The GPU Depreciation Debate: Burry's 1968 Parallel vs Nvidia's Residual-Value Slide

The GPU Depreciation Debate: Burry's 1968 Parallel vs Nvidia's Residual-Value Slide

The accounting question underneath the whole AI buildout — how fast GPUs actually depreciate — has escalated into a direct fight between Nvidia and Michael Burry, and it is now the sharpest bear argument against the basket's reported margins.

The trigger: a Nvidia investor-presentation slide from late September 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. Bulls celebrated it as the definitive counter to "depreciation bears."

Burry, writing in his Substack Cassandra Unchained ("Don't Believe Your Lyin' Eyes, GPU Depreciation & Useful Lives"), says the chart — derived from analytics firm Silicon Data — is a DCF of forward rental income over an assumed eight-year chip life, not actual secondary-market resale prices, so setting it against a five-year depreciation curve is an "apples-to-oranges" illusion. He contends Nvidia's reliance on accelerated depreciation is "an implicit admission of how severely each new product generation devalues its predecessors." On why legacy GPUs still earn high rental rates: "A chip earning more than it costs is a sign of scarcity due to memory and power shortages, not GPU die shortages and not durability."

The historical parallel is 1968: third-party computer-leasing firms undercut IBM by using eight-year depreciation schedules against IBM's conservative four-year timeline. They were right that mainframes physically lasted — but rental yields collapsed when IBM launched System/370. "By 1970, computer leasing equities fell by roughly 80%, wiping out heavily leveraged funds and sparking industry-wide insolvencies." Burry warns the modern analogue — GPU-backed debt, asset-backed loans and private credit financing neoclouds and PE-backed data centers — shifts the losses to institutional lenders and private insurers when rental yields compress.

Positioning: Burry covered short common-stock positions in Micron, Nebius, Caterpillar, SOXX, CoreWeave, Nvidia and Palantir, and "replaced the Nvidia short with September 2027 puts at strikes in the mid $100s."

Context per the market view: NVDA trades at a $5.65T market cap with RSI(14) of 83 and 98.2% of its 52-week range, on a 74.7% gross margin and TTM revenue of $302.97B (+105.9%) — the tape is as extended as the accounting debate is unresolved.

Why it matters for this basket: if GPU useful lives are really 2–3 years rather than 5–8, hyperscaler operating margins are overstated, the neocloud credit stack is mispriced, and NVDA's own replacement-demand curve is flattered¹. This is the operating-income complement to the paper-gain story The Paper-Gain Ledger: Alphabet's $98B SpaceX Mark and Amazon's $53B Anthropic Mark — Now Facing IPO Price Discovery and the off-balance-sheet commitments ledger The $3 Trillion Shadow Ledger: Off-Balance-Sheet AI Commitments Dwarf Reported Debt — three separate ways headline AI earnings overstate economic profit. The November 20 NVDA print, where gross margin is guided down to 74% on memory scarcity, is the next hard data point (watch registered).