What lives the spenders assign, and how they have moved
Every large spender now books servers and network equipment over five to six years. All of them lengthened those lives between 2022 and 2026. The table gives each company's current policy as its latest filing states it.
| Company | Servers and network life now | Where it stands | What it means for the reader |
|---|---|---|---|
| Microsoft | Two to six years | In force, fiscal year ended June 30, 2026 | Widest stated range among the hyperscalers; no split by asset type |
| Alphabet | Generally six years | In force, 2025 annual report | Single headline life |
| Oracle | Six years (servers and networking, within a one-to-six-year equipment class) | Confirmed in quarter ended August 31, 2026 | Longest end of the range |
| Meta | Five to 5.5 years | In force; was four to five years at end-2024 | Most servers moved to 5.5 years in 2025 |
| Amazon | Five to six years | In force; lengthened to six in 2024 | Only spender with a range created by shortening part of the fleet |
| CoreWeave | Technology equipment six years | In force, 2025 annual report | GPU cloud on the longest life |
| Nebius | Five years from January 2026 | Was 4.0 years at end-2025 | Most recent extension |
| IREN | High-performance computing hardware five years | Fiscal year ended June 30, 2026 | Bitcoin mining hardware kept at four years |
None of the five hyperscalers reports a life for GPUs as such. Their property notes group GPUs with CPUs, networking gear and, at Microsoft, software. A review of five AI-infrastructure annual reports found no line labeled GPU. Nebius is the exception. It describes GPUs as its main assets and depreciated them over four years until its January 2026 extension to five. It expected that change to cut 2026 depreciation by about $167.6 million.11,12
The stated reasons rest on the companies' own operating experience rather than on market prices for used hardware. Alphabet says it sets technical-infrastructure lives from historical asset performance, expected technology advancements and its future deployment plans. It applies any change prospectively, meaning from the date of change without restating past years. Meta's finance chief presented the extension as an efficiency gain. The reasons for each individual change, including Amazon's move in the opposite direction, appear with their effects below.13,14
What the changes did to reported profit
Each lengthening lifted first-year net income by about 4-5%, and Amazon's partial reversal took back about 1.3%. In the table, our calculation divides each disclosed effect by the same year's reported result. A compilation by WireSift independently reaches the same 4.1-4.6% range. The figures are not fully like-for-like. Microsoft, Oracle and Meta measured their effects only on the fleet at the start of the year, which leaves out equipment bought during the year of change. Amazon's 2025 figure includes that year's purchases.15
| Change (effective) | Stated reason | Measured on | Depreciation or expense effect | Net income; diluted EPS | Share of that year's results (our calculation) |
|---|---|---|---|---|---|
| Microsoft, servers and network 4 to 6 years (fiscal 2023, from July 2022) | Software investments that increased operating efficiency, and advances in technology | Carrying amount at June 30, 2022 | Operating income +$3.7 billion | +$3.0 billion; +$0.40 | 4.2% of $88,523 million operating income; 4.1% of $72,361 million net income |
| Alphabet, servers 4 to 6, certain network equipment 5 to 6 (2023) | Not stated in results release | Not stated | Depreciation -$3.9 billion | +$3.0 billion; +$0.24 | 4.6% of $84,293 million operating income |
| Oracle, servers 4 to 5 (fiscal 2023) | Not stated | Carrying value at May 31, 2022 | Operating expenses -$434 million | Not disclosed | - |
| Oracle, servers and networking 5 to 6 (fiscal 2025) | Not stated | Carrying value at May 31, 2024 | Operating expenses -$733 million | +$573 million; +$0.20 | 4.1% of $17,678 million operating income; 4.6% of $12,443 million net income |
| Amazon, servers 5 to 6 (2024) | Not stated | Not stated | Depreciation about -$3.2 billion (AEA Capital Research); Q1 2024 operating income benefit guided at about $900 million | +$2.5 billion (AEA Capital Research) | About 4.7% of $68,593 million operating income; 4.2% of $59,248 million net income |
| Meta, most servers and network to 5.5 years (2025) | Efficiency (finance chief) | Assets in service at December 31, 2024 | Depreciation -$2.92 billion | +$2.59 billion; +$1.00 | 4.3% of $60.46 billion net income |
| Amazon, subset of servers and networking 6 to 5 (2025) | Faster pace of technology development, particularly AI and machine learning | Assets at December 31, 2024 plus 2025 purchases | Depreciation and amortization +$1.4 billion, mainly AWS | -$1.0 billion; -$0.10 | 1.75% of $79,975 million operating income; 1.3% of $77,670 million net income |
| CoreWeave, computing equipment 5 to 6 (2023) | Advances in hardware performance, software optimization and data-center design | Equipment on the balance sheet at December 31, 2023 | Total expenses -$20 million | +$0.10 per share | - |
| Nebius, servers and network 4 to 5 (January 2026) | Updated information and usage patterns | Assets in service at December 31, 2025 | Q1 2026 depreciation -$43.1 million | Q1 2026 +$41.6 million | Q1 operating loss of $128.0 million would have been about $171 million on four years, 34% larger |
Amazon's reversal came with a second cost. Mainly in the fourth quarter of 2024, it decided to retire some servers and networking gear early. It recorded about $920 million of accelerated depreciation and related charges in that quarter, and expected a further hit of about $0.6 billion to 2025 operating income. It attributed both moves to the pace of AI and machine-learning development.32
Why the one-year figures understate today's stakes
The life assumption now governs far more money than when the changes were made. One analysis estimates that the five hyperscalers' FY2025 operating income of $439bn would have been $395bn had they kept the pre-2022 four-year life. That is an uplift of $44bn, or 11%. It assumes a full year's depreciation in the purchase year and zero residual value. This cumulative figure, not the 4-5% one-year effects, is the right measure of where reported profit stands now.33,34
Capital spending is running three to seven times depreciation. Most of that spending has not yet reached the income statement, and the useful-life choice decides how fast it will.
| Company | Period | Capital spending | Depreciation | Ratio (our calculation) | What it signals |
|---|---|---|---|---|---|
| Alphabet | Nine months 2026 | $132.0 billion | $22.0 billion | About 6.0x (4.2x in nine months 2025) | Expects to significantly increase 2026 infrastructure investment |
| Amazon | Nine months 2026 | $172.2 billion cash capex | $41,845 million (AWS $24,583 million, up from $14,882 million) | About 4.1x | Expects capex to meaningfully increase in 2027; depreciation also covers fulfillment assets |
| Microsoft | Fiscal 2026 | $115,948 million additions to property and equipment | $34.3 billion (fiscal 2025 $22.0 billion; fiscal 2024 $15.2 billion) | About 3.4x | A further $26.7 billion of purchases sat in accounts payable at year-end |
| Oracle | Fiscal 2026 | $55.7 billion | $7.6 billion | About 7.3x | Gross computer and network equipment rose from $30,345 million to $59,634 million; construction in progress from $16,510 million to $39,973 million |
| CoreWeave | First half 2026 | $14.1 billion cash paid | $2.5 billion | About 5.6x | Expects to increase investment relative to 2025 |
Microsoft depreciation expense ($ billion) by year
Source: Microsoft Corporation Form 10-K For the Fiscal Year Ended June 30, 2026 (30 Jun 2026)
Meta shows how much depreciation is still to come. In the first nine months of 2026, its gross servers and network assets rose from $98,040 million to $133,626 million, up 36% on our calculation. Construction in progress, which is not yet depreciated, rose from $50,521 million to $89,499 million. Meta guides 2026 capital spending at about $135 billion to $145 billion. Depreciation starts only when assets enter service and is then spread over years, so it lags spending: Alphabet more than doubled first-half capital spending while its depreciation grew 43%. In February 2025, Bank of America strategists estimated that the spending would be a 1.6 percentage point drag on margins for earnings before interest and taxes in 2026, compared with the fourth quarter of 2024.45,46,47,48
How much profit would change under 3, 4, 5 or 6 years
On current depreciation run-rates, moving from about six years to four would cut operating income by roughly 5-11% at Alphabet and Microsoft and about 9% at Meta. Three years would roughly double those cuts, to around 10-22%. Oracle and CoreWeave are far more sensitive, because depreciation is large relative to their operating income.
Our method uses straight-line depreciation with zero salvage value. On that basis, $1 billion of equipment costs $333 million a year over three years, $250 million over four, $200 million over five and $167 million over six. We scale each company's current depreciation by its current life divided by the alternative life. That is exact for a fleet in which no asset has yet passed the shorter life. Fast-growing fleets come close to this. The three-year results overstate somewhat, because some older equipment would already be fully depreciated.49
In the table, each cell shows the change in annual pre-tax operating income, in dollars and as a percentage of the operating income shown. A negative figure means lower reported profit. Microsoft's depreciation base is the fiscal 2026 figure given above, and Oracle's six-year life is the one in its latest filing.
| Company | Depreciation base used (our estimate) | Compared with | 6 years | 5 years | 4 years | 3 years | How firm |
|---|---|---|---|---|---|---|---|
| Meta | Server and network depreciation of $14.34 billion for nine months 2026, about $19.1 billion annualized, mostly at 5.5 years | 2025 operating income $83.28 billion | +$1.6 billion (+1.9%) | -$1.9 billion (-2.3%) | -$7.2 billion (-8.6%) | -$15.9 billion (-19.1%) | Firmest: server depreciation is disclosed |
| Alphabet | About $12.7 billion: 60% of 2025 depreciation of $21.1 billion, at six years | 2025 operating income $129,039 million | Current | -$2.5 billion (-2.0%) | -$6.3 billion (-4.9%); upper bound -$10.6 billion (-8.2%) | -$12.7 billion (-9.8%); upper bound -$21.1 billion (-16.4%) | Proxy |
| Microsoft | Total fiscal 2026 depreciation of $34.3 billion, treated as six-year | Fiscal 2026 operating income $155,237 million | Current | -$6.9 billion (-4.4%) | -$17.2 billion (-11.0%) | -$34.3 billion (-22.1%) | Upper bound |
| Oracle | Q1 fiscal 2027 depreciation of $3.2 billion, treated as six-year | Q1 fiscal 2027 operating income $6,728 million | Current | -$0.64 billion (-9.5%) | -$1.6 billion (-24%) | -$3.2 billion (-48%) | Upper bound; one quarter |
| CoreWeave | Technology equipment of $33,823 million at June 30, 2026, over six years: about $5.6 billion a year | First-half 2026 operating loss $193 million; Q2 2026 revenue $2.6 billion; Q2 adjusted operating income $128 million | Current | -$1.1 billion a year | -$2.8 billion a year | -$5.6 billion a year | Large charge against a loss-making base |
The Meta row is the firmest, because Meta discloses server depreciation separately. Alphabet says only that about 60% of its technical-infrastructure assets are servers and network equipment. Servers depreciate faster than buildings, so their true share of depreciation may be higher, which is why the table also gives an upper bound. Microsoft's server line also carries software on a three-year life, and servers on lives from two to six years. Dividing that line by any single life "produces a fake depreciation number", so we treat total depreciation as an upper bound. Oracle's figure likewise includes buildings. All figures are pre-tax; the effect on EPS is smaller after tax.61
Outside estimates sit at the low end of these ranges. A three-year stress test attributed to J.P. Morgan, reported by INDmoney, cuts EPS and operating margins by roughly 6-8% for most hyperscalers, with Oracle the bigger exception. INDmoney's own calculation puts three-year EPS about 15% lower for Meta and about 8% lower for Microsoft and Alphabet.62,63
Product cadence and technical obsolescence
New accelerator generations now arrive about yearly, with large gains in performance per dollar and per watt. That is the strongest argument that a frontier GPU's economic life is shorter than its accounting life, even though the chip keeps working. Epoch AI finds that the performance a dollar of AI chips buys has grown about 49% a year since 2023 (90% confidence interval 36% to 66%), doubling every 1.7 years. Growth has come in spurts: nearly flat through mid-2024, then roughly doubling as Blackwell chips became most of new spending. NVIDIA's GB300 costs about 5.5 times the P100's 2016 launch price but delivers roughly 200 times the performance. These are peak spec-sheet figures, and realized gains on real workloads can differ substantially.64,65,66
Goldman Sachs' Global Institute describes the core tension. Rapid gains in performance per dollar push companies to replace hardware quickly, while the widening range of AI tasks keeps older chips valuable, and NVIDIA's annual release cadence sharpens the tension. Research Affiliates frames the test economically. Older chips rarely fail physically; they become obsolete economically. The question is whether a chip still earns enough to cover depreciation, operating costs and the cost of capital.67,68
Vendors' own claims for their latest generations show the scale of the step-ups.
| Chip | Compared with | Claimed gain | Who claims it |
|---|---|---|---|
| Google TPU 8t | Ironwood TPU | Up to 2.7x training performance per dollar; up to 2x performance per watt | |
| Google TPU 8i | Ironwood TPU | Up to 80% better inference performance per dollar | |
| AWS Trainium3 (Trn3) | Trn2 UltraServers | Up to 4.4x performance; 4x performance per watt | AWS |
| NVIDIA Vera Rubin | Prior platform | 10x agentic inference performance (internal workload) | NVIDIA |
| NVIDIA Vera Rubin NVL72 | GB300 NVL72 | Over 30x throughput per megawatt; up to 45x lower cost per million tokens | NVIDIA, citing SemiAnalysis data |
Performance per watt matters most where data centers are short of power. There, an older chip that uses more electricity for less output can become uneconomic long before it stops working. That is economic obsolescence rather than physical wear, and it is the reason Amazon gave for shortening part of its server lives. Microsoft's chief executive put the same concern in terms of depreciation, even though Microsoft's book lives run to six years.73
"I didn't want to go get stuck with four or five years of depreciation on one generation."74
Satya Nadella, Microsoft chief executive
Software support ages too, but more slowly. CUDA 13, NVIDIA's latest programming toolkit, removed offline compilation and library support for the Maxwell, Pascal and Volta architectures. NVIDIA says applications built with older toolkits keep running on supported drivers. The claim that heavy use physically wears GPUs out in one to two years, three at most, rests on an unnamed Google architect quoted in a Princeton CITP blog post.75,76,77
Market evidence: rental rates, resale prices and repurposing
Older GPUs plainly stay in paid service well beyond three years, and that use rests heavily on today's shortage. CoreWeave says it recently signed an A100 contract extending into 2029; the A100 was introduced in 2020. The company says it remains largely sold out of prior generations, and that prior-generation pricing is at or above where it was years ago. In November 2025, its chief executive said a batch of 2022 H100s freed by an expired contract was rebooked at 95% of the original price. CoreWeave's investor presentation, labeled illustrative, shows V100s in market use about eight years after release. AWS still offers T4-based instances for inference, the work of running trained models to answer queries. Silicon Data, which tracks GPU prices, says A100 rental rates held up after a strong 2026 rebound and that GPU economic life "certainly doesn't appear to be 2-3 years".78,79,80,81,82,83
Rental prices fell from 2024 peaks and recovered in 2026. Published figures differ partly because they measure different things: list prices, executed trades, the cheapest offers or contract rates.
| Measure | Date | H100 ($ per GPU-hour) | A100 ($ per GPU-hour) | What it captures |
|---|---|---|---|---|
| Compute Census median on-demand | October 16, 2026 | $3.49, up 79% from $1.95 on October 1, 2025; 12-month reserved $3.26 | $1.75 (80GB) | Published offers across 29 providers |
| Ornn index, executed trades | Settled October 30, 2026 | $3.03 (SXM) | $1.08 (SXM4) | Transacted prices |
| GPU Economy reconstructed median | January 2025 to October 2026 | $7.24 to $4.25 | - | Cheapest on-demand offer per provider, 4 to 7 providers |
| SemiAnalysis one-year contracts | October 2025 to March 2026 | $1.70 to $2.35 | - | Contract rates, responding to capacity shortage |
| Research Affiliates observed rates | 2024 to early 2026 | Near $8 peaks in 2024; below $3 by late 2025; below $1 in early 2026 | - | Low end of observed hourly rates |
Median on-demand price ($ per GPU-hour) by gpu
Source: GPU Rental Price Index - Compute Census, snapshot 16 Oct 2026 (16 Oct 2026)
Resale prices look much weaker once measured against what buyers originally paid. Silicon Data finds that used H100s traded at a median of about 61% of the contemporaneous new price in 2024 and about 69% in 2025. Against the historical peak new price, however, two-year-old units trade at 30-40% and three-year-old units at 20-30%. Other quotes fall within that spread, and differ mainly in what they measure against. A marketplace that opened in July 2026 quoted used H100s at $6,000 to $22,000 against $25,000 to $40,000 new, according to Aethir, which has an interest in GPU clusters. NVIDIA, citing Silicon Data, says a six-year-old A100 is still worth a quarter of its cost. A GPU residual-value insurer's conservative band puts a used H100 at 11-19% of its original list price in 2027.89,90,91,92,93
Long paid use and a short profitable life can both be true. Research Affiliates models H100 revenue as the hourly rate times 8,760 hours times 60% utilization. Against that it sets five-year straight-line depreciation, operating expense and an 8% capital charge on book value. On falling rental rates, revenue drops below total economic cost in year four. The chip stays in service because it covers its running costs, but the productive life is "closer to three years". Goldman's reading of the same rental market is that trailing-edge A100 and H100 prices imply lives of five to six-plus years. It cautions that this may reflect today's extreme capacity constraints rather than durable value.94,95,96
GPU clouds are where the assumption bites hardest
For CoreWeave and similar operators, contracts are shorter than book lives. The six-year life therefore rests on re-contracting or redeploying hardware after the first contract ends. CoreWeave's committed contracts had a weighted-average duration of about five years at December 31, 2025. The company warns that any inability to redeploy infrastructure beyond its contracted life could significantly affect results. Management counters that initial contracts fully repay the asset-level debt and generate free cash flow, so any renewal or resale adds to returns already earned.97,98,99
Depreciation is large relative to revenue. CoreWeave's 2025 depreciation and amortization of $2.4 billion equalled about 47% of its $5,131 million revenue (our calculation), against an operating loss of $46 million. Our calculation for a five-year life: $2.4 billion × 6 ÷ 5 less $2.4 billion adds about $0.48 billion. That would have turned 2025's small operating loss into a loss of roughly $0.5 billion.100,101
Revenue ($ million) and Depreciation and amortization on property and equipment ($ million) by year
Source: CoreWeave, Inc. Annual Report on Form 10-K (31 Dec 2025)
The same pattern runs across the other GPU clouds. Nebius's depreciation and amortization rose 332% to $212.0 million in the first quarter of 2026, reaching 40% of total operating expenses. IREN warns that new GPU generations on short development cycles may shorten the economic life and lower the residual value of its equipment. It also says contracts shorter than asset lives expose it to re-contracting and pricing risk. Nebius warns of pricing pressure on older GPU generations as newer ones arrive.102,103,104,105
The bookkeeping itself carries risk at the younger operators. Nebius's auditor found a material weakness in its controls over property and equipment, meaning a control gap serious enough that a material misstatement might not be caught. The weakness includes how depreciation start dates are set. CoreWeave reports that its disclosure controls were not effective at year-end 2025 because of material weaknesses, but says these did not result in a material misstatement.106,107,108,109
What analysts, critics and auditors say
No one credible alleges that these changes break US accounting rules; the dispute is over how much they matter. A useful-life change is a change in accounting estimate under US GAAP (ASC 250-10-45-17). It is applied prospectively, without restating earlier periods. As a result, earnings after a change cannot be compared like-for-like with years before it. An auditor's report must flag a change in accounting principle or a correction of a misstatement, and a change in estimate is neither. Instead, auditing standard AS 2501 requires auditors to check that significant assumptions are consistent with existing market information and recent experience. SEC guidance says a favorable estimate change should be disclosed so readers do not mistake it for operational improvement.110,111,112,113,114,115
How far the estimate can stretch shows in one quarter: Meta lengthened lives while Amazon shortened part of its fleet. Both changes came from formal useful-life studies, were reviewed by audit committees and were concurred in by the auditors. Buxton Helmsley concludes that the resulting difference in operating income "is not driven by economics - it is driven by the elasticity of the estimate."116,117
The loudest critic's number measures something different from the disclosed effects. Starting in November 2025, Michael Burry estimated that five companies would understate depreciation by about $176 billion over 2026-2028. On his figures, Oracle's 2028 earnings would be overstated by 26.9% and Meta's by 20.8%. That is a forward projection on a much larger future fleet. The 4-5% figures are one-year effects on existing fleets, and the 11% is the cumulative 2025 level. The 11% is best read as where profit stands now, and Burry's figure as one bearish scenario. A forensic accounting firm declines to endorse his figures without a model tying hardware vintages to each company's depreciation schedules.118,119
| Voice | Position | Figure | How to weigh it |
|---|---|---|---|
| Barclays (as reported by a Princeton CITP blog post) | Accounting lives run too long | Cutting AI firms' 2025 earnings forecasts by up to 10% | Second-hand report |
| NVIDIA (memo to analysts) | Lives are grounded in real longevity | Customers depreciate GPUs over four to six years | Vendor with a commercial interest |
| Andy Jassy, Amazon chief executive (shareholder letter, as reported by Fortune) | Current lives are reasonable | AWS chips, servers and networking last 5-6 years | Management view, despite Amazon's partial shortening |
| Defenders of longer lives | Software gains and inference reuse extend life | Servers often outlasted three-to-four-year assumptions | Rests on pre-AI-cycle experience |
| Columnist (via Tiger Brokers) | Depreciation is not the real AI risk | The debate "won't move the needle" | Opinion |
Whatever life is chosen, it changes only accrual measures. EBITDA excludes depreciation and is unaffected, and so is cash flow. US tax depreciation is set separately: Amazon's taxable income, for example, is reduced by accelerated tax depreciation. Operating margins, price-to-earnings multiples and returns on capital all move with the useful-life assumption. Comparing earnings with capital spending and free cash flow is therefore the cleaner check.125,126,127
What would change the answer
The clearest sign that lives have run too long would be early write-offs, of the kind Amazon took in late 2024. Depreciation deferred by a long life does not disappear. It can return abruptly through higher depreciation, accelerated depreciation or impairments, which are write-downs of an asset's book value. Microsoft and Meta both warn that underused or excess infrastructure could lead to impairments.128,129,130,131
| Signal | Points toward | What to watch |
|---|---|---|
| Accelerated-depreciation or early-retirement charges | Shorter lives | Quarterly filings, especially as new chip generations ship |
| Further shortenings like Amazon's 2025 subset | Shorter lives | Changes in estimate in annual reports |
| Prior-generation rental and re-contract prices falling as new supply arrives | Shorter lives, impairment risk | Rental indices; GPU cloud renewal pricing |
| Older-generation prices holding at or above year-ago levels; inference demand absorbing retired fleets | Current lives hold | CoreWeave says older-generation prices remain at or above levels of about a year ago, and expects about $250 million of managed-inference annual recurring revenue by year-end |
| EBITDA per dollar of gross equipment | Whether returns keep pace with the life assumption | At CoreWeave, on one analyst's measure, the quarterly trend ran 12.5% to 11.7% to 8.6% to 9.8% |
| Adverse technological or market change; evidence of obsolescence | Impairment testing | Listed among the triggers for impairment testing in the international standard, IAS 36 |
The swing factor is whether shortage-era pricing persists. Research Affiliates notes that spot rental prices can overshoot in both directions. It adds that hyperscalers may achieve higher utilization or lower financing costs than third-party renters. Its point is not that GPUs fail after three years, but that their ability to earn excess returns may fade before their accounting life ends.136
What the sources gathered do not establish
- The useful lives and depreciation of GPUs themselves at Microsoft, Alphabet, Amazon, Meta and Oracle. Without these, company-level GPU sensitivity can only be approximated, and the true cuts could sit nearer the lower or upper bounds above.
- Whether Microsoft's latest annual report records any estimate change or rationale beyond its two-to-six-year policy. A further change would alter Microsoft's row in the sensitivity table.
- A vintage-by-vintage sensitivity built on each company's disclosed server and GPU balances. This would tighten the run-rate ranges, especially the three-year cases.
- Any sensitivity disclosed by the companies themselves. None quantifies how profits would change under alternative lives.
- Hyperscalers' actual retirement ages and realized resale proceeds for AI hardware, and how large a part of its fleet Amazon moved to five years. These would show whether current lives are being borne out.
- The model behind the $176 billion understatement estimate. Without it, that scenario cannot be tested against filed asset vintages.
- Whether today's compute shortage will persist as new chip generations ship. This decides whether five-to-six-year lives hold or give way to shortening and impairments.