Skip to main content

Deeperer editorial

Published

Changed: 21 sections, 65 sources

71 min read

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

The largest AI infrastructure spenders now depreciate servers and networking gear over five to six years, up from four (or five) years before 2022. Every lengthening raised reported profit: in the first year, net income rose by about 4-5% at Microsoft, Alphabet, Oracle, Meta and Amazon. Those one-year disclosures understate today's stakes, because capital spending now runs three to seven times depreciation. One estimate puts the combined 2025 uplift for the five, against four-year lives, at $44 billion, or 11% of operating income. On straight-line arithmetic, a three-year life would cut operating income by roughly a tenth to a fifth at Alphabet, Meta and Microsoft. It would cut up to about half of Oracle's quarterly operating income, as an upper bound, and add more than $5 billion a year of depreciation at CoreWeave, which is near operating breakeven. Cash flow, EBITDA and cash tax would not change. Whether five to six years is right is genuinely unsettled. Older chips keep earning: A100s are contracted into 2029, and H100 rental prices rose sharply in 2026. But resale values fall to 20-40% of peak price by years two to three, and chip performance per dollar improves about 49% a year. Lives that today's compute shortage supports could later give way to shortening, accelerated depreciation or impairments, and Amazon has already moved part of its fleet that way.

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.

CompanyServers and network life nowWhere it standsWhat it means for the reader
MicrosoftTwo to six yearsIn force, fiscal year ended June 30, 2026Widest stated range among the hyperscalers; no split by asset type
AlphabetGenerally six yearsIn force, 2025 annual reportSingle headline life
OracleSix years (servers and networking, within a one-to-six-year equipment class)Confirmed in quarter ended August 31, 2026Longest end of the range
MetaFive to 5.5 yearsIn force; was four to five years at end-2024Most servers moved to 5.5 years in 2025
AmazonFive to six yearsIn force; lengthened to six in 2024Only spender with a range created by shortening part of the fleet
CoreWeaveTechnology equipment six yearsIn force, 2025 annual reportGPU cloud on the longest life
NebiusFive years from January 2026Was 4.0 years at end-2025Most recent extension
IRENHigh-performance computing hardware five yearsFiscal year ended June 30, 2026Bitcoin mining hardware kept at four years

1,2,3,4,5,6,7,8,9,10

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 reasonMeasured onDepreciation or expense effectNet income; diluted EPSShare 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 technologyCarrying amount at June 30, 2022Operating income +$3.7 billion+$3.0 billion; +$0.404.2% 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 releaseNot statedDepreciation -$3.9 billion+$3.0 billion; +$0.244.6% of $84,293 million operating income
Oracle, servers 4 to 5 (fiscal 2023)Not statedCarrying value at May 31, 2022Operating expenses -$434 millionNot disclosed-
Oracle, servers and networking 5 to 6 (fiscal 2025)Not statedCarrying value at May 31, 2024Operating expenses -$733 million+$573 million; +$0.204.1% of $17,678 million operating income; 4.6% of $12,443 million net income
Amazon, servers 5 to 6 (2024)Not statedNot statedDepreciation 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, 2024Depreciation -$2.92 billion+$2.59 billion; +$1.004.3% 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 learningAssets at December 31, 2024 plus 2025 purchasesDepreciation and amortization +$1.4 billion, mainly AWS-$1.0 billion; -$0.101.75% 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 designEquipment on the balance sheet at December 31, 2023Total expenses -$20 million+$0.10 per share-
Nebius, servers and network 4 to 5 (January 2026)Updated information and usage patternsAssets in service at December 31, 2025Q1 2026 depreciation -$43.1 millionQ1 2026 +$41.6 millionQ1 operating loss of $128.0 million would have been about $171 million on four years, 34% larger

16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31

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.

CompanyPeriodCapital spendingDepreciationRatio (our calculation)What it signals
AlphabetNine months 2026$132.0 billion$22.0 billionAbout 6.0x (4.2x in nine months 2025)Expects to significantly increase 2026 infrastructure investment
AmazonNine months 2026$172.2 billion cash capex$41,845 million (AWS $24,583 million, up from $14,882 million)About 4.1xExpects capex to meaningfully increase in 2027; depreciation also covers fulfillment assets
MicrosoftFiscal 2026$115,948 million additions to property and equipment$34.3 billion (fiscal 2025 $22.0 billion; fiscal 2024 $15.2 billion)About 3.4xA further $26.7 billion of purchases sat in accounts payable at year-end
OracleFiscal 2026$55.7 billion$7.6 billionAbout 7.3xGross computer and network equipment rose from $30,345 million to $59,634 million; construction in progress from $16,510 million to $39,973 million
CoreWeaveFirst half 2026$14.1 billion cash paid$2.5 billionAbout 5.6xExpects to increase investment relative to 2025

35,36,37,38,39,40,41,42,43,44

Microsoft depreciation expense ($ billion) by year

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.

CompanyDepreciation base used (our estimate)Compared with6 years5 years4 years3 yearsHow firm
MetaServer and network depreciation of $14.34 billion for nine months 2026, about $19.1 billion annualized, mostly at 5.5 years2025 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
AlphabetAbout $12.7 billion: 60% of 2025 depreciation of $21.1 billion, at six years2025 operating income $129,039 millionCurrent-$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
MicrosoftTotal fiscal 2026 depreciation of $34.3 billion, treated as six-yearFiscal 2026 operating income $155,237 millionCurrent-$6.9 billion (-4.4%)-$17.2 billion (-11.0%)-$34.3 billion (-22.1%)Upper bound
OracleQ1 fiscal 2027 depreciation of $3.2 billion, treated as six-yearQ1 fiscal 2027 operating income $6,728 millionCurrent-$0.64 billion (-9.5%)-$1.6 billion (-24%)-$3.2 billion (-48%)Upper bound; one quarter
CoreWeaveTechnology equipment of $33,823 million at June 30, 2026, over six years: about $5.6 billion a yearFirst-half 2026 operating loss $193 million; Q2 2026 revenue $2.6 billion; Q2 adjusted operating income $128 millionCurrent-$1.1 billion a year-$2.8 billion a year-$5.6 billion a yearLarge charge against a loss-making base

50,51,52,53,54,55,56,57,58,59,60

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.

ChipCompared withClaimed gainWho claims it
Google TPU 8tIronwood TPUUp to 2.7x training performance per dollar; up to 2x performance per wattGoogle
Google TPU 8iIronwood TPUUp to 80% better inference performance per dollarGoogle
AWS Trainium3 (Trn3)Trn2 UltraServersUp to 4.4x performance; 4x performance per wattAWS
NVIDIA Vera RubinPrior platform10x agentic inference performance (internal workload)NVIDIA
NVIDIA Vera Rubin NVL72GB300 NVL72Over 30x throughput per megawatt; up to 45x lower cost per million tokensNVIDIA, citing SemiAnalysis data

69,70,71,72

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.

MeasureDateH100 ($ per GPU-hour)A100 ($ per GPU-hour)What it captures
Compute Census median on-demandOctober 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 tradesSettled October 30, 2026$3.03 (SXM)$1.08 (SXM4)Transacted prices
GPU Economy reconstructed medianJanuary 2025 to October 2026$7.24 to $4.25-Cheapest on-demand offer per provider, 4 to 7 providers
SemiAnalysis one-year contractsOctober 2025 to March 2026$1.70 to $2.35-Contract rates, responding to capacity shortage
Research Affiliates observed rates2024 to early 2026Near $8 peaks in 2024; below $3 by late 2025; below $1 in early 2026-Low end of observed hourly rates

84,85,86,87,88

Median on-demand price ($ per GPU-hour) by gpu

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

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

VoicePositionFigureHow to weigh it
Barclays (as reported by a Princeton CITP blog post)Accounting lives run too longCutting AI firms' 2025 earnings forecasts by up to 10%Second-hand report
NVIDIA (memo to analysts)Lives are grounded in real longevityCustomers depreciate GPUs over four to six yearsVendor with a commercial interest
Andy Jassy, Amazon chief executive (shareholder letter, as reported by Fortune)Current lives are reasonableAWS chips, servers and networking last 5-6 yearsManagement view, despite Amazon's partial shortening
Defenders of longer livesSoftware gains and inference reuse extend lifeServers often outlasted three-to-four-year assumptionsRests on pre-AI-cycle experience
Columnist (via Tiger Brokers)Depreciation is not the real AI riskThe debate "won't move the needle"Opinion

120,121,122,123,124

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

SignalPoints towardWhat to watch
Accelerated-depreciation or early-retirement chargesShorter livesQuarterly filings, especially as new chip generations ship
Further shortenings like Amazon's 2025 subsetShorter livesChanges in estimate in annual reports
Prior-generation rental and re-contract prices falling as new supply arrivesShorter lives, impairment riskRental indices; GPU cloud renewal pricing
Older-generation prices holding at or above year-ago levels; inference demand absorbing retired fleetsCurrent lives holdCoreWeave 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 equipmentWhether returns keep pace with the life assumptionAt 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 obsolescenceImpairment testingListed among the triggers for impairment testing in the international standard, IAS 36

132,133,134,135

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.

Sources (136)

  1. 1
    www.sec.gov
    Property and equipment is stated at cost less accumulated depreciation and depreciated using the straight-line method over the shorter of the estimated useful life of the asset or the lease term. The estimated useful lives of our property and equipment are generally as follows: software developed or acquired for internal use, three years ; servers and network equipment, two to six years ; buildings and improvements, five to 15 years ; leasehold improvements, three to 15 years ; and furniture and equipment, one to 10 years . Land is not depreciated.
  2. 2
    www.sec.gov
    Property and Equipment Property and equipment is comprised of technical infrastructure, office space, corporate and other assets currently in service, and assets not yet in service. Technical infrastructure includes data center land, buildings and leasehold improvements, and servers and network equipment. Office space includes office land, buildings, and leasehold improvements. Assets not yet in service are those that are not ready for their intended use, including data center buildings and servers in the process of construction or assembly. Property and equipment are stated at cost less accumulated depreciation. Depreciation commences once assets are ready for their intended use and is recorded using the straight-line method over the estimated useful lives of the assets, which we regularly evaluate for factors such as technological obsolescence and our planned use and utilization. We depreciate data center and office buildings over periods of seven to 40 years. We depreciate servers and network equipment generally over a period of six years . We depreciate corporate and other assets over periods of two to 25 years. We depreciate leasehold improvements over the shorter of the remaining lease term or the estimated useful lives of the assets. Land is not depreciated.
  3. 3
    www.sec.gov
    (1) Comprised primarily of servers and networking equipment with estimated useful life of six years .
  4. 4
    www.sec.gov
    Property and Equipment Useful Life/ Amortization period Servers and network assets Five to 5.5 years Buildings 25 to 30 years Equipment and other One to 25 years Finance lease right-of-use assets Five to 20 years Leasehold improvements Lesser of estimated useful life or remaining lease term
  5. 5
    www.sec.gov
    Property and equipment, including finance leases, are depreciated and stated at cost less accumulated depreciation. Depreciation is computed using the straight-line method over the estimated useful lives of the assets or the remaining lease term, whichever is shorter. The estimated useful lives of property and equipment and amortization periods of finance lease right-of-use (ROU) assets as of December 31, 2024 are described below: Property and Equipment Useful Life/ Amortization period Servers and network assets Four to Five years (1) Buildings 25 to 30 years Equipment and other One to 25 years Finance lease right-of-use assets Five to 20 years Leasehold improvements Lesser of estimated useful life or remaining lease term _______________________ (1) Effective January 2025, the useful lives of certain servers and network assets are extended to 5.5 years.
  6. 6
    www.sec.gov
    The estimated useful lives as of December 31, 2025, are as follows: Property and equipment Estimated useful life Buildings Lesser of forty years or the remaining life of the underlying building Servers and networking equipment Five to six years (1)
  7. 7
    www.sec.gov
    Effective January 1, 2024, we changed our estimate of the useful lives for our servers from five to six years
  8. 8
    www.sec.gov
    The estimated useful lives of the Company's property and equipment are as follows: Technology equipment 6 years Software 3 - 6 years Data center equipment and leasehold improvements Shorter of remaining lease term or estimated useful life of up to 12 years Furniture, fixtures, and other assets 3 - 5 years
  9. 9
    www.sec.gov
    Property and equipment are recorded at cost and depreciated using the straight-line method over their useful lives. Capital expenditures incurred before property and equipment are ready for their intended use are capitalized as assets not yet in use. The depreciable amount of property and equipment is its cost less its residual (salvage) value (if applicable). Depreciation is computed under the straight-line method using estimated useful lives as follows: Estimated useful lives Server and network equipment 4.0 years Infrastructure systems and equipment 3.0 - 10.0 years Office furniture and equipment 3.0 years Buildings 20.0 years Other property and equipment 2.0 - 10.0 years Land is not depreciated. Depreciation of assets included in assets not yet in use commences when they are ready for the intended use.
  10. 10
    www.sec.gov
    The estimated useful lives of the Group’s property, plant and equipment are generally as follows: Useful life (in years) Buildings 20 Other PPE 3 - 10 Mining hardware 4 HPC hardware 5 Leasehold improvements Lesser of the life of the lease or the useful life of the improvement
  11. 11
    gpuinsights.net
    None of these lines is labeled “GPU.”
  12. 12
    www.sec.gov
    Our main assets, GPUs, are currently depreciated over four years. In January 2026, the Company completed an assessment of the useful lives of servers and network equipment based on updated information and usage patterns obtained subsequent to year-end and concluded that the estimated useful lives of such assets should be extended from four to five years. Management is considering the application of this change in accounting estimate prospectively beginning January 1, 2026. Based on the servers and network equipment placed in service as of December 31, 2025, we expect this change in accounting estimate will reduce the depreciation expenses for fiscal year 2026 by approximately $167.6 million. The actual impact in future periods will depend on the timing and level of future capital expenditures and asset retirements. We will continue to evaluate the appropriateness of our estimated useful lives and may revise them in future periods if warranted. Any such revisions will be accounted for prospectively.
  13. 13
    www.sec.gov
    We assess the reasonableness of the useful lives of our property and equipment periodically or when events indicate a change is necessary. To determine the useful lives of our technical infrastructure, we rely on multiple inputs, including historical asset performance, expected technology advancements, and our future infrastructure deployment plans. Any change in the estimated useful lives is recognized on a prospective basis.
  14. 14
    www.accountingtoday.com
    Meta's chief financial officer, Susan Li, said on the most recent earnings call that the company is making efficiency gains "by extending the useful lives of our servers and associated networking equipment."
  15. 15
    wiresift.com
    Five companies, five accounting teams, one number: ~4 to 5% of profit. In the year each company extended its server lives, the change lifted net income by a strikingly consistent ~4 to 4.6%: Microsoft +$3.0B (4.1%), Alphabet +$3.0B (4.1%), Amazon +$2.5B (4.2%), Meta +$2.59B (4.3%), Oracle +$0.57B (4.6%). Real money, and a real single-year boost to reported earnings per share. But notice what that is and is not. It is a one-time, disclosed tailwind in a single year, around 4 to 5% of profit. It is not a claim that earnings are fabricated, and it is an order of magnitude smaller than the cumulative figures circulating in the bubble debate.
  16. 16
    www.sec.gov
    Change in Accounting Estimate In July 2022, we completed an assessment of the useful lives of our server and network equipment. Due to investments in software that increased efficiencies in how we operate our server and network equipment, as well as advances in technology, we determined we should increase the estimated useful lives of both server and network equipment from four years to six years. This change in accounting estimate was effective beginning fiscal year 2023. Based on the carrying amount of server and network equipment included in property and equipment, net as of June 30, 2022, the effect of this change in estimate for fiscal year 2023 was an increase in operating income of $3.7 billion and net income of $3.0 billion, or $0.40 per both basic and diluted share.
  17. 17
    www.sec.gov
    (In millions, except percentages and per share amounts) 2023 2022 Percentage Change Revenue $ 211,915 $ 198,270 7% Gross margin 146,052 135,620 8% Operating income 88,523 83,383 6% Net income 72,361 72,738 (1)% Diluted earnings per share 9.68 9.65 0%
  18. 18
    www.sec.gov
    In January 2023, we completed an assessment of the useful lives of our servers and network equipment and adjusted the estimated useful life of our servers from four years to six years and the estimated useful life of certain network equipment from five years to six years. This change in accounting estimate was effective beginning in fiscal year 2023, and the effect was a reduction in depreciation expense of $983 million and $3.9 billion and an increase in net income of $765 million and $3.0 billion, or $0.06 and $0.24 per basic and $0.06 and $0.24 per diluted share for the three and twelve months ended December 31, 2023, respectively.
  19. 19
    www.sec.gov
    Operating income $ 18,160 $ 23,697 $ 74,842 $ 84,293
  20. 20
    www.sec.gov
    During the first quarter of fiscal 2023, we completed an assessment of the useful lives of our servers and increased the estimate of the useful lives from four years to five years effective at the beginning of fiscal 2023. Based on the carrying value of our servers as of May 31, 2022, this change in accounting estimate decreased our total operating expenses by $ 434 million during fiscal 2023.
  21. 21
    www.sec.gov
    During the first quarter of fiscal 2025, we completed an assessment of the useful lives of our servers and networking equipment and increased the estimate of the useful lives from five years to six years , effective at the beginning of fiscal 2025. Based on the carrying value of our servers and networking equipment as of May 31, 2024, this change in accounting estimate decreased our total operating expenses by $ 733 million and increased our net income by $ 573 million, or $ 0.21 per basic and $ 0.20 per diluted share, during fiscal 2025.
  22. 22
    www.sec.gov
    Total operating expenses 39,721 37,608 36,861 Operating income 17,678 15,353 13,093 Interest expense ( 3,578 ) ( 3,514 ) ( 3,505 ) Non-operating income (expenses), net 60 ( 98 ) ( 462 ) Income before income taxes 14,160 11,741 9,126 Provision for income taxes 1,717 1,274 623 Net income $ 12,443 $ 10,467 $ 8,503 Earnings per share: Basic $ 4.46 $ 3.82 $ 3.15 Diluted $ 4.34 $ 3.71 $ 3.07
  23. 23
    aeacapitalresearch.com
    AMZN extended server lives from five to six years in 2024, reducing depreciation by $3.2 billion and increasing net income by $2.5 billion.
  24. 24
    www.buxtonhelmsley.com
    In its fourth quarter 2023 earnings release, dated February 1, 2024, the company disclosed that effective January 2024, it was extending the useful life of its servers from five years to six years, an action that Chief Financial Officer Brian Olsavsky stated on the corresponding earnings call was expected to benefit first quarter 2024 operating income by approximately $900 million.
  25. 25
    www.sec.gov
    In January 2025, we completed an assessment of the useful lives of property and equipment, which resulted in an increase in the estimated useful lives of most servers and network assets to 5.5 years, effective January 1, 2025. Based on the servers and network assets placed in service as of December 31, 2024, the financial impact of this change in estimate included a reduction in depreciation expense of $ 2.92 billion and an increase in net income of $ 2.59 billion, or $ 1.00 per diluted share, for the year ended December 31, 2025.
  26. 26
    www.sec.gov
    Net income was $60.46 billion, with diluted earnings per share (EPS) of $23.49 for the year ended December 31, 2025.
  27. 27
    www.sec.gov
    We review the useful lives of equipment on an ongoing basis. Effective January 1, 2025 we changed our estimate of the useful lives of a subset of our servers and networking equipment from six years to five years . The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning. The effect of this change in estimate for the year ended December 31, 2025, based on servers and networking equipment that were included in “Property and equipment, net” as of December 31, 2024 and those acquired during the year ended December 31, 2025, was an increase in depreciation and amortization expense of $ 1.4 billion and a reduction in net income of $ 1.0 billion, or $ 0.10 per basic share and $ 0.10 per diluted share, which primarily impacted our AWS segment.
  28. 28
    www.sec.gov
    Information on reportable segments and reconciliation to consolidated net income is as follows (in millions): Year Ended December 31, 2023 2024 2025 North America Net sales $ 352,828 $ 387,497 $ 426,305 Operating expenses 337,951 362,530 396,686 Operating income $ 14,877 $ 24,967 $ 29,619 International Net sales $ 131,200 $ 142,906 $ 161,894 Operating expenses 133,856 139,114 157,144 Operating income (loss) $ ( 2,656 ) $ 3,792 $ 4,750 AWS Net sales $ 90,757 $ 107,556 $ 128,725 Operating expenses 66,126 67,722 83,119 Operating income $ 24,631 $ 39,834 $ 45,606 Consolidated Net sales $ 574,785 $ 637,959 $ 716,924 Operating expenses 537,933 569,366 636,949 Operating income 36,852 68,593 79,975 Total non-operating income 705 21 17,336 Provision for income taxes ( 7,120 ) ( 9,265 ) ( 19,087 ) Equity-method investment activity, net of tax ( 12 ) ( 101 ) ( 554 ) Net income $ 30,425 $ 59,248 $ 77,670
  29. 29
    www.sec.gov
    Effective January 1, 2023, the Company changed its estimate of the useful life for its computing equipment utilized in data centers from five to six years , reflecting continuous advancements in hardware performance, software optimization, and data center design improvements. The effects of this change in estimate for the year ended December 31, 2023 on computing equipment that was included in property and equipment, net on the consolidated balance sheets as of December 31, 2023 was a reduction in total expenses of $ 20 million. The per share impact of the change in estimate was a $ 0.10 increase for the year ended December 31, 2023.
  30. 30
    www.sec.gov
    In January 2026, the Company completed an assessment of the useful lives of servers and network equipment based on updated information and usage patterns obtained and concluded that the estimated useful lives of such assets should be extended from four to five years. Management applied this change in accounting estimate prospectively beginning January 1, 2026. Based on the servers and network equipment placed in service as of December 31, 2025, the financial impact of this change in estimate included a reduction in depreciation expense of $43.1 and an increase in net income of $41.6, for the three months ended March 31, 2026.
  31. 31
    www.sec.gov
    Three months ended March 31, Notes 2025* 2026 Revenues 50.9 399.0 Operating costs and expenses: Cost of revenues (1) 24.7 103.8 Product development (1) 36.5 67.4 Sales, general and administrative (1) 60.9 143.8 Depreciation and amortization 49.1 212.0 Total operating costs and expenses 171.2 527.0 Loss from operations (120.3) (128.0)
  32. 32
    www.sec.gov
    In 2024, we also determined, primarily in the fourth quarter, to retire early certain of our servers and networking equipment. We recorded approximately $ 920 million of accelerated depreciation and related charges for the quarter ended December 31, 2024 related to these decisions. The accelerated depreciation will continue into 2025 and decrease operating income by approximately $ 0.6 billion in 2025. These two changes above are due to an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.
  33. 33
    suttonfinancialreporting.com
    The five companies reported aggregate operating income of $439bn in FY2025. If they had continued to depreciate servers and network assets over the pre-2022 estimated life of four years, their aggregate FY2025 operating income would have been $395bn. Thus, slower depreciation lifted their operating income by $44bn in FY2025, a material 11% increase. The uplift to income will increase in future years as data centre capex grows (it is forecast to reach $1.7tn by 2030).3 Note there is no tax cost from extending these assets’ life since US tax authorities determine tax depreciation separately.
  34. 34
    suttonfinancialreporting.com
    The impact of longer expected life of data servers and network assets on the five companies’ FY2025 operating income is estimated in the exhibit below. Key assumptions in the calculation are that a full year’s depreciation is charged in the year of asset purchase and data servers and network assets have an estimated residual value of zero. (All five companies depreciate these assets on a straight-line basis.)
  35. 35
    www.sec.gov
    During the nine months ended September 30, 2025 and 2026, we spent $63.6 billion and $132.0 billion on capital expenditures, respectively. In 2026, we expect to significantly increase, relative to 2025, our investment in our technical infrastructure, including servers and network equipment and data centers.
  36. 36
    www.sec.gov
    Depreciation of our property and equipment commences when such assets are ready for their intended use. For the nine months ended September 30, 2025 and 2026, our depreciation on property and equipment was $15.1 billion and $22.0 billion, respectively.
  37. 37
    www.sec.gov
    Cash capital expenditures were $34.2 billion and $75.9 billion during Q3 2025 and Q3 2026, and $89.9 billion and $172.2 billion for the nine months ended September 30, 2025 and 2026, which primarily reflect investments in technology infrastructure (the majority of which is to support AWS business growth) and in additional capacity to support our fulfillment network, both of which we expect to increase in 2026. We expect cash capital expenditures to meaningfully increase in 2027, primarily driven by investments in technology infrastructure.
  38. 38
    www.sec.gov
    Depreciation and amortization expense on property and equipment, including corporate property and equipment, are allocated to all segments based on usage. Total depreciation and amortization expense, by segment, is as follows (in millions): Three Months Ended September 30, Nine Months Ended September 30, 2025 2026 2025 2026 North America $ 4,001 $ 4,566 $ 11,273 $ 13,346 International 1,276 1,346 3,592 3,916 AWS 5,648 9,230 14,882 24,583 Consolidated $ 10,925 $ 15,142 $ 29,747 $ 41,845
  39. 39
    www.sec.gov
    Investing Additions to property and equipment ( 115,948 ) ( 64,551 ) ( 44,477 )
  40. 40
    www.sec.gov
    NOTE 6 - PROPERTY AND EQUIPMENT The components of property and equipment were as follows: (In millions) June 30, 2026 2025 Land $ 10,546 $ 9,338 Buildings and improvements 182,749 137,921 Leasehold improvements 16,348 12,117 Servers, network equipment, and software 215,874 132,836 Furniture and equipment 6,250 6,407 Total, at cost 431,767 298,619 Accumulated depreciation ( 118,691 ) ( 93,653 ) Total, net $ 313,076 $ 204,966 During fiscal years 2026, 2025, and 2024, depreciation expense was $ 34.3 billion, $ 22.0 billion, and $ 15.2 billion, respectively. As of June 30, 2026, 2025, and 2024, purchases of property and equipment remaining in accounts payable were $ 26.7 billion, $ 6.9 billion, and $ 4.3 billion, respectively. As of June 30, 2026, we have committed $ 34.6 billion for the construction of new buildings, building improvements, and leasehold improvements, primarily related to datacenters.
  41. 41
    www.sec.gov
    Capital Expenditures : Cash used for capital expenditures increased from $21.2 billion in fiscal 2025 to $55.7 billion in fiscal 2026 primarily due to the expansion of our data centers. We expect this upward trend to continue during fiscal 2027 and in the following fiscal years as we increase our existing data center capacity and establish data centers in new geographic locations in order to meet current and expected customer demand for our cloud offerings. We believe that our current cash, cash equivalents and marketable securities balances, together with cash generated from operations and available financing arrangements, will be sufficient to meet our working capital, committed capital expenditures and contractual obligations for at least the next twelve months. Thereafter, we expect that our existing sources of liquidity, together with potential access to additional financing, will continue to be sufficient for the foreseeable future. Further, we have flexibility in managing the timing of certain discretionary capital expenditures.
  42. 42
    www.sec.gov
    Estimated May 31, (Dollars in millions) Useful Life 2026 2025 Computer, network, machinery and equipment 1 - 6 years (1) $ 59,634 $ 30,345 Buildings and improvements 1 - 40 years 21,263 10,881 Furniture, fixtures and other 5 - 15 years 452 466 Land - 1,329 1,352 Construction in progress (2) - 39,973 16,510 Total property, plant and equipment 1 - 40 years 122,651 59,554 Accumulated depreciation ( 22,694 ) ( 16,032 ) Total property, plant and equipment, net $ 99,957 $ 43,522 (1) Comprised primarily of servers and networking equipment with estimated useful life of six years . (2) Comprised primarily of servers, networking equipment and leasehold improvements to be deployed at our data centers. Depreciation expense on property, plant and equipment in fiscal 2026, 2025 and 2024 were $ 7.6 billion, $ 3.9 billion and $ 3.1 billion, respectively. Property, plant and equipment, net includes ROU assets recorded in connection with lease arrangements that are accounted for as finance leases, totaling $ 7.5 billion and $ 2.9 billion as of May 31, 2026 and 2025, respectively.
  43. 43
    www.sec.gov
    During the six months ended June 30, 2026 and 2025, cash paid for property and equipment was $14.1 billion and $3.9 billion, respectively. We expect to increase, relative to 2025, our investment in our technology and infrastructure, including servers, network equipment, and data center related expenses, to support the growth of our business and our long-term initiatives.
  44. 44
    www.sec.gov
    Depreciation and amortization on property and equipment was $ 1.4 billion and $ 553 million for the three months ended June 30, 2026 and 2025, respectively, and $ 2.5 billion and $ 996 million for the six months ended June 30, 2026 and 2025, respectively.
  45. 45
    d18rn0p25nwr6d.cloudfront.net
    Property and equipment, net consists of the following (in millions): September 30, 2026 December 31, 2025 Land $ 3,764 $ 3,687 Servers and network assets 133,626 98,040 Buildings 68,223 55,568 Leasehold improvements 8,855 8,346 Equipment and other 10,818 9,377 Finance lease right-of-use assets 11,704 8,187 Construction in progress 89,499 50,521 Property and equipment, gross 326,489 233,726 Less: Accumulated depreciation (73,267) (57,326) Property and equipment, net $ 253,222 $ 176,400
  46. 46
    d18rn0p25nwr6d.cloudfront.net
    We anticipate making capital expenditures of approximately $135 billion to $145 billion in 2026 to support our AI efforts and core business.
  47. 47
    aeacapitalresearch.com
    ORCL nearly doubled gross computer and network equipment between fiscal 2025 and fiscal 2026 while construction in progress more than doubled. GOOGL more than doubled first-half capital expenditures year over year, while depreciation grew 43%. That gap is partly timing: assets are not depreciated until they enter service, and then their cost is spread across years.
  48. 48
    www.accountingtoday.com
    Bank of America estimates the spending will be a 1.6 percentage point drag on the companies' margins for earnings before interest and taxes in 2026 compared with the fourth quarter of 2024, strategists Ohsung Kwon and Savita Subramanian wrote in a research note on Feb. 10.
  49. 49
    www.amcompute.com
    Useful life Annual depreciation Book value after 3 years After 5 years 3 years $ 333 M $ 0 M $ 0 M 4 years $ 250 M $ 250 M $ 0 M 5 years $ 200 M $ 400 M $ 0 M 6 years $ 167 M $ 500 M $ 167 M Straight line, no salvage value, $1 billion of equipment placed in service on day one.
  50. 50
    d18rn0p25nwr6d.cloudfront.net
    Depreciation expense on property and equipment was $6.81 billion and $4.76 billion for the three months ended September 30, 2026 and 2025, respectively, and $18.49 billion and $12.88 billion for the nine months ended September 30, 2026 and 2025, respectively. Within property and equipment, our servers and network assets depreciation expense was $5.34 billion and $3.60 billion for the three months ended September 30, 2026 and 2025, respectively, and $14.34 billion and $9.34 billion for the nine months ended September 30, 2026 and 2025, respectively.
  51. 51
    www.sec.gov
    Income from operations for 2025 was $83.28 billion, an increase of $13.90 billion, or 20%, compared to 2024, driven by an increase in advertising revenue, partially offset by an increase in costs and expenses. The increase in costs and expenses was mainly due to increases in employee compensation and infrastructure costs.
  52. 52
    www.sec.gov
    Depreciation of our property and equipment commences when such assets are ready for their intended use. For the years ended December 31, 2024 and 2025, our depreciation on property and equipment was $15.3 billion and $21.1 billion, respectively.
  53. 53
    www.sec.gov
    Property and Equipment, Net Property and equipment, net, consisted of the following (in millions): As of December 31, 2024 2025 Technical infrastructure (1) $ 141,852 $ 203,679 Office space 45,403 48,348 Corporate and other assets 12,574 14,463 Property and equipment, in service 199,829 266,490 Less: accumulated depreciation ( 79,390 ) ( 98,485 ) Add: assets not yet in service 50,597 78,592 Property and equipment, net $ 171,036 $ 246,597 (1) As of December 31, 2024 and 2025, approximately 60 % of technical infrastructure assets were comprised of servers and network equipment. The remaining balance was comprised of data center land and buildings and related assets.
  54. 54
    www.sec.gov
    The following table summarizes consolidated financial results (in millions, except for per share information and percentages): Year Ended December 31, 2024 2025 $ Change % Change Consolidated revenues $ 350,018 $ 402,836 $ 52,818 15 % Cost of revenues $ 146,306 $ 162,535 $ 16,229 11 % Operating expenses $ 91,322 $ 111,262 $ 19,940 22 % Operating income $ 112,390 $ 129,039 $ 16,649 15 % Operating margin 32 % 32 % 0 % Other income (expense), net $ 7,425 $ 29,787 $ 22,362 301 % Net income $ 100,118 $ 132,170 $ 32,052 32 % Diluted net income per share (1) $ 8.04 $ 10.81 $ 2.77 34 %
  55. 55
    www.sec.gov
    SUMMARY RESULTS OF OPERATIONS (In millions, except percentages and per share amounts) 2026 2025 Percentage Change Revenue $ 331,839 $ 281,724 18% Gross margin 225,465 193,893 16% Operating income 155,237 128,528 21% Net income 133,749 101,832 31% Diluted earnings per share 17.95 13.64 32%
  56. 56
    www.sec.gov
    Depreciation expenses on property, plant and equipment during the three months ended August 31, 2026 and 2025 were $ 3.2 billion and $ 1.4 billion, respectively.
  57. 57
    www.sec.gov
    Operating income 6,728 4,277 Interest expense ( 1,428 ) ( 923 ) Non-operating income, net 307 73 Income before income taxes 5,607 3,427 Provision for income taxes 847 500 Net income $ 4,760 $ 2,927 Preferred stock dividends 81 - Net income available to common shareholders $ 4,679 $ 2,927 Earnings per share attributable to common shareholders: Basic $ 1.58 $ 1.04 Diluted $ 1.56 $ 1.01
  58. 58
    www.sec.gov
    Property and equipment, net, consisted of the following (in millions): June 30, 2026 December 31, 2025 Technology equipment $ 33,823 $ 20,903 Software 859 802 Data center equipment and leasehold improvements 5,997 2,842 Furniture, fixtures, and other assets 25 18 Construction in progress 11,918 9,376 Total property and equipment 52,622 33,941 Less: accumulated depreciation and amortization ( 5,886 ) ( 3,384 ) Total property and equipment, net $ 46,736 $ 30,557
  59. 59
    d18rn0p25nwr6d.cloudfront.net
    Operating income (loss) (49) 19 (193) (8)
  60. 60
    s205.q4cdn.com
    Revenue Backlog1 $ 104.2B Revenue $ 2.6B +112% YoY Does not include more than $25 billion of net new customer commitments added in early Q3 +246% YoY Adj. Operating Income 1 $ 128M 5% Adj. Operating Margin
  61. 61
    gpuinsights.net
    Microsoft’s property note, in millions, shows servers, network equipment, and software at a gross carrying amount of $215,874 million on June 30, 2026, against $132,836 million a year earlier. The gross balance rose by $83,038 million. That line includes software, whose disclosed life is three years, and servers, whose disclosed life is two to six years. Dividing $215,874 million by any single life produces a fake depreciation number. We do not do it.
  62. 62
    www.indmoney.com
    J.P. Morgan’s stress test also found that applying a 3-year depreciation life to AI hardware would cut EPS and operating margins by roughly 6-8% for most hyperscalers, with Oracle as the bigger exception. That is material, but not collapse-level.
  63. 63
    www.indmoney.com
    Here is what the reported earnings per share (EPS) of each company would look like if they used a 3-year depreciation schedule, consistent with Nvidia's actual product cycle, instead of their current 5 to 6-year assumptions. Company Current Life FY25 EPS Est. EPS at 3-Year Life Earnings Impact Meta (META) 5.5 years $23.49 ~$19.92 ~15% lower Oracle (ORCL) 5 years $4.34 ~$3.87 ~11% lower Microsoft (MSFT) 6 years ~$11.80 ~$10.89 ~8% lower Alphabet (GOOGL) 6 years $10.81 ~$9.98 ~8% lower Source: SEC 10-K Filings, SEC 10-Q Filings), Internal Calculation | Note: Servers and network equipment represent roughly 60 to 70% of each company's total annual depreciation, based on their asset mix disclosures. This is an estimate, since companies don't separately break out server depreciation in their filings.
  64. 64
    epoch.ai
    Since 2023, the average dollar spent on AI chips each quarter has yielded about 49% more performance each year. At that rate, the performance a dollar buys doubles every 1.7 years. Growth has come in spurts, as spending shifts to the newest generation of chips. Price-performance was nearly flat through mid-2024, then roughly doubled as Blackwell-generation chips grew to a majority of new spending.
  65. 65
    epoch.ai
    Performance per dollar rose from about 5.6 × 10 11 bit-operations per second per dollar in the first quarter of 2023 to about 1.4 × 10 12 by the end of 2025, in constant 2025 dollars. Fitting an exponential trend gives an average growth rate of about 49% per year (90% CI: 36% to 66% per year). The gains are largely due to chips getting exponentially more powerful with each new generation. While recent chips have become more expensive, they have grown more powerful at an even faster rate. In 2025 dollars, NVIDIA’s GB300 costs about 5.5 times the P100’s 2016 launch price, yet it delivers roughly 200 times the performance, making it about 37 times more cost effective.
  66. 66
    epoch.ai
    We use peak theoretical performance as given by each chip’s spec sheet, converted to peak bit-operations per second. Real-world performance depends heavily on software support for the hardware, the model architecture, the workload shape, the rack-scale system design, the networking topology, and memory and interconnect bandwidth. Realized performance per dollar on any given workload can differ substantially from on-paper specs.
  67. 67
    web.archive.org
    Uncertainty around AI silicon’s useful life reflects a core tension: Rapid improvements in performance per dollar between generations of AI silicon push companies to replace hardware quickly, while the growing range of AI tasks means older chips can still deliver value for longer. This tension is sharpened by NVIDIA's unprecedented annual release cadence for GPU architectures, with each generation delivering step-function leaps in capability rather than incremental improvements.
  68. 68
    media-stg.researchaffiliates.com
    Economically, AI hardware ages much faster. Nvidia now operates on an annual product cadence. Each generation of new chips delivers materially more compute per watt than the last. Older chips rarely fail physically. They become obsolete economically. The relevant test is not whether a GPU still runs, but whether it earns a return sufficient to cover depreciation, operating costs, and the cost of capital.
  69. 69
    cloud.google.com
    Generation over generation: The performance leap Our commitment to co-designing hardware and software continues to pay dividends. When compared to seventh-generation Ironwood TPU, the eighthgeneration TPUs deliver massive gains: Training price-performance : TPU 8t delivers up to 2.7x performance-per-dollar improvement over Ironwood TPU for large-scale training. Inference price-performance : TPU 8i delivers up to 80% performance-per-dollar improvement over Ironwood TPU, particularly at low-latency targets for large MoE models. Energy efficiency : Both chips deliver up to 2x better performance-per-watt , critical for scaling the next generation of AI sustainably.
  70. 70
    aws.amazon.com
    Trn3 delivers up to 4.4x higher performance, 3.9x higher memory bandwidth and 4x better performance/watt compared to our Trn2 UltraServers, providing the best price-performance for training and serving frontier-scale models, including reinforcement learning, Mixture-of-Experts (MoE), reasoning, and long-context architectures.
  71. 71
    developer.nvidia.com
    Figure 1. Pareto frontiers illustrating the 10x generational uplift in agentic inference performance of the Vera Rubin platform (internal 2T MoE workload)
  72. 72
    blogs.nvidia.com
    SemiAnalysis AgentX data shows NVIDIA Vera Rubin NVL72 systems deliver over 30x higher throughput per megawatt than NVIDIA GB300 NVL72, and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model.
  73. 73
    ibinterviewquestions.com
    The opposing argument is that AI hardware is changing faster than any previous generation of servers. Leading chip designers now release new generations roughly every year, each delivering large gains in performance per watt, and in data centers constrained by power, 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.
  74. 74
    www.cnbc.com
    Although Microsoft plans to build AI infrastructure aggressively, CEO Satya Nadella said this week that his company is trying to space out its AI chip purchases and not overinvest in a single generation of processors. He added that the biggest competitor for any new Nvidia AI chip is its predecessor. "One of the biggest learnings we had even with Nvidia is that their pace increased in terms of their migrations," Nadella said. "That was a big factor. I didn't want to go get stuck with four or five years of depreciation on one generation."
  75. 75
    www.stateof.ai
    Software support can also constrain reuse: CUDA 13 removed offline compilation and library support for Maxwell, Pascal, and Volta architectures.
  76. 76
    www.stateof.ai
    NVIDIA says applications built with older toolkits can continue running on supported drivers; the change does not make those chips stop working.
  77. 77
    blog.citp.princeton.edu
    Technical analyses have converged on estimating the useful lifespan of AI chips at one to three years. One unnamed Google architect assessed that GPUs running at 60-70% utilization-standard for AI workloads-survive one to two years, with three years as a maximum. The reason: thermal and electrical stress is simply too high. But physical failure isn’t the only concern. Technological obsolescence drives replacement cycles. Nvidia’s GB200 (“Blackwell”) chip provides 4-5x faster inference than the H100. When competitors deploy hardware with significantly better performance, three-year-old chips become economically obsolete even if they still function.
  78. 78
    s205.q4cdn.com
    What we are seeing today is that the upside of recontracting is real as we remain largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs. So as our earlier generation fleets roll off their original contracts, they offer the potential to deliver strong returns in the subsequent years. We are seeing this across our [ph] Ampere (26:43) and Hopper fleet. As an example, we recently signed an A100 contract that extends into 2029 at an attractive price. As a reminder, this SKU was introduced in 2020. Clusters of prior generations of architecture offer installed energized production grade compute already running at scale. They come with a proven ROI for customers. In a market where new capacity is supply constrained and costs are rising, AI cloud infrastructure in production is a scarce, valuable asset. While we have built a business whose economics do not rely on recontracting after initial customer term, increasingly we are seeing longer utilization at higher prices, offering the potential for significant further upside.
  79. 79
    s205.q4cdn.com
    Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs while pricing for prior generation SKUs is at or above where it was years ago. Our near-term capacity remains effectively sold out.
  80. 80
    www.cnbc.com
    Some of Nvidia's customers say AI chips will retain value for a long time and that customers will continue to pay for access to older processors because they'll still be useful for other tasks. CoreWeave , which buys GPUs and rents them out to clients, has used six-year depreciation cycles for its infrastructure since 2023. CoreWeave CEO Michael Intrator told CNBC this week, following quarterly earnings, that his company is being "data driven" about GPU shelf life. Intrator said that CoreWeave's Nvidia A100 chips, which were announced in 2020, are all fully booked. He also added that a batch of Nvidia H100 chips from 2022 became available because a contract expired, and they were immediately booked at 95% of their original price. "All of the data points that I'm getting are telling me that the infrastructure retains value," Intrator said.
  81. 81
    s205.q4cdn.com
    Average contract Recent Large Wins GAAP GPU useful life V100 GPUs T4 GPUs A100 GPUs H100 GPUs Contract duration Observed usage in market vs. GAAP useful life (2) ~5 years 5+ years ~6 years ~8 years since release ~7 years since release ~5 years since release ~3 years since release ~1 year of GAAP useful life post-initial contract Observed usage in market since release > GAAP useful life ~6yr GAAP useful life Illustrative extrapolation based on V100 GPUs Illustrative extrapolation based on V100 GPUs (1) Recent wins have 5-6 yr durations (1) (1) Illustrative useful life is based on V100 GPUs and is being shown for illustrative purposes only. Actual useful life may vary significantly (2) Observed usage commencement date is based on announcement of NVIDIA GPUs; May 2017 for Tesla V100 GPUs; September 2018 for Tesla T4 GPU; May 2020 for Ampere A100 GPUs and March 2022 for Hopper H100 GPUs
  82. 82
    www.stateof.ai
    AWS still offers NVIDIA T4-based G4dn instances for machine learning inference and small-scale training, including applications such as image classification and speech recognition.
  83. 83
    www.businessinsider.com
    Rental-market data backs up the company's comments. Silicon Data, which tracks GPU prices, says A100 rental rates have held up well after a strong rebound in 2026. "We are still learning when it comes to the question of economic lifespan of GPUs. It certainly doesn't appear to be 2-3 years as some seem to casually assume," Silicon Data wrote in a post on X on Tuesday.
  84. 84
    computecensus.com
    As of 16 Oct 2026, renting an NVIDIA H100 80GB costs $3.49 per GPU-hour at the median on-demand rate, with published offers from $1.30 to $12.29. As of 16 Oct 2026, the median on-demand price is $4.40 per GPU-hour for the H200 141GB, $6.25 for the B200 and $1.75 for the A100 80GB. A published 12-month reserved H100 rate of $3.26 per GPU-hour is -7% against the on-demand median (16 Oct 2026). The H100 median moved +79% between 1 Oct 2025 ($1.95) and 16 Oct 2026. The Compute Census GPU index covers 10 models and 29 providers with a recorded price (snapshot 16 Oct 2026).
  85. 85
    data.ornn.com
    Current GPU Rental Prices What renting a GPU costs, in USD per GPU-hour, from executed compute trades. Settled October 30, 2026 . GPU Price Day % 1M % Chart ( 1M ) H100 SXMBloomberg: ORNNH100 $3.03 +3.77% +12.64% H200 $5.49 -3.85% -4.02% B200Bloomberg: ORNNB200 $9.77 -0.71% +21.37% A100 SXM4 $1.08 -1.82% +18.68% RTX PRO 6000 $1.54 +0.00% +4.05%
  86. 86
    gpueconomy.com
    Our own series starts in the week of September 7, 2026. For the time before it we publish a second-hand series rebuilt from a public, Apache-licensed price catalog (its source is named in the methodology) that bots rewrite every few hours: for each ISO week we take the cheapest per-GPU on-demand H100 SXM instance per provider as the catalog listed it at the end of that week, and publish the median across providers when at least 4 contributed. We vouch that the catalog published each number on that date - every row links to the exact commit - not that the provider’s own page did. It is a different panel from the series above (the catalog covers at most 7 of our providers) and is never merged with it. From the week of January 6, 2025 to the week of October 19, 2026, the reconstructed median moved from $7.24 to $4.25 per H100 GPU-hour, with the panel growing from 4 to 6 providers (between 4 and 7 over the period). Read each point against its provider count: a cheaper provider joining the catalog lowers the median without any listed price moving.
  87. 87
    intuitionlabs.ai
    Research firm SemiAnalysis reported that one-year H100 rental contract pricing actually reversed course, having "shot up almost 40% from a low of $1.70/hr/GPU in October 2025 to $2.35/hr/GPU" ( [32] ) by March 2026, evidence that even a H100 chip's rental price responds to capacity shortages, not just to its age.
  88. 88
    media-stg.researchaffiliates.com
    Observed hourly rates for Nvidia H100 GPUs fell from scarcity-driven peaks near $8 in 2024 to below $3 by late 2025, and to below $1 in early 2026. As new generations arrive, pricing converges toward the marginal cost of the most efficient hardware.
  89. 89
    www.silicondata.com
    In 2024 (roughly ~2 years post-launch), USED H100s traded at a median of ~61% of the contemporaneous NEW price. By 2025 (~3 years post-launch), the median moved to ~69% - a reminder that secondary-market pricing is often non-linear and can shift with supply waves, listing mix, and buyer expectations, not just hardware age. Rather than behaving like a smooth “depreciation curve,” the USED segment is best understood as market repricing: values reset quickly as supply loosens, alternative GPUs enter the conversation (e.g., H200 and Blackwell-class systems), and buyers become more selective about risk, warranty, and remaining service life.
  90. 90
    www.silicondata.com
    2-year-old H100s now trade at just 30-40% of their historical maximum NEW price. 3-year-old units fall further, with resale values clustering between 20-30% of their peak. These numbers show that even GPUs with seemingly high resale prices (e.g., $12K-$16K) have already lost the majority of their peak value if they once commanded $40K+.
  91. 91
    aethir.com
    The institutional support for the secondary enterprise GPU market arrived in July 2026, when a marketplace for used, refurbished, and surplus accelerators opened alongside an existing reserved-capacity venue, quoting used H100 units at $6,000 to $22,000 against $25,000 to $40,000 new, and used A100 80GB cards at $4,000 to $9,000 against $7,000 to $15,000 new. Equipment in that channel was described as selling 40% to 60% below original list.
  92. 92
    blogs.nvidia.com
    Silicon Data shows a six-year-old A100 GPU is still worth a quarter of what it cost, where a five-year depreciation schedule had it at zero more than a year ago.
  93. 93
    www.amcompute.com
    Book life, resale value and physical life A useful life is an accounting choice. It sets how fast a server's cost reaches earnings, and says nothing about what the GPU will sell for. Resale value follows the market. Our 2026 residual report puts a used H100 at 11-19% of its original list price in 2027 in the conservative band.
  94. 94
    media-stg.researchaffiliates.com
    During the first three years, returns on capital are substantial. In year four, revenue falls below total economic cost. The chip remains in service because it still covers marginal operating expenses. But it no longer earns a return on invested capital. Economically, the productive life of the asset is closer to three years than to the five-year accounting for depreciation.
  95. 95
    media-stg.researchaffiliates.com
    Annual revenue equals hourly rental rate × 8,760 hours × 60% utilization. Total cost equals straight-line depreciation (five-year life), operating expense, and an 8% capital charge applied to beginning-of-year book value. Economic profit equals revenue minus total cost; ROCE equals economic profit divided by beginning-of-year book value.
  96. 96
    web.archive.org
    Today, the rental price of trailing-edge NVIDIA devices such as A100s and H100s remains high enough to suggest useful lives of five to six-plus years. That could be a consequence of the extreme capacity constraints model providers are operating under today, or it could be a signal about the sustained value of silicon in the AI era-and thus the appropriateness of its prevailing depreciation timelines.
  97. 97
    www.sec.gov
    As of December 31, 2025, our committed contracts had a weighted-average contract duration of approximately five years.
  98. 98
    www.sec.gov
    We continually work to upgrade and enhance our platform, solutions, and services in response to customer demand and to keep up with technological changes. Part of this process entails cycling out older components of our infrastructure and replacing them with the latest technology available. This requires us to make certain estimates with respect to the useful life of the components of our infrastructure and to maximize the value of the components of our infrastructure, including our GPUs, to the fullest extent possible. We cannot guarantee that our estimates will be accurate or that our attempts at maximizing value will be successful. Any changes to the significant assumptions underlying our estimates or to the estimates of our components' useful lives, or any inability to redeploy components of our existing infrastructure to extend past their contracted life could significantly affect our business, operating results, financial condition, and prospects.
  99. 99
    s205.q4cdn.com
    The deployment delivers attractive returns, fully repaying asset level debt used to fund the CapEx while generating significant additional free cash flow. So when an initial contract ends, the cluster no longer has any leverage and we are free to recontract that cloud infrastructure or offer it to the market. We will have generated an attractive return even before the prospect of further monetizing the cloud infrastructure. Every resale or renewal is incremental on top of the returns already earned within the initial term.
  100. 100
    www.sec.gov
    Depreciation and amortization on property and equipment was $ 2.4 billion, $ 861 million, and $ 101 million for the years ended December 31, 2025, 2024, and 2023, respectively.
  101. 101
    www.sec.gov
    Year Ended December 31, 2025 2024 2023 (in millions) Revenue $ 5,131 $ 1,915 $ 229 Operating expenses: Cost of revenue (1) 1,453 493 69 Technology and infrastructure (1) 2,929 961 131 Sales and marketing (1) 144 18 13 General and administrative (1) 651 119 30 Total operating expenses 5,177 1,591 243 Operating income (loss) (46) 324 (14) Gain (loss) on fair value adjustments 27 (756) (534) Interest expense, net (1,229) (361) (28) Other income, net 33 49 18 Loss before income taxes (1,215) (744) (558) Provision for (benefit from) income taxes (48) 119 36 Net loss and comprehensive loss $ (1,167) $ (863) $ (594)
  102. 102
    www.sec.gov
    Depreciation and amortization expenses increased by $162.9 million, or 332%, from $49.1 million in the first quarter of 2025 to $212.0 million in the first quarter of 2026. The primary driver of the increase in depreciation and amortization expenses was the continued investments in GPU-related capital expenditures and related data center hardware for the Nebius AI cloud business. Starting the first quarter of 2026, we revised the useful life for our server and network equipment from four years to five years to reflect usage patterns and current utilization commitments. The change in accounting estimate has been applied prospectively from 2026. As a percentage of total operating expenses, depreciation and amortization increased from 29% in the first quarter of 2025 to 40% in the first quarter of 2026, highlighting the growing capital intensity of our infrastructure operations.
  103. 103
    www.sec.gov
    In addition, the equipment we deploy is subject to rapid technological change. Successive generations of GPUs and supporting infrastructure are being released on short development cycles, and customer demand may concentrate on the newest generations of hardware. As a result, the economic useful life and residual value of our deployed equipment may be shorter or lower than we anticipate, pricing for services delivered on older generations of hardware may decline, and we may be required to make substantial ongoing capital expenditures to refresh our fleet and to modify our facilities to meet the power density, cooling and other requirements of new hardware. If we do not manage technological transitions effectively, our competitive position, the utilization and value of our assets, our depreciation profile and our results of operations and financial condition could be materially and adversely affected.
  104. 104
    www.sec.gov
    If the term of our customer contracts is shorter than the useful life of the associated assets or the tenor of the associated power commitments and financing, we are exposed to recontracting and pricing risk at the end of contract terms.
  105. 105
    www.sec.gov
    We may also encounter pricing pressure in respect of capacity for older generations of GPUs as newer generations are introduced. We may be unable to effectively calibrate our prices, whether through increases or decreases, in order to remain competitive and attract new customers and develop our existing customer base. Given our relatively early stage of development and the immaturity of the market, there is limited experience with respect to determining the most favorable prices and pricing models for our offerings. As customer demand shifts to inference and other use cases, we may experience changing pricing dynamics. In addition, larger competitors with more diverse offerings may reduce the price of any offerings that compete with ours or may bundle them with other solutions and services. This could lead customers to demand greater price concessions or additional functionality at the same price levels. These risks may reduce our margins and adversely affect our business, operating results, financial condition, and future prospects.
  106. 106
    www.sec.gov
    Our controls related to fixed assets were not adequately designed and were not operating effectively. Specifically, we did not fully implement and ensure the effectiveness of the relevant controls and procedures over depreciation start dates, and timely reconciliation around the asset count process. As a result of this deficiency, we were not able to rely on certain data and reports used in the accounting for fixed assets, including server and network equipment, to ensure the completeness and accuracy of such information;
  107. 107
    www.sec.gov
    In our opinion, because of the effect of the material weaknesses described below on the achievement of the objectives of the control criteria, the Company has not maintained effective internal control over financial reporting as of December 31, 2025, based on criteria established in Internal Control -Integrated Framework (2013) issued by COSO. A material weakness is a deficiency, or combination of deficiencies, in internal control over financial reporting, such that there is a reasonable possibility that a material misstatement of the company’s annual or interim financial statements will not be prevented or detected on a timely basis. The following material weaknesses have been identified and included in management’s assessment: ● Material weakness in internal controls over property and equipment. ● Material weakness in internal controls over the revenue process for the TripleTen segment.
  108. 108
    www.sec.gov
    Based on the evaluation of our disclosure controls and procedures, our Chief Executive Officer and Chief Financial Officer concluded that our disclosure controls and procedures were not effective at the reasonable assurance level as of December 31, 2025 due to the material weaknesses in our internal control over financial reporting described below.
  109. 109
    www.sec.gov
    The deficiencies identified did not result in a material misstatement to our financial statements.
  110. 110
    suttonfinancialreporting.com
    1 All five companies prepare their accounts according to US GAAP. Under US GAAP, a change in the estimated life of depreciable assets is applied prospectively (FASB, ASC 250-10-45-17 Change in accounting estimate).
  111. 111
    ibinterviewquestions.com
    A change in useful life is a change in accounting estimate. It is applied prospectively, from the date of the change, without restating earlier periods, and its effect on income and earnings per share is disclosed when material, the treatment described in restatements, accounting changes and estimates. The estimate itself is a judgment based on expected use, technological change and the company's own experience, as explained in useful lives and salvage value.
  112. 112
    www.buxtonhelmsley.com
    If an issuer extends the useful life of an asset class from four years to six years, reported operating income increases in the period of change and in subsequent periods, while there is no corresponding adjustment to historical results that would allow an investor to compare current earnings to a prior baseline on like-for-like terms.
  113. 113
    pcaobus.org
    .04 The auditor should recognize the following matters relating to the consistency of the company's financial statements in the auditor's report if those matters have a material effect on the financial statements: 1. A change in accounting principle 2. An adjustment to correct a misstatement in previously issued financial statements.
  114. 114
    pcaobus.org
    The auditor should evaluate the reasonableness of the significant assumptions used by the company to develop the estimate, both individually and in combination. This includes evaluating whether: 1. The company has a reasonable basis for the significant assumptions used and, when applicable, for its selection of assumptions from a range of potential assumptions; and 2. The significant assumptions are consistent with the following, when applicable: 1. Relevant industry, regulatory, and other external factors, including economic conditions; 2. The company's objectives, strategies, and related business risks;17 3. Existing market information; 4. Historical or recent experience, taking into account changes in conditions and events affecting the company; and 5. Other significant assumptions used by the company in other estimates tested.
  115. 115
    www.sec.gov
    If there is a reasonable likelihood that reported financial information is not indicative of a company's future financial condition or future operating performance due, for example, to the levels of subjectivity and judgment necessary to account for highly uncertain matters and the susceptibility of such matters to change, appropriate disclosure in MD&A should be considered and may be required. For example, if a change in an estimate has a material favorable impact on earnings, the change and the underlying reasons should be disclosed so that readers do not incorrectly attribute the effect to operational improvements.
  116. 116
    www.buxtonhelmsley.com
    Both changes were characterized by management as the product of formal useful-life studies. Both were reviewed by audit committees and concurred in by independent registered public accounting firms. Both were disclosed under the same accounting standard.
  117. 117
    www.buxtonhelmsley.com
    When two issuers operating substantially similar hardware fleets arrive at directionally opposite conclusions on useful life in the same fiscal quarter, the variance in reported operating income that flows from those conclusions is not driven by economics-it is driven by the elasticity of the estimate.
  118. 118
    www.buxtonhelmsley.com
    In a series of public posts beginning November 11, 2025, the investor Michael Burry-best known for his pre-crisis short of the United States housing market-accused five public companies of having understated depreciation expense across the period from 2026 through 2028 by approximately $176 billion, with the result that, in his estimate, Oracle Corporation’s reported earnings for 2028 will be overstated by 26.9 percent and Meta Platforms, Inc.’s by 20.8 percent.
  119. 119
    www.buxtonhelmsley.com
    Buxton Helmsley does not adopt or endorse Mr. Burry’s specific dollar estimates, and would caution against doing so without a closed-form model that ties projected hardware vintages to amortization schedules at each named company.
  120. 120
    blog.citp.princeton.edu
    Investment analysts have run the numbers on what happens if accounting catches up to reality. Barclays, for example, is cutting earnings forecasts of AI firms by up to 10% for 2025 to account for more realistic depreciation assumptions. The physical evidence, the technological trajectory, and the emerging financial data all point in the same direction: actual useful life is at least half the accounting life, and this does not even account for the significant operational expenditures required.
  121. 121
    ibinterviewquestions.com
    Nvidia, whose chips fill many of those servers, disputed the claim in a memo to analysts, saying its customers depreciate graphics processors over four to six years based on actual longevity and use.
  122. 122
    fortune.com
    AWS CEO Andy Jassy has a very different take. In his latest annual letter to shareholders, he stated that AWS chips, servers and networking gear have a useful life of 5-6 years.
  123. 123
    ibinterviewquestions.com
    The companies that extended lives point to their own operating data. Before the current cycle, servers often ran longer than the three or four years originally assumed, as software improvements made older hardware more efficient and companies redeployed it to less demanding work. Alphabet, which moved to six years for servers and certain network equipment in 2023, reported that the change reduced its depreciation by $3.9 billion that year. Defenders of longer lives also argue that AI chips, once replaced at the frontier of training new models, can remain economically useful for years running inference, the work of answering queries, where older hardware can be good enough. Microsoft had made the same move six months earlier, extending server and network equipment from four to six years from July 2022, the start of its 2023 fiscal year, because investments in software had made the equipment more efficient to operate; the change added $3.7 billion to that year's operating income. Amazon lengthened its server lives from five to six years in 2024, cutting that year's depreciation by $3.2 billion, and then shortened a subset back to five years only twelve months later. A useful life is a running estimate, revisited as each company learns how long its equipment actually earns its keep.
  124. 124
    www.itiger.com
    If investors someday conclude that much AI investment is being wasted, it won't be due to the depreciation schedules companies have picked. While there are other good reasons to be skeptical of how technology giants are accounting for their AI adventures, the depreciation debate won't move the needle.
  125. 125
    ibinterviewquestions.com
    Because earnings before interest, taxes, depreciation and amortization (EBITDA) excludes depreciation, useful life assumptions do not affect it at all, and enterprise value to EBITDA multiples cannot capture the question. Price to earnings multiples, operating margins and returns on capital all depend on it. For capital-intensive technology companies, compare earnings-based measures with capital expenditure and free cash flow before drawing conclusions.
  126. 126
    d18rn0p25nwr6d.cloudfront.net
    Our U.S. taxable income is reduced by accelerated depreciation deductions and the resulting U.S. tax liability is reduced by tax credits, primarily related to the U.S. federal research and development credit.
  127. 127
    ornn.com
    Because most issuers use straight-line depreciation for data center equipment, increasing useful life mechanically reduces the periodic depreciation charge and increases reported operating income, while leaving EBITDA and cash flow unchanged. The relevant investor question is not whether longer lives are ‘right’ in the abstract, but whether disclosures provide enough evidence to underwrite (i) service-life reality, (ii) retirement behavior, and (iii) residual value realizations.
  128. 128
    aeacapitalresearch.com
    Early write-offs are the clearest evidence that a prior life estimate ran too long.
  129. 129
    ornn.com
    If management later reverses the estimate or retires assets early, the ‘missing’ depreciation can reappear abruptly via higher future depreciation, accelerated depreciation, or impairments.
  130. 130
    www.sec.gov
    Demand for cloud-based and AI products and services is evolving and difficult to forecast. Overestimation of demand or misalignment of capacity investments may result in underutilization of infrastructure and may lead to impairment of assets on our balance sheet.
  131. 131
    d18rn0p25nwr6d.cloudfront.net
    Further, any technical infrastructure in excess of our actual needs may lead to impairment of assets on our balance sheet or otherwise adversely affect our financial condition.
  132. 132
    s205.q4cdn.com
    Yeah. In terms of the capacity that's coming up for renewal, Samik, that is a very limited part of our fleet and the ASPs of the older generation remain higher or at levels that we've seen about a year ago. The second part of the piece that is very interesting in our business is as these fleets come off maturity, it allows us to have a great product in terms of managed inference to sell for our customers, which, as Mike noted, is a very fast evolving nature of our business, part of our business, which we expect to continue to grow rapidly and expect to have about $250 million of ARR by the end of the year.
  133. 133
    capefearadvisors.substack.com
    One ratio grades the whole: does EBITDA per dollar of in-service gross equipment rise, quarter over quarter, faster than the useful life of the equipment falls? Both numbers are filed every ninety days. The trend so far runs the wrong way on the broader measure: 12.5 percent to 11.7 to 8.6 to 9.8 on total gross equipment across the four quarters the filings cover, with the fourth quarter derived from the annual report less its nine months. The single upturn arrives in the strongest reported quarter, and the August print says whether it was a turn or a quarter. The in-service measure is the better gauge, and the construction-in-progress conversion is the bull case’s whole first leg.
  134. 134
    www.ifrs.org
    (b) significant changes with an adverse effect on the entity have taken place during the period, or will take place in the near future, in the technological, market, economic or legal environment in which the entity operates or in the market to which an asset is dedicated.
  135. 135
    www.ifrs.org
    (e) evidence is available of obsolescence or physical damage of an asset.
  136. 136
    media-stg.researchaffiliates.com
    Rental markets are imperfect signals. Spot pricing can overshoot in both directions, and hyperscalers may achieve higher utilization or lower financing costs than third-party renters. The point is not that a GPU fails after three years, but that its ability to earn excess returns may fade well before its accounting life ends. Economic obsolescence precedes physical obsolescence by several years.

Made with Deeperer. Editorial-grade research, cited inline

Ask your own question. Get a report like this one.

Start a report

More reports

Cite as: Deeperer. "GPU Depreciation Lives and Their Effect on Reported AI-Sector Profits." Deeperer, 31 Oct 2026. https://deeperer.com/r/gpu-depreciation-lives-and-their-effect-on-reported-ai-2

Produced by AI from the cited sources. Check it against them before you act on it.