AI Benefits Outrun Capex Only If GPUs Last Six Years. Burry Says Three.
There is a respectable bull case that benefits eventually grow faster than capital expenditure, and it does not depend on optimism. It depends on arithmetic. The trouble is that the same arithmetic reduces the entire thesis to a single contested number, and that number is currently being attacked by the man who shorted the last bubble.
The Mechanical Case Is Stronger Than the Hype Case
Strip away the cheerleading and the bull argument is structural. Capex is a step function: compute, data centers, substations, and grid interconnects are sunk before a dollar of inference revenue arrives. Monetization is an S-curve that begins only once capacity is utilized. The two curves are offset in time by construction, which means peak capex intensity is not the steady state of the benefit-to-cost ratio. It is the trough. The ratio improves mechanically the moment spending shifts from building capacity to running it, even if technological progress stopped entirely.
Three forces then compound on top of that timing. The marginal cost of serving an additional user is trivial against the value delivered, which is operating leverage of the kind that made software the highest-margin business ever constructed, now pointed at cognition. Capability per dollar is falling faster than any industrial input in history, so each buildout cycle buys multiples more deliverable intelligence than the last. And Jevons does the demand-side work: as cost per task collapses, the set of economically automatable tasks expands convexly, because work that was not worth automating at one price becomes worth it at a hundredth of that price. Benefits do not scale linearly with capability. They scale with the integral of newly viable use cases.
Every prior general-purpose buildout rhymes with this. Railroads, electrification, and the late-1990s fiber glut all produced catastrophic benefit-to-capex ratios during the overbuild and decades of returns off the laid infrastructure afterward. The dark fiber stranded in 2001 carried the traffic of the 2010s. Solow’s quip about computers appearing everywhere except the productivity statistics resolved roughly a decade later, once the complementary reorganization caught up. Measured benefit lags real benefit through the investment phase, then snaps.
The Whole Thesis Rests on One Hinge
Here is the part the bull case usually elides. Every analogy above assumes the installed base keeps producing after the spending stops. Stranded fiber paid off because glass in the ground has a thirty-year life and amortizes against three decades of traffic. The benefits-outrun-capex argument is, at bottom, the claim that an AI buildout behaves the same way: spend now, harvest the installed base for years.
That claim is false if the installed base does not last. If a GPU cluster has a two-to-three-year economic life because each silicon generation renders the prior buildout uncompetitive, there is never an “installed base” phase to harvest. There is only perpetual rebuild. The capex S-curve never flattens into a utilization plateau; it resets every cycle. Thirty-year infrastructure amortizes. Three-year infrastructure may never get the chance.
So the thesis is not “benefits always win.” It is strictly conditional: benefits outrun capex if and only if the depreciation cycle lengthens faster than the buildout cycle repeats. That is one variable, and it is the only one that matters.
The Numbers Make the Stakes Concrete
The capex side of the ledger is no longer an estimate. The four largest hyperscalers spent roughly $410 billion in 2025 and have guided to around $725 billion for 2026, a 77 percent increase: Amazon near $200 billion, Alphabet $175 to $185 billion, Meta $115 to $135 billion, Microsoft tracking toward $120 billion or more. Goldman now models $5.3 trillion of combined hyperscaler capex from 2025 through 2030. Tech equipment and software investment reached roughly 4.4 percent of GDP in 2025, near the dotcom peak. This is not a sector allocating capital. It is a sector at war over it.
The depreciation side is where the bull thesis lives or dies, because useful life is the accounting expression of exactly the conditional above. Between 2020 and 2024 the big three quietly extended assumed server life from three or four years to six. Amazon moved from three to four years in 2020, then to six by 2023. Microsoft and the others followed. Lengthening useful life is not a footnote; it is the income statement asserting that today’s silicon will generate revenue for six years. That assertion is the benefits-outrun-capex thesis, written in GAAP. The collective effect of the extensions was an estimated $18 billion in annual depreciation expense removed from the cohort’s costs.
The tell came in early 2025, when the consensus cracked. Amazon shortened the life of a subset of its servers and networking gear back to five years, citing the increased pace of technology development. Meta moved the other way, stretching most of its server and network assets toward five and a half years and some classes far longer. The same hardware, the same physics, opposite accounting. When two of the most sophisticated operators on earth diverge on the single load-bearing assumption, the variable is not settled. It is contested inside the buildout itself. Satya Nadella said the quiet part aloud, admitting he did not want to be stuck with several years of depreciation on one chip generation.
Burry Found the Pressure Point
In November 2025 Michael Burry broke a long silence to argue that the hyperscalers are systematically overstating profit by depreciating Nvidia hardware over five to six years when its real economic life is closer to two or three. His arithmetic put the gap at roughly $176 billion of understated depreciation and overstated earnings across the industry between 2026 and 2028. The framing was inflammatory; the inputs were not. They are public, audited, and sitting in the filings.
His mechanism is simple and hard to dismiss. Nvidia now ships a new architecture on roughly an annual cadence, each generation two to three times more efficient per watt, which makes prior-generation silicon uneconomic for frontier training almost immediately. A six-year schedule against a one-year refresh asserts that five or six chip generations coexist productively. The bull rebuttal is the value cascade: a GPU does training at the frontier for a year or two, drops to high-value real-time inference for a couple more, then ages into batch inference and analytics at the long tail, generating revenue the whole way down. If the cascade holds, six years is defensible and the benefits thesis survives. If frontier inference also demands the newest silicon, the cascade collapses, useful life is genuinely two to three years, and the thesis dies with it. This is not an accounting argument dressed as economics. It is the economics, with the accounting as its only public scoreboard.
What Actually Decides It
The benefits-outrun-capex case and the six-year-useful-life assumption are the same claim wearing different clothes. One is an economic forecast; the other is its representation on the balance sheet. You cannot believe the first and dismiss the second as arcana, because they stand or fall together. Burry’s $176 billion is simply the bear case, priced.
That reframes what to watch. The instinct is to wait for AI revenue to validate the spending, or for the productivity statistics to confirm the Solow paradox is resolving. But those are lagging signals, and the leading one is already public. The cascade either holds or it does not, and the first place it breaks will not be a revenue miss. It will be a depreciation disclosure. Amazon shortening a subset to five years was the bull case quietly cracking in real time, ahead of any income statement. The next hyperscaler to shorten useful life, or to take an asset retirement charge on stranded silicon, is telling you the cascade failed before the revenue ever gets the chance to.
So track one ratio above all others: the gap between Nvidia’s refresh cadence and the assumed useful life on the filings. It is currently running near six years of depreciation against roughly twelve months of obsolescence. Five generations of silicon are being booked as if they earn together. That spread is the entire bet. The benefits will outrun the capex precisely when, and only when, that number stops widening.