The GPU Depreciation Debate: Burry Answers Nvidia's Retained-Value Slide With Adam Smith's 1968 "Money Game" — "We Have All Been Here Before"

The GPU Depreciation Debate: Burry Answers Nvidia's Retained-Value Slide With Adam Smith's 1968 "Money Game" — "We Have All Been Here Before"

Update 2026-10-05: the public ROI/depreciation fight has escalated from dueling data to a historical-narrative clash, and Burry has widened it into a full forensic attack on hyperscaler accounting (see The Circular-Financing Web Adds Its Largest Layers Yet: Broadcom's $50B+ OpenAI Chip Package, Oracle's Independent Chip-Lessor — and SpaceX's $40B Now Priced at 194bps CDS).

The spark. On Sept 27, Nvidia presented a slide to investors titled "NVIDIA AI Infrastructure Retains Value Beyond Accelerated Depreciation Schedules," plotting the retained value of A100, H100 and B200 chips against a five-year accelerated depreciation curve — implying companies are over-depreciating because the hardware holds value across generations (Benzinga). Burry noted bulls celebrated it as a definitive counter to the critics.

The answer. In a post titled "Don't Believe Your Lyin' Eyes, GPU Depreciation & Useful Lives," Burry reached back to "Adam Smith's" 1968 The Money Game — the Great Winfield mocking a critic for asking "how fast these computers are written off" (such questions mark you as "middle-aged"), and Billy the Kid explaining he buys data-processing stocks on leverage, putting up "at least three percent cash." Per Benzinga: "the 1960s investors argued the need for computers was 'practically infinite,' predicting that earnings would double year after year because 'the surface has barely been scratched.'" Burry's verdict, in the article's words: "Michael Burry compares Nvidia's GPU depreciation defense to the 1960s computer leasing bubble, warning 'we have all been here before.'"

Why it matters. The 1960s parallel is aimed at the exact seam in the capex story: if GPUs economically live 2–3 years but are depreciated over 5–6, reported AI profits are borrowed from the future¹ — and Nvidia's own retained-value slide is now the bull case's load-bearing exhibit. The debate's structure this cycle: Nvidia supplies the retained-value data; Burry supplies the historical analogy that the data itself is the bubble tell. Neither side has moved the orders: TSMC's 2nm customers just raised orders 10–20% (Nvidia's AI Capex Cycle Sustainability: Verdict 2026-10-08 — Intact for a Ninth Straight Read; Records on the Supply Leg, and the Debt Funding It Gets Priced Harder). The empirical referees remain the rental indexes — Silicon Data puts H100 at $2.77/hr and B200 at $5.86/hr, with the B200 term curve below spot, pricing in some easing once Rubin volume arrives — now with a public forward curve after today's CME listing (Compute Becomes a Tradable Commodity: CME Lists the First GPU Rental Futures Today (Oct 5) — a Public Forward Curve for AI Compute, With the Term Curve Already Pricing Rubin Easing).

Standing counter-evidence to the fast-obsolescence thesis (unchanged): H100 rentals contracted through 2029, CoreWeave's ~95% contract-rebooking rate, and the A100 (2020 chip) still renting at or above where it started.

What would decide it: any hyperscaler shortening GPU useful lives or taking GPU impairments (tracked watch), or a sustained break in rental rates — not more essays from either side.