The three-year clock hiding inside a thirty-year building

Take a $40,000 GPU. Depreciate it over three years and you book roughly $13,333 a year in expense. Depreciate the identical chip, same cash outlay, same box, over six years, and you book about $6,667. Nothing about the hardware changed. What changed is a single assumption on a disclosure schedule, and that assumption is worth twice the reported profit on every GPU a hyperscaler owns.

That is the arithmetic Michael Burry put in front of Wall Street in November 2025, and it is why a spreadsheet input most people never read has become the most contested number in the entire buildout.

What did the hyperscalers actually change, and when?

A lot, and recently. Amazon moved server depreciation from three years to four in 2020, then to six by 2023, arguing its servers "serve a diverse set of use cases" and earn revenue longer than the old schedule assumed. Microsoft and Google made similar moves in 2022 and 2023. By 2023 all three had converged on six years, and the useful-life extensions had collectively saved the industry an estimated $18 billion a year in depreciation expense compared with the old three-to-four-year schedules.

Then the schedules started moving again, in opposite directions, inside the same reporting quarter. In February 2025, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, taking a $700 million hit to operating income and roughly $920 million in accelerated depreciation, citing "the increased pace of technology development" behind its accelerator fleet. The same quarter, Meta extended the useful life of most of its servers and network equipment to five and a half years, booking a $2.9 billion reduction in depreciation expense, about 4% of its estimated pre-tax profit. Same industry, same three months, opposite conclusions about how fast the hardware ages.

Is Burry right that the real number is three years?

His claim, and it is a claim rather than an audited fact, is that Nvidia's generational cadence makes the real economic life of frontier accelerator silicon two to three years, not five or six, because the newest chip out-competes the old one on cost-per-token so badly that the old one stops earning its keep long before it physically fails. Burry estimates this understates industry depreciation by roughly $176 billion between 2026 and 2028 and that by 2028 Oracle's reported earnings could run 26.9% above the economic reality and Meta's 20.8% above it, if his three-year assumption is the correct one and six years is not.

Nvidia disputed the framing directly, telling analysts the comparison to accounting frauds like Enron or WorldCom does not hold, and the industry's own counter-argument has a name: the value cascade. The idea, laid out by David Vellante at theCUBE Research, is that a GPU does not retire when a newer model ships. It cascades: years one and two on frontier model training, years three and four on high-value real-time inference, years five and six on batch inference and analytics work that does not need the newest silicon. Under that framework six years of revenue-generating life is not aggressive accounting. It is a description of how the fleet actually gets used, and Amazon made the identical argument in 2020 about EC2's compute mix.

Both stories can be true for different fleets and still leave you with the same problem: nobody outside the finance department actually knows which one applies to the GPUs sitting in a given campus today.

Satya Nadella has said as much, in his own words. Discussing Microsoft's chip strategy in November 2025, he explained the company deliberately paces its Nvidia purchases to avoid over-committing to a single generation: "I didn't want to go get stuck with four or five years of depreciation on one generation." That is a hyperscaler CEO describing depreciation risk as a purchasing input, not an accounting afterthought, months before Burry's public estimate landed.

What is actually driving the change: hardware, or earnings?

Both, and the honest version of this post says so plainly. Nvidia's own product cadence has genuinely accelerated, Blackwell to Vera Rubin inside about eighteen months, so the argument that useful life should move at all is not manufactured. What is worth watching skeptically is direction and timing: useful-life extensions land almost exclusively in quarters when reported earnings need the help, and shortenings land when a company can absorb the hit without moving the stock. That pattern is not proof of misconduct. It is a description of management discretion operating at the edge of what GAAP allows, and GAAP allows quite a lot of edge.

What does the building look like next to all of this?

Almost embarrassingly stable. Commercial developers routinely underwrite a shell on a 30-to-40-year economic life, and nonresidential real property depreciates on a 39-year straight-line schedule under IRS Publication 946, roughly 2.56% of basis per year, the same rulebook that ages an office park, a warehouse, and a data-center shell alike. Epoch AI's cost model for a one-gigawatt campus assumes a 14-year facility life against a 5-year server life as its baseline for capital-recovery purposes, and its own sensitivity table shows why the server assumption is where all the risk concentrates: move server life to three years and the annualized cost of the whole campus rises from $8.5 billion to $12 billion. Move it to seven years and it falls to $7 billion. A five-billion-dollar swing, inside one input, on hardware that sits inside a shell nobody argues about.

A 39-year building holds still. A five-versus-three-year server assumption alone moves a one-gigawatt campus's annualized cost by roughly $5 billion.

The secondary market already has an opinion, and it sides closer to Burry than to the six-year camp. Used H100 systems, three years into their working life, have traded at roughly 45% of new unit cost. That is not a rounding error against a six-year straight-line assumption; it is a fleet losing more than half its value on the exact timeline Burry described as the real economic clock. Lenders have noticed. GPU-backed debt facilities, structured as delayed-draw term loans secured by the hardware itself plus customer contracts, went from roughly $2.3 billion at about 15% floating in August 2023 to an $8.5 billion facility in March 2026 priced at SOFR plus 2.25% floating or 5.9% fixed, rated investment-grade by Moody's and DBRS, running to 2032. The credit market is pricing GPU collateral out to 2032 while the accounting department depreciates the same chips over five to six years. Both numbers cannot be the honest one.

CoreWeave's business at its Kenilworth, New Jersey campus makes the stakes concrete rather than theoretical. The company's entire model is leasing GPU compute on multi-year contracts against equipment it is simultaneously depreciating on its own schedule, inside a building whose useful life nobody disputes. If the GPUs age out of their most valuable use in three years but the lease and the depreciation schedule both assume five or six, the mismatch does not stay a footnote. It becomes a refinancing conversation. The same tension sits underneath the Abilene Stargate campus, an $11.6 billion financing built around a roughly 400,000-GPU target: the loan covenants and depreciation assumptions embedded in that structure were set against somebody's view of chip life, and that view is exactly the number in dispute.

My read, stated so it can be checked

I think both sides are describing real phenomena and neither is describing the whole fleet. My bet is that by the 2027 or 2028 reporting cycle, at least one major hyperscaler quietly shortens useful-life assumptions again, not because Burry won the argument in public, but because a specific cohort of training-era GPUs stops earning inference-grade revenue faster than the value-cascade model assumed, and the write-down shows up as a targeted charge rather than a company-wide policy reversal. If 2028 arrives and every hyperscaler is still depreciating on a flat five-to-six-year schedule with no cohort-level exceptions, the value-cascade defenders were right and I underestimated how real the cascade is. Either way, if you are underwriting a data-center asset today, run your model at both ends of this range, three years and six, and look at what happens to the equity return. If the deal only works at six, you are not underwriting a data center. You are underwriting an accounting policy.

This analysis is a source-cited research summary drawn from public records, not legal advice. It can contain errors and should be verified independently before any investment decision.