The most revealing thing about Nvidia's September 27 slide is who made it. In 1968 the manufacturer was the force that eventually destroyed the computer lessors' residual values. In 2026 the manufacturer is the one vouching for them. That inversion is the rhyme worth hearing in Michael Burry's latest History Rhymes piece. The consensus reading of the slide, that it settles the depreciation debate, mistakes evidence of scarcity for evidence of durability.
A slide built for someone else's income statement
The scene was an investor presentation. Nvidia, by then the most valuable company in the world, was on the road explaining itself. Partway through the deck came a slide titled "Nvidia AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules." It plotted the purported retained value of three generations of accelerators, the A100, the H100 and the B200, against a five-year accelerated depreciation curve. The message was that GPUs hold their value well enough across generations that the companies buying them are, if anything, writing them off too fast.
Nvidia's admirers took it as a decisive blow against the bears. Tae Kim, author of The Nvidia Way, was among those celebrating. The critics the slide seemed aimed at had spent the better part of a year asking how fast these chips and the machinery around them should be expensed. Burry counts himself among them.
Two features of the slide deserve more attention than they received.
The first is the benchmark. An accelerated schedule front-loads expense, so it is the easiest curve to beat. The largest buyers do not book their servers that way. Over 2022 and 2023 several of them lengthened server useful lives on a straight-line basis, to six years in some cases, and Amazon moved part of its fleet back toward five years in 2025. Showing that retained value clears the most aggressive write-off curve available does not show that it clears the schedules actually used to compute reported earnings.
The second feature is the speaker. Nvidia does not depreciate the GPUs it sells. Its customers do. A slide arguing that customers over-depreciate is, in effect, a manufacturer defending the earnings of its buyers, because those earnings fund the next order. In 1968 the incentives ran the other way, and that difference is the one genuinely new element in an otherwise familiar pattern.
Billy the Kid and the long life of the System/360
The episode Burry chose comes from The Money Game, published in 1968 under the pseudonym Adam Smith by George Goodman. Warren Buffett recommended the book in his mid-year partnership letter that year, calling it "a magnificent account of the current financial scene" and "loaded with insights and supreme wit."
In its most famous scene, Smith walks into the office of the Great Winfield, a tape-reading speculator who mocks him on sight for "framing questions about depreciation, about how fast these computers are written off." Such questions, Winfield says, show you are middle-aged. His answer to the market is kids: Billy the Kid, Johnny the Kid and Sheldon the Kid. He stakes them, they pick the stocks, and they split the profits. Billy had turned five thousand dollars into more than half a million in six months.
His method was computer leasing stocks. He named Leasco Data Processing, Data Processing and Financial General, and Randolph Computer. He bought the convertibles, borrowed against them, and bought more, putting up three percent cash, or five when he felt conservative. His thesis was clean and, on its own terms, correct: the need for computers was practically infinite, leasing was the only practical way to sell them, and the manufacturers lacked the capital to finance their own customers. Earnings would double, then double again. When Smith raised the obvious objection, Winfield answered it before it was asked: what makes a finance company worth fifty times earnings?
That question was the right one, and the business model explains why. The lessors bought IBM System/360 machines, announced in April 1964, and leased them to corporate users on terms that undercut IBM's own rentals. Their reported profits depended on a single assumption: that the machines would keep earning rent for many years, longer than IBM's product cadence would later allow. A long useful life spread the purchase price thinly across the years and produced the earnings the Kids extrapolated. A short one would have turned the same cash flows into much smaller profits, or into losses.
For a while, the evidence seemed to support the long lives. Demand for computing was real and growing, used 360s were valued accordingly, and the lessors' stocks rose with each doubling of reported earnings. Leasco's shares became currency, which Saul Steinberg used to acquire Reliance Insurance in 1968 and then to bid for Chemical Bank in 1969, a bid that failed.
The turn came from the manufacturer. On June 30, 1970, IBM announced the System/370. The lessors' inventory did not stop working that day. It stopped being what customers wanted, and its rental value began to fall toward what the newer machines made it worth. The residual values in the lessors' depreciation schedules had been estimated during a period of scarcity and growth, and they were repriced in an environment of replacement. The leasing stocks had already fallen sharply through 1969 and 1970. Billy the Kid's three percent margin left no room for that repricing.
Smith's question, the middle-aged one, had been the only question that mattered. The Kid had been right that the need for computers was practically infinite. He had been wrong to treat that as the same thing as the value of the particular computers his companies owned.
Fiber that lasted and companies that did not
The pattern recurred a generation later in a form that sharpens it. Between the mid-1990s and 2001, carriers laid an enormous amount of fiber-optic cable on the premise that internet traffic would grow without limit. The premise was largely correct. Traffic did grow enormously, and much of that glass is still carrying data today. On the narrow question of physical useful life, the optimists were vindicated beyond anything their schedules assumed.
The companies were not. Global Crossing filed for Chapter 11 in January 2002, and WorldCom followed in July 2002 after disclosing that it had capitalized billions of dollars of ordinary line costs. The fiber lasted, but its price did not hold. Wavelength-division multiplexing allowed each existing strand to carry many times more traffic, so the supply of bandwidth grew faster than demand, and the price per unit of capacity fell accordingly. An asset can outlive its depreciation schedule and still lose most of its economic value, because value is set by the cost of the newest equivalent capacity, not by whether the old capacity still functions.
This is the distinction the Nvidia slide blurs. A GPU from 2020 can run for years. Whether it earns its book value depends on what a customer will pay for its compute when the alternative is a newer chip offering several times the work per watt, from a manufacturer that now ships a new generation roughly once a year. The fiber carriers learned that physical life and economic life are separate quantities. The 1968 lessors learned that the manufacturer's product calendar governs economic life. Both lessons point to the same blind spot.
Shale oil, or being right decades too soon
Smith's scene contains a second Kid whose story is easy to skip. Sheldon owned Western Oil Shale, which had gone from three to thirty. The western United States, he said, sat on shale oil reserves five times all the known reserves in the world. Equity Oil could earn seven hundred and fifty dollars a share. A commercial underground nuclear test was coming that would unlock it. "Shale oil! Shale oil!" Winfield cried, as if the idea were already old.
Sheldon was right about the resource. American shale did transform world oil supply, roughly four decades later, after horizontal drilling and hydraulic fracturing made it economic, and the United States became the world's largest producer. He was wrong about the technology, wrong about the timing, and wrong about which shareholders would benefit. Nuclear stimulation never became the method. The fortunes of the 2010s went to different companies, using different tools, financed by different capital.
Sheldon belongs in this story because he isolates the variable that causes the most confusion in the GPU debate. The bulls tend to defend AI demand, as if the depreciation critics were attacking it. They need not be. Billy the Kid's demand thesis for computers came true. Sheldon's resource thesis for shale came true. Neither outcome rescued the equity that had been priced on the near-term version of the story. Being right about the size of a technology and being right about the value of a specific asset, financed on a specific schedule, are separate calls. Markets have often paid handsomely for the first and settled accounts on the second.
Where the 2026 tape rhymes, and where it does not
The parallels are close enough to name directly. There is a scarce, rapidly improving piece of capital equipment. There is a claim that demand for it is practically infinite. There is a reported earnings stream whose size depends heavily on assumed useful lives. There is a secondary market price, observed during a shortage, offered as proof that those lives are conservative. And there is a growing class of intermediaries that buy the equipment with borrowed money and rent it out, the role Leasco and its peers played in 1968. The specialist GPU clouds are the closest modern counterparts, though which of them, if any, carries Billy the Kid's leverage is a balance-sheet question for each company, not something the analogy can settle.
The financing environment makes that last link more than a literary echo. A lessor's spread is the gap between rental yield and the cost of carrying the asset. With the 10-year Treasury at 5.26% and the 2-year at 4.89% at the end of September, the carry is not cheap. Any asset whose rental value falls on the manufacturer's schedule, rather than the lessor's, compresses that spread from both sides at once.
The retained-value slide fits this sequence at a recognizable stage. Used-equipment prices during a shortage are real prices, but they measure scarcity. Used 360s held their value while IBM's backlog was long. Leased fiber held its value until multiplexing arrived. A chart of A100 and H100 resale or rental values drawn during a period when buyers cannot get enough compute is evidence that compute is scarce in 2026. It is not evidence that a 2023 accelerator will earn its carrying value in 2029, after two or three more product generations from the company that drew the chart.
The difference from 1968 is real, and it cuts in two directions. Nvidia is not a leveraged lessor. It is the manufacturer, with margins and a balance sheet that bear no resemblance to Leasco's, and the hyperscalers that buy most of its output have historically funded capital spending from operating cash rather than from three percent margin. On that basis, the most direct exposure to a residual-value repricing sits with the intermediaries and their lenders, not with the chipmaker.
The inversion of incentives, however, removes the brake that existed in 1970. IBM had every reason to obsolete the 360 installed base, because the lessors were competing with its own rentals and the 370 won that customer base back. Nvidia's position is double-edged. Its annual cadence is what ages the installed base, yet its order book depends on customers whose reported earnings look healthiest when that installed base is assumed to age slowly. A manufacturer publicly certifying the residual value of chips it is actively superseding is acting less like IBM in 1970 and more like an underwriter of the lessors' assumptions.
The sequence that recurs
Across the computer leases of the 1960s, the fiber of the 1990s, and, so far, the accelerators of the 2020s, the order of events has been strikingly similar. A genuine technology arrives with demand that is described, accurately, as enormous. Secondary prices formed during the shortage are taken as proof of durability. Useful lives are stretched to match, and reported earnings rise. Leverage then capitalizes those earnings, and the next generation of equipment arrives on the manufacturer's timetable rather than the owner's, repricing the residual all at once.
The episodes do not always end the same way, and nothing in the pattern dictates when the repricing comes. The recurring feature is that the decisive variable was the one the optimists dismissed as a middle-aged question. In 1968 the Great Winfield mocked it. In 2026 it is answered with a slide. In both cases, how quickly the machines should be written off is the question that determines what the earnings are worth.
