Michael Burry, Nvidia, and the Compute Question: Is the AI Boom Built on Sand?

On Thanksgiving week, hedge-fund contrarian Michael Burry put a new target on the AI rally: Nvidia. He published bearish positions, launched a paid Substack, and accused the AI supply chain of overstating demand and hiding depreciation. Markets, engineers, and startups woke up to a question that has become more technical than tribal: how durable is demand for massive GPU fleets?

Burry’s argument matters because it ties accounting choices-how companies amortize GPUs and account for stock compensation-to the core economics of AI compute. If his claim that Nvidia’s compensation and customers’ useful-life assumptions have inflated profitability holds any weight, the ripple effects reach valuations, capital spending, and whether startups can afford to train state-of-the-art models. This story intersects finance, ML engineering, and procurement in a way that will decide who keeps building and who runs out of cash.

The short and the sharp: what Burry is saying

Michael Burry’s public case is simple: Nvidia’s ascent is priced on future demand that may not materialize, and accounting around GPUs masks how quickly those assets actually lose value. In filings and on his new Substack, Burry flagged what he says are $112.5 billion of stock-based compensation charges baked into Nvidia’s shareholder calculus, a figure Nvidia disputes as inflated by tax treatment and RSU math.

Burry went further, suggesting customers are stretching the useful life of GPUs to justify massive capital expenditures. If companies treat a GPU as a three-year asset rather than an eighteen-month one, reported depreciation falls and near-term profits look healthier. That matters because valuation multiples for hardware-centric firms are driven by near-term margins and forecasted reinvestment needs.

Why GPU accounting changes the math for AI

The tone has rattled markets because Burry is not anonymous; his newsletter launched with a reported 90,000 subscribers at $400 a year, giving him a big megaphone. The spat escalated after Palantir CEO Alex Karp called Burry “batshit crazy” on CNBC, and Nvidia’s investor-relations team fired off a seven-page memo disputing the math, saying Burry “incorrectly included RSU taxes” and that the real buyback figure is closer to $91 billion.

GPUs are the cost center of modern AI. Training a frontier generative model can consume millions of GPU-hours and devour tens of millions of dollars in electricity and rack space. How an organization amortizes that hardware-over one year, three years, or more-directly alters reported operating margins and the implied return on capital.

The ripple effects: startups, cloud providers, and capital flows

Think of it like a factory retooling for a new product. If you say the new machines will last five years, your annual charge is small; if they last one year, your annual charge jumps fivefold. For AI, the vintage of chips matters quickly. New GPU generations deliver step changes in performance per watt, so hardware can become functionally obsolete faster than traditional enterprise servers.

That decay is compounded by software. Model optimizations, sparsity tricks, and runtime compilers can squeeze more useful work out of older hardware, but only up to a point. The net effect is that the useful-life assumption is both an engineering judgment and a financial lever.

Parsing the claims: numbers, noise, and what we can verify

If Burry’s narrative gains traction, the immediate victims would be companies whose business cases hinge on relentless GPU spending: startups training big models, cloud providers offering GPU fleets, and enterprises planning capex-heavy AI initiatives. A pullback in expectations of perpetual growth would tighten funding, raise the cost of capital, and force reprioritization of projects.

Venture economics already show strain. Startups that booked multi-month to multi-year rental agreements for tens of thousands of GPU-hours can see unit economics flip if discounts evaporate. Large cloud providers and hyperscalers could respond by changing pricing tiers, gating access to high-end accelerators, or offering more aggressive managed services tied to usage rather than upfront capacity.

There are positive knock-ons, too. Pressure on capex will accelerate work on efficiency: model distillation, quantization, sparse training, and better scheduling to squeeze more useful FLOPS per dollar. Those technical levers are real and measurable; they change the denominator in the compute-cost equation and could blunt a demand collapse.

Parsing the claims: numbers, noise, and what we can verify

• Palantir CEO Alex Karp calls Michael Burry 'batshit crazy' in exchange over put options - CNBC, 2025-11-24

• Nvidia investor-relations memo rebuts Michael Burry’s claims - Barron's, 2025-11-26

• Moving toward LessOps with VMware-to-cloud migrations - MIT Technology Review, 2025-11-27