better name is Sam Brannan, who ran the store nearest the diggings, quietly cornered every pan and shovel in San Francisco, and only then went into the streets shouting that there was gold in the American River. He became, by most accounts, California's first millionaire while the men who did the panning mostly went home broke. All of that is true, and it is where the comfortable version of the story stops — one chapter early. Brannan died broke decades later, having mastered the entry and never the exit. Selling shovels is the right move for one phase of an infrastructure boom. Mistaking the phase for a position is how fortunes that looked permanent ended in receivership. The pattern has a second act. The second act pays better than the first, and it pays somebody else.

A shovel is a phase, not a position

The clearest tell is that the people who got richest selling shovels knew when to stop selling them. Carnegie supplied the steel the rails were made of and built one of the great American fortunes doing it. Then, in 1901, he sold the entire works to Morgan at the top and walked off into philanthropy. He did not hold the shovel through the cycle. He sold the shovel and the shovel business at the moment both were worth the most — a different act from owning them, and a much rarer one.

The investors who forgot the difference built themselves a monument. Its name is Cisco. Cisco was the Nvidia of 1999, the company selling picks for the internet gold rush, and for a few weeks in the spring of 2000 it was the most valuable company on earth. The internet it sold gear for went on to eat the economy exactly as promised. The stock peaked that spring and, a quarter century later, has never seen the price again. Sit with that. Everything the bulls believed about the technology came true, and the people who bought the shovel-seller at the top were still wrong, because a correct thesis about the tool is not a correct price for the toolmaker. Those are different bets. They come apart at the peak. Every time.

They come apart for reasons that are structural, not moral, and the structure is worth holding in your hands. Infrastructure at this scale is built against a forecast, and the forecast is always the current growth rate extrapolated to a horizon it cannot survive. Capital gets raised when growth is steepest and belief is strongest, which is precisely when the forecast sits furthest from what a mature market will absorb. So capacity gets sized to the peak of the belief, not the plateau of the demand. Then the build lags the decision by years. Foundries and substations do not arrive the day they are ordered; they arrive after the growth rate has already begun to bend, because growth rates bend before the people extrapolating them can admit it. The supply lands late, and it lands in a wave, timed almost perfectly to the moment the market needs it least. And underneath all of it, the asset changes character without moving. Scarce, it prices like a miracle and mints money. Abundant, the same physical object is a carrying cost. The steel is identical. The position is upside down.

The pattern has a second act

The railroads are the cleanest version, because the boom, the bust, and the payoff sit far enough back to see whole. Britain's railway mania in the 1840s and the American booms that followed were financed the way this one is, on speculative equity chasing a story that was real and overstated at the same time. Real, because rail remade the economy. Overstated in its timing, and in what it returned to the people who built first. When the panics came, in 1873 and again in 1893, the overbuild collapsed into the largest wave of corporate failure the country had seen. By the middle of the 1890s something like a quarter of American rail mileage sat in receivership. The equity was ash.

The steel in the ground did not move. Hold on to that, because it is the entire thesis in one image. The physical network the speculators financed outlived the speculators' claims on it and passed, at cents on the dollar, to whoever stood positioned to take it. Some of the prize went to the reorganizers, the men who assembled the modern system out of the bankrupt roads. The larger prize went one layer over, to operators who never laid a mile of track. Standard Oil did not get rich owning railroads; it got rich using its scale to squeeze secret rebates out of them, converting somebody else's overbuild into a private discount. Sears built a retail empire on cheap rail and rural free delivery, selling a nation its goods through a catalog that was only possible because the track was already there and nobody could charge much for it anymore. The fortune was in the retooling. The building was the tuition.

The nearest example is still warm

The version closest to now is close enough to touch. In the late 1990s the carriers laid fiber on the conviction that internet traffic would grow without a ceiling, and they laid far more than the era could use — the great majority of the strands they buried were never lit. Dark fiber, the industry called it. When the bubble broke, the builders broke with it: Global Crossing and WorldCom into bankruptcy, hundreds of billions of telecom capital gone, and the fiber itself selling out of receivership for a fraction of what it had cost to bury.

Then the second act ran on schedule. The cheap, abundant, already-buried bandwidth became the substrate of the following decade, and it made possible exactly the things that had been unthinkable while a bit was scarce and metered at boom prices. Streaming video was not a business when bandwidth was expensive; it became inevitable the moment the overbuild drove the marginal cost of a bit toward zero. YouTube in 2005. Netflix over the wire in 2007. The cloud, the whole second web — all of it rode on infrastructure that had bankrupted the firms that built it, while Google spent those years quietly buying dark fiber at distressed prices, which is a large part of why it could stand up a global backbone for so little. The bulls had told a lie that eventually came true. The traffic really did arrive to fill the pipe. It arrived a decade late, and it paid a completely different set of owners: the ones who bought the wreck, not the ones who built it.

Why this rhymes with rail, not the web

Here is where most of the current commentary runs the wrong comp, and the wrong comp decides who you think captures the value. The instinct is to reach for dot-com, because the last mania was the internet. But the internet was capital-light at the layer that mattered — serving one more user of a website cost roughly nothing — which is why its value pooled in software and asset-light aggregation. AI does not have that shape. Inference carries a real, recurring marginal cost in compute and energy, and the binding constraint on the whole buildout has already stopped being money and become power. That one fact drags the analogy off the internet and onto the railroad and the electrical grid: physical capital, real marginal cost, and value that accrues to whoever uses the commoditized infrastructure to do something new, not to whoever owns the infrastructure itself.

Electrification carries the other half of the lesson, which is the timeline. The dynamo was a finished technology decades before it showed up in the productivity statistics, because factories had to be physically torn down and rebuilt around electric drive before the gains could appear. That took the better part of forty years, and the technology being real did nothing to save the people who financed its first wave on a schedule the payoff could not meet. This is the reconciliation the bubble arguments keep missing. You do not have to believe AI is a fraud to believe the current capital structure is mispriced. The technology can be everything its advocates claim, and the equity funding this year's buildout can still be wiped, because the payoff diffuses on a timeline longer than the financing can survive. The tech is not the bubble. The timing is.

Now point it at the compute

So set the pattern down on what is being poured right now. The overbuild is the accelerator fleet, the hyperscale campuses, and the power and grid work behind them — the nuclear restarts, the interconnect queue — all of which only pencil out at the demand the forecasts assume. Combined hyperscaler capital spending is running well past half a trillion dollars a year against a revenue base that has not remotely caught up, a divergence now wider than at the peak of the 2001 telecom bubble, with a web of vendor and circular financing holding much of it together and power availability, not capital, deciding what actually gets built. The consensus trade prices all of this as though the current growth rate is the resting state of the world. No growth rate is.

And the saturation risk does not live where the bulls are looking for it. The market for intelligence is as large as anyone claims; that is not the problem. The problem is that the demand curve for intelligence is not the demand curve for this generation of silicon or this footprint of building. Efficiency compounds, and every gain in how much useful work a model wrings from a chip is deflationary for the precise asset being financed. A GPU is not a rail. Rails sat in the ground for a century; a training cluster tuned to this year's frontier is a fast-depreciating, purpose-narrow thing, and the revenue that has to justify it must arrive on the exact slope the forecast drew, or the asset strands. None of this requires the technology to disappoint. It only requires the growth rate to do what growth rates do.

Now run it forward. The phase change turns today's scarce, must-have, priced-for-scarcity compute into tomorrow's abundant, distressed, priced-for-nothing compute, and the question that decides the next decade is the dark-fiber question asked again. What is obviously too compute-hungry to be a business today, precisely because compute is scarce and metered, that becomes trivial the day a cluster that cost billions clears at cents and inference runs near free? Whoever is holding a real answer to that when the overbuild breaks is standing where Google stood in 2003. Not selling the shovel. Buying the wreck, and lighting it for a use the builders could never have served at their cost basis.

It will not arrive as a clean pop; the likelier shape is rolling dispersion. The leveraged pure-plays and the circular-financing web break first and hardest. The hyperscalers with fortress balance sheets survive the drawdown and use it to consolidate the distressed assets around them — the same move Morgan ran on the railroads, the same move the survivors always run. The power and grid buildout probably outlives the specific thesis that summoned it and becomes durable substrate for electrification at large: the one shovel that keeps paying after the rush it was dug for has cooled. And the durable position through the whole cycle was never ownership of the shovel through the peak. It is the standing capacity to acquire and retool on the far side of it. Those are different businesses. They reward different temperaments, and the second one is lonelier, because it asks you to stay liquid and patient at the exact moment the tape is punishing patience.

I am not going to hand you the month the shovels stop paying. Anyone who names that number is selling something. The honest output is a mechanism and its tells. Watch the spread between committed capital spending and realized revenue. Watch efficiency gains outrun demand for a given class of hardware. Watch for the moment the second derivative of the growth rate turns over while the forecasts are still bending it upward. That is the phase change beginning, and it always begins before consensus has a word for it.

The shovel you are holding

Step back far enough and the mistake stops being about technology at all. It is a mistake about constraints. The investor who misses the second act has quietly assumed that today's scarcity is a permanent feature of the world, and has therefore mispriced everything that becomes possible the day the scarcity inverts. Treat the current constraint as the fixed shape of things and the future will keep arriving as a surprise you sold too cheap.

I built a career on watching that error from the outside. A risk system misses migrating fraud for the same reason a capital allocator misses the salvage: it treats the current surface as the permanent one. A solved check reads as a closed problem; an attack that has not appeared yet reads as an attack that never will. Fifteen years on the merchant side of the payments stack taught me that the failure is identical in every room it happens in — a confidence in the present shape of the pattern that the pattern has never once earned.

But there is one more place to point the lens, and I am not allowed to exempt it. The input being overbuilt this time is not bandwidth or track. It is cognition. Intelligence is the commodity being stamped out at scale and driven toward the price of the electricity that runs it, and nearly everyone reading this — myself included — is long their own intelligence at a price set when it was scarce. If the pattern holds, the premium on having the answer collapses the way the premium on a lit strand collapsed. What survives an abundance is never the input. It is what the input cannot generate for itself: the question worth pointing it at, the nerve to act while acting still looks wrong, a name willing to sit under the decision. A machine can produce a contrarian sentence for nothing now. A contrarian position still costs what it always cost. When every desk runs the consensus at zero, consensus gets crowded faster and breaks harder, and the whole remaining spread migrates to the one asset the machine cannot mint.

The consensus trade is the majority report. The salvage thesis is the dissent.