On November 13, 2025, Michael Burry filed the paperwork to shut down Scion Asset Management, the firm that made him a legend for seeing the 2008 mortgage crisis coming before almost anyone else. He didn't go quietly. In the weeks prior, he had disclosed more than $1 billion in bearish bets against Nvidia and Palantir, built on an unusually specific argument: that America's biggest tech companies were overstating profits — by his estimate, $176 billion worth between 2026 and 2028 — through how they depreciate their GPUs.
On its face, it was a bookkeeping dispute. Nvidia's chips, Burry argued, behave like commodities with two-to-three-year useful lives, not the five- or six-year assets some hyperscalers carry them as. Meta had just stretched its server depreciation schedule to 5.5 years, trimming $2.9 billion off expenses — nearly 4% of pretax profit — in a single stroke. Amazon, staring at the same silicon, went the other way, shortening its schedule and taking a $700 million hit. Two of the most sophisticated finance organizations on earth, looking at identical hardware, reached opposite conclusions about how fast it wears out. That divergence should worry an investor more than any keynote about superintelligence. It suggests nobody actually knows.
That's the real argument to have about AI, and it isn't the one most people are having. That the technology will reshape the economy is no longer seriously contested. Whether the sums being spent to build it will earn back anything like what today's valuations assume is a separate, harder question.
History isn't comforting here. The internet transformed commerce; most companies that raised money on that promise in 1999 no longer exist. Railroads reordered how goods moved across continents while bankrupting the men who financed the track. A technology can be indispensable and still be a poor place to have put your capital. A market can be enormous without the companies selling into it capturing much of that value as profit. The size of the pie says nothing about who gets the slice.
AI has a structural problem neither the internet nor enterprise software faced: it's expensive to serve, not just to build. Traditional software, once written, cost almost nothing to sell again. Every AI query burns real, ongoinghyperscalersower. That's one reason the biggest hyperscalers are on pace to spend $700 billion to $760 billion on AI infrastructure in 2026 alone — capital that has to be serviced with cash flow the technology hasn't yet reliably produced.
Increasingly, that capital comes from debt. Oracle's credit-default-swap spreads — the price of insuring against an Oracle default — hit an 18-year high in July 2026, driven by anxiety over how much of its buildout, including its roughly $300 billion commitment tied to OpenAI, is debt-financed rather than cash-funded. That debt sits inside a tangle of related-party dealing: Nvidia invests in OpenAI, OpenAI commits to buy compute from Oracle and CoreWeave, Oracle buys chips from Nvidia to build that compute — and each transaction gets booked as revenue somewhere in the loop. None of it is illegal. But when the same dollars appear to circulate among a small number of counterparties, it gets harder to tell how much real external demand sits underneath the numbers — the same question examiners were asking about mortgage securitizations in 2007, long before anyone said "crisis."
None of this requires believing the technology doesn't work. It likely does. The trouble is that working and paying for itself aren't the same achievement. A widely cited MIT study found that roughly 95% of generative-AI pilots at large companies failed to show a measurable return in 2025 — not evidence AI is a dead end, but evidence that converting a genuinely useful technology into a bottom-line result is its own unsolved problem. And even where AI does make a bank or a retailer more productive, competition tends to push the gain toward customers as lower prices, not toward the seller as margin. That's been true of nearly every general-purpose technology in economic history. There's no obvious reason AI repeals it.
None of this settles the argument. Inference costs have fallen sharply since 2023 and could keep falling. Enterprises could eventually redesign whole workflows around AI rather than bolting it onto what already exists. If that happens, today's data-center spending will look, in hindsight, like the railroads that did get built profitably.
But that's a thesis, not a conclusion — and the more money committed on the assumption it's already proven, the costlier it gets to discover otherwise. By the time hyperscalers close their books on 2028, the GPUs at the center of Burry's argument will be reaching the end of the shorter working life he insisted was the honest one. Somewhere between his number and theirs sits $176 billion. Whether that gap turns out to be an asterisk in an annual report or the first line of the next one will say more about this era than any amount of enthusiasm for the technology ever could.
No comments:
Post a Comment