Engineers Are Underwriting $4 Trillion of Debt and Don't Know It

A few weeks back I was listening to Cory Doctorow on the Lawfare podcast, and he got onto the economics of the AI datacenter boom, which he thinks is madness. His sharpest point was about depreciation. The hyperscalers book their GPUs on five-year schedules, but Doctorow argues that a card run flat out can burn itself out in under two months, and that the honest duty cycle is closer to two or three years. Booking a two-year asset as a five-year one, he says, is a species of accounting fraud, and it’s propping up the most expensive build-out in the history of the industry. Somewhere in the middle of that I stopped thinking about the bubble and started thinking about a server my team used to run: four GTX 1080s, kept alive from 2018 to 2024. By Doctorow’s math those cards should have been dead, or at least worthless, long before we finally retired them. They weren’t. That gap is the whole essay.

The useful life of an AI accelerator is an engineering output that accountants transcribe after the fact. It equals the depth of the workload cascade the chip can fall through: frontier training at the top, then ordinary training, then live inference, then distilled and quantized inference, then compiled batch work at the bottom, where our 1080s spent their last four years. The deeper the cascade, the longer the chip earns, and the depth of the cascade is set by the people building distillation and compilation tooling, three layers below the capital markets pricing the asset. In an earlier post I argued that leaving a model in a hot inference loop is running an un-compiled interpreter, and that the compiled tail is where the efficiency lives. This is that argument pointed at a balance sheet. Doctorow is right that the five-year number is fiction. Where I’d turn his argument is on the reason: he treats the short life as a hidden hardware fact, the chips quietly dying faster than the books admit, and I think the truth runs the other way. Useful life was never a property of the silicon in the first place, which is why there’s no hidden hardware fact for the books to catch up to. Engineers set the number, and mostly they set it longer than the market assumes.

The mechanism is principal-agent with hidden information. A principal hires an agent to act for them, but the agent knows something the principal can’t see, and that private knowledge decides the outcome the principal is exposed to. Here the principal is whoever underwrites the debt: the bank modeling a depreciation schedule to decide whether trillions in data-center bonds get repaid. The agent is the engineer deciding whether a distillation pipeline exists to catch last generation’s workload. The bank can audit everything physical about the chip: its spec sheet, its power draw, its resale price. What it can’t see is whether someone two org charts away is about to ship the tooling that keeps that chip earning for another two years. The number that secures the collateral lives with the agent, and the principal prices it as though it fell from the sky.

The scale is what makes the asymmetry costly. JPMorgan put AI capex through 2030 at $5.5 trillion, with roughly $4.1 trillion of it debt-financed. Goldman Sachs showed that depreciating this hardware over three years instead of seven swings cumulative cost by $1.76 trillion, which makes useful life the single largest swing variable in the whole build-out. The market and Doctorow are fighting over that variable from opposite ends of the same table. The bulls hold out for six or seven years because the schedule has to pencil out, while Doctorow says two because the silicon physically wears out, and both are arguing about the chip when the number actually lives in the cascade underneath it.

Picture a GPU’s working life as a fall through tiers of decreasing value per hour. At the top it trains frontier models, the only work that justifies buying an H100 at the front of the queue, and the work that runs the fans hard enough to make Doctorow’s burnout math real. Chips don’t stay at the top. When the next generation lands, that card drops a tier: training smaller models, fine-tuning, work that doesn’t need the frontier. Below that is live inference, still latency-sensitive and still valuable. One tier down, inference runs on distilled and quantized models, where a card half the speed of current hardware is fine because the model shrank to fit it. At the bottom is batch work with no latency budget at all, running cool and cheap, which is exactly where our 1080s lived for four years. Running a card flat out at the top is where Doctorow’s burnout holds. It says nothing about the bottom.

Each tier is a rung of useful life, and how many rungs a chip falls through is a question about the tooling downstream: whether something exists to catch the workload as it drops. Distillation shrinks a model to fit slower hardware; quantization does the same job in a different currency, trading precision for throughput; the loop-compilation I’ve written about before strips interpreter overhead so batch jobs run on whatever’s cheap. On our 1080s, quantizing to 8-bit bought three to four times the throughput on at least one workload, which is close to what a Pascal card can do once you feed it the four-way integer instruction it was built around. A model that needed an A100 to serve at speed in 2024 can, distilled and quantized, serve fine on hardware two generations older. The workload came down to meet the cheaper silicon, and that card’s earning life stretched by years without anyone booking the change.

Doctorow’s own favorite analogy is the one that undoes him here. Twenty-odd years ago the telecom bubble laid far more fiber than anyone could use, WorldCom cooked its books on the back of it, and the sector collapsed. The fiber stayed in the ground. A decade later it got lit for pennies on the dollar, and the cheap bandwidth under streaming, cloud, and the modern internet was the overbuild’s second life. Doctorow expects the same ending for AI: the bubble pops, the datacenters go dark, and the GPUs reach climate scientists and indie animators and open-source hackers at scrap prices.

He’s got the destination right and the timing wrong. The second life doesn’t wait for the crash. The cascade is that conversion market, and it runs continuously, inside the datacenters, right now. Every distilled model shipped to production is a chip being handed its next tier of work while the bubble is still inflating. Our four 1080s were dark fiber that never went dark.

The best argument against all of this is that the real constraint is power, and no amount of clever software changes that. If each GPU generation delivers dramatically better performance per watt, then keeping old cards busy is a losing trade whatever your tooling does, because the old chip burns more energy per unit of work than a new one would, and energy is the bill that never stops. On that view useful life really does collapse into an electricity question, and the cascade lever can’t reach it. This is the same reason a six-year schedule keeps getting harder to defend the more each new generation improves on power, and it’s the strongest thing anyone can say against me.

It’s correct wherever power is the binding cost, and I won’t pretend otherwise. It costs me the clean universal claim. What survives is narrower and still worth a lot: the cascade sets useful life wherever energy is cheap, or wherever the fleet is already built and already powered, running on marginal electricity close to free. A sunk, colocated cluster with idle capacity is exactly where depreciation becomes an engineering choice again, because the only open question is whether there’s work cheap enough to run on it. Our 1080s sat in a room that was already lit and already cooled. The marginal cost of one more overnight batch job rounded to nothing, so the cascade paid.

The number that decides whether $4 trillion of debt is money-good gets set a long way from any boardroom or ratings model, by whoever decides, this quarter, whether to build the distillation pipeline that keeps last year’s cluster earning. That engineer thinks the job is saving on compute. A distillation pipeline is a depreciation schedule with a different job title, and nobody told them they’re the one signing it.