GPU Depreciation Lives and Their Effect on Reported AI-Sector Profits

1. Overview: the useful-life lever

The single largest discretionary input into the reported profits of the five biggest AI-infrastructure spenders is not a market price but an accounting estimate: how many years a server is assumed to last. Between 2020 and 2024, Microsoft, Alphabet, Amazon and Oracle each extended the stated useful lives of servers and network equipment from roughly 3 years to 6 years; Meta extended the same assets to a 4–5-year range by 2024, reaching 5.5 years only in a further change effective January 2025. The cumulative net-income benefit those companies themselves disclosed for the extensions is roughly $19.4bn (sum of the per-change impacts filed under ASC 250 and itemised in section 2). Over the same period their combined capital expenditure rose roughly 2.6-fold, from about $100bn to about $259bn per year, meaning each year of assumed life now spreads a far larger pool of capitalised cost.

This report is strictly descriptive. It documents what the filings say, quantifies what stated net income would look like under compressed lives of 5, 4 and 3 years (with a 2-year outer bound), and sets the spending against the one historical benchmark of comparable scale, the 1996–2000 telecom buildout. It offers no view on valuations and no verdict on whether current lives are right — a question the evidence, presented in section 4, does not settle. Four charts accompany the text.

2. Stated useful lives: what the 10-Ks disclose

Microsoft (fiscal year ends 30 June) depreciated servers over 3 years and network equipment over 2 as of FY2020. Effective July 2020, both moved to 4 years; effective July 2022, both moved to 6 years — the disclosed operating-income and net-income impacts of each change are itemised in the table below.^1,2

Alphabet moved servers from 3 to 4 years and network equipment from 3 to 5 years effective January 2021, then extended both to 6 years effective January 2023 — the disclosed depreciation and net-income impacts of each change are itemised in the table below, including the $988m less depreciation and $770m more net income recorded in Q1 2023 alone.^3,4,5

Amazon made three extensions and then the sector's only reversal. Servers went from 3 to 4 years in 2020, to 5 years (networking 5→6) in January 2022, and to 6 years in January 2024 — impacts of each change are itemised in the table below. Then, effective 1 January 2025, Amazon cut a subset of server and networking lives back to 5 years, citing the increased pace of AI-driven technology development.^6,7,8

The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.

Amazon Form 10-Q, Q1 2025

⁹

The reversal cut net income by $677m in the nine months to 30 September 2025 ($889m of additional D&A). Amazon also recorded roughly $920m of accelerated depreciation and related charges in the quarter ended 31 December 2024 for servers and networking equipment it decided to retire early, with a further ~$0.6bn expected in 2025 — partly offset by extending heavy equipment from 10 to 13 years (+$0.9bn expected 2025 net income).^10,11

Meta moved servers and network assets from 3 to 4 years in 2021, then to 5 years for affected assets in mid-2022, producing a blended 4.5-year life. In January 2025 it extended certain servers and network assets to 5.5 years, expecting the change to "reduce our full year 2025 depreciation expense by approximately $2.9 billion."^12,13,14

Oracle (fiscal year ends 31 May) extended servers and networking equipment from 5 to 6 years effective the start of FY2025, cutting operating expenses by $733m and lifting net income by $573m ($0.21 basic EPS).¹⁵

Disclosure has thinned just as the numbers grew. Microsoft's FY2025 10-K no longer states a server life at all, collapsing the category into "computer equipment" with a 2–6-year range, and Alphabet's FY2024 10-K dropped the separate "Information technology assets" line from its PP&E note in favour of a combined "technical infrastructure" category that also includes land and buildings. This regression arrived just as the underlying asset bases reached record size — Microsoft's net PP&E stood at $205.0bn at 30 June 2025 and Alphabet's at $171.0bn at 31 December 2024.^16,17

Oracle is omitted from the chart because comparable useful-life disclosures are only available from FY2023 onward (5 years in FY2023, 6 years in FY2024 and FY2025, per its 10-K/10-Q filings on SEC EDGAR); plotting it alongside the full FY2020–FY2025 series would require data points the filings do not provide.

3. The capex–depreciation gap

The five companies' combined capital expenditure rose from roughly $100bn to roughly $259bn between their most recent fiscal years and the years four earlier — about 2.6 times — with every company contributing: Microsoft $20.6bn (FY2021) to $64.6bn (FY2025), Alphabet $22.3bn to $52.5bn, Amazon $40.1bn to $83.0bn, Meta $15.1bn to $37.3bn, and Oracle $2.1bn to $21.2bn.¹⁸

Depreciation has not kept pace, by construction. Bucketing the five companies by comparable periods, combined capex went from $132.7bn (2021 bucket) to $258.6bn (2024 bucket) while combined D&A as filed went from $70.4bn to $121.6bn — a capex-over-D&A gap that widened from roughly $62bn to roughly $137bn — against operating cash flow of $294bn rising to $489bn (Microsoft and Oracle fiscal years mapped to the calendar bucket containing their midpoints; D&A definitions vary by company as filed). Calendar-2025 actuals extend the trend: Alphabet $91.4bn, Amazon $131.8bn, and Meta $72.2bn including finance-lease principal. Microsoft reports only fiscal years (FY2025: $64.6bn).^19,20,21

The mix matters as much as the level. CRE Finance Council notes the composition of this spending is shifting toward shorter-lived assets — servers, GPUs, networking — rather than building shells. The mechanism is simple: capitalised spending reaches the income statement only as depreciation over the assumed life. With a ~$137bn annual wedge of spending not yet expensed, the life assumption controls more prospective earnings than at any point in these companies' histories. That is why the estimate itself, examined next, has become contested ground.²²

4. How fast does AI hardware actually age?

The cleanest evidence that useful life is a management estimate rather than an observable fact is the January 2025 divergence. In the same month, Amazon shortened a subset of server lives from 6 to 5 years, citing the pace of AI-driven technology development, while Meta extended certain servers to 5.5 years. Both changes were characterised by management as products of formal useful-life studies, reviewed by audit committees, under the same accounting standard. Two studies, one month, opposite answers. Oracle (fiscal year ends 31 May) had separately moved in the extending direction some seven months earlier, taking servers and networking equipment from 5 to 6 years effective the beginning of its fiscal 2025, around June 2024.^23,24,25,26

The public dispute is anchored by Michael Burry, who wrote in November 2025 that extending useful lives against Nvidia's "2-3 yr product cycle" would understate depreciation by about $176bn over 2026–28, overstating Oracle's 2028 earnings by ~27% and Meta's by ~21%.²⁷

Nvidia's rebuttal memo to sell-side analysts countered that customers depreciate GPUs over 4–6 years "based on real-world longevity and utilization patterns," and that A100s released in 2020 still run at high utilisation. Burry's rejoinder: "I stand by my analysis. I am not claiming Nvidia is Enron. It is clearly Cisco." The dispute is unresolved; neither side has produced auditable fleet-wide usage or retirement data.²⁸

There is also an internal tension on Nvidia's side: at its March 2025 conference, Jensen Huang reflected candidly on the prior generation of chips.

When Blackwell starts shipping in volume, you couldn't give Hoppers away… There are circumstances where Hopper is fine. Not many.

Jensen Huang, Nvidia GTC, March 2025

²⁹

The evidence for faster economic ageing is real. Nvidia now releases new AI chips annually, versus its previous two-year cadence, and AMD has followed. H100 rental rates that ran $7–10 per GPU-hour during the 2023–24 shortage had fallen to roughly $2–3.50 per hour by late 2025, with AWS cutting H100 prices ~44% in June 2025. Used H100 SXM5 cards that sold for ~$40,000 in late 2023 traded at $12,000–22,000 on secondary markets by 2026. But the ledger's own caveat applies: rental-price decline chiefly reflects supply catching up with scarcity pricing and new-generation competition, not hardware failing or becoming unusable. On reliability, the only quantified public datapoint is Meta's Llama-3 training run: 419 unexpected interruptions in 54 days on a 16,384-GPU H100 cluster, about 30% attributed to GPU failures and 17% to HBM3 memory. That is a single-cluster interruption metric, not a fleet-wide economic-life statistic, and cannot be extrapolated into one.^{30,31,32,33,34}

The evidence against is equally concrete. CoreWeave, whose business is renting GPUs, depreciates them over 6 years per its S-1; its chief executive Michael Intrator said in November 2025 that its 2020-vintage A100s are fully booked and that a tranche of H100s freed by an expiring contract was immediately re-let at 95% of the original price.³⁵

All of the data points that I'm getting are telling me that the infrastructure retains value.

Michael Intrator, CoreWeave CEO, November 2025

³⁶

He is a self-interested speaker, but the datapoints are specific. JPMorgan's own bull rationale in the same analysis that models the downside notes that "older chips remain in use for several years; even NVIDIA A100 chips still run at high utilization and generate positive margins beyond 2-3 years," and that superseded processors are repurposed for inference or resold. V100s — 2017 silicon — still rented at a median ~$1.29/hour across 19 providers in mid-2026. And American Compute, drawing on 76,775 secondary-market transactions covering 622,098 units since 2023, argues a GPU's useful life can reach 8 years with residuals above 10% of cost at year five — though as a GPU-financing business it has a direct financial interest in high residuals.^37,38,39

JPMorgan's two-sided treatment bridges to the arithmetic: applying 3-year depreciation to GPUs and networking equipment added since late 2022, it estimates EPS and operating-margin revisions of −6% to −8% for the hyperscalers, with larger declines for Oracle. The next section runs the equivalent exercise directly from the filings.⁴⁰

5. Sensitivity: net income under 6-, 5-, 4- and 3-year lives

The method is an illustrative steady-state approximation, not a GAAP restatement — under ASC 250 a change in estimate applies prospectively, so no company would ever report these numbers. For each company, take the most granular disclosed gross PP&E category containing servers, assume straight-line depreciation to zero salvage, and approximate annual depreciation as gross base ÷ assumed life. The incremental pre-tax expense of a shorter life L′ versus the current stated life is Gross × (1/L′ − 1/L₀), taxed at the company's most-recent-year effective rate. Current lives L₀ are 6 years for Microsoft, Alphabet, Amazon and Oracle, and 5.5 for Meta. Worked example — Microsoft at 4 years: $132,836m × (1/4 − 1/6) = $11,070m pre-tax; × (1 − 0.176) = $9,121m after tax; $101,832m − $9,121m ≈ $92.7bn restated.⁴¹

The approximation is nonetheless consistent with the companies' own filed arithmetic: Microsoft's disclosed one-to-two-year extensions were worth $2.3–3.0bn of net income per change, Alphabet's $2.0–3.0bn, Amazon's $2.0–2.8bn per year of extension, and Oracle's $0.57bn — the same order of magnitude the table produces per year of life, per company.

The headline: at 4-year lives, combined net income falls roughly $37bn (−11%); at 3 years, roughly $75bn (−22%), with Amazon (−53%, as approximated) and Oracle (−36%) most exposed. At 2-year lives — the outer bound matching claims that an annual product cadence makes hardware economically stale in two years — the same method with the same caveats yields a combined after-tax increment of roughly $150bn, cutting combined net income by about 45% and pushing Amazon's restated figure negative. Burry's $176bn cumulative, pre-tax estimate of understated depreciation over 2026–28, whose derivation he has not fully disclosed, is close to this model's pre-tax combined increment at a 2-year life for a single year (≈$176bn) — but the two are not directly comparable (a cumulative three-year figure versus a single steady-state year; pre-tax versus the after-tax range quoted above), so the proximity is a rough coincidence, not a validation.⁴²

6. Macro scale: AI capex against the telecom precedent

The nearest historical analogue by scale is the telecom and fibre overbuild. US telecom capex peaked at roughly $120bn nominal in 2000 (~$213bn inflation-adjusted), or 1.0–1.2% of GDP, with more than $500bn spent cumulatively over 1996–2000, most of it debt-financed.^43,44,45

Estimates of 2025 AI-related capex as a share of US GDP range from 1.28% (Magnificent-7 scope, Q2 2025 annualised, 7GC) through 1.4% (data-centre capex, Apollo) to about 1.6% (broader AI capex, Goldman Sachs via secondary reporting) — the spread reflects differing scopes, not a resolvable disagreement, and is presented here as a range. Apollo forecasts the data-centre share rising from 1.4% in 2025 to 3.1% by 2027.^46,47,48

Official data corroborate the scale. The St. Louis Fed, using BEA and Census series, finds information-processing-equipment investment contributed 0.90 percentage points to real GDP growth in Q1 2025 — more than two standard deviations above its long-run average — and software investment contributed 0.57 points in Q2 2025, a historical peak against prior highs of 0.44 points in 1997 and 0.49 in 2021.⁴⁹

Fed Board staff put nominal spending on data-centre structures at ~$10bn in 2021 rising to ~$40bn in the first half of 2025. As a rough approximation mixing fiscal-year company figures against a quarterly annualised denominator, the five companies' combined ~$258.6bn capex equates to about 5.9% of US private nonresidential fixed investment ($4,364bn SAAR, Q3 2025).^50,51

Bond issuance signals the capital commitment. Meta sold $30bn of public bonds on 30 October 2025 — the year's biggest corporate offering, against $125bn of orders — separate from the $27bn of debt (plus $2.5bn equity) in its Blue Owl Hyperion data-centre SPV, in which Meta retains 20%. Oracle sold an $18bn six-tranche bond in September 2025, upsized from $15bn on ~$88bn of peak demand, separate from Stargate-related project finance including a $38bn JPMorgan-led facility and an $18bn New Mexico mini-perm loan. Hyperscalers issued $121bn of bonds in 2025 against a five-year average of $28bn, per one analyst compilation. Data-centre ABS and CMBS issuance in 2025 ran between ~$21bn (through Q3) and ~$25–30bn (preliminary full-year), depending on source and cutoff. Morgan Stanley estimates $2.9tn of global data-centre capex over 2025–28, of which the four biggest hyperscalers self-fund ~$1.4tn, leaving a ~$1.5tn financing gap ($800bn private credit, $200bn corporate bonds, $150bn securitised). These are scale comparisons only; this report draws no conclusion about whether the spending is over- or under-built.^{52,53,54,55,56,57,58}

7. What the numbers do and do not say

What the record shows is narrow but firm: reported AI-sector earnings are highly sensitive to an estimate no outsider can observe, the disclosed extensions of that estimate added roughly $19.4bn to net income across five companies, and in January 2025 two formal useful-life studies at two hyperscalers reached opposite conclusions in the same month. Meanwhile disclosure is moving the wrong way for anyone wishing to check: only Meta still discloses a server-specific gross asset base ($68.4bn at 31 December 2024), Microsoft has folded servers into a 2–6-year "computer equipment" range, and Alphabet has dropped its IT-assets line.

The limitations are equally firm. No fleet-wide GPU failure or retirement data exists in the public record — the Llama-3 cluster figure is a single 54-day snapshot. The sensitivity in section 5 is a steady-state approximation with named directional biases, not a restatement. And accounting depreciation is not economic depreciation; this report measures only the first. What would settle the question is disclosure the companies do not currently provide: server-specific asset bases and lives, retirement and redeployment schedules, and utilisation data for older fleets. Until then, the difference between a 3-year and a 6-year assumption — worth roughly $75bn a year of combined reported profit on this arithmetic — remains a judgement taken inside the companies, disclosed only when it changes. This report offers no investment conclusion.