In 2025, Meta extended the estimated useful lives of most servers and network assets to 5.5 years. The change reduced annual depreciation expense by $2.92 billion and increased net income by $2.59 billion, calculated using the relevant assets already in service at the end of 2024.1
In the same year, Amazon shortened the estimated lives of a subset of servers and networking equipment from six years to five. That increased depreciation and amortization expense by $1.4 billion and reduced net income by $1.0 billion, primarily affecting AWS. Amazon attributed the change to faster technological development, particularly in artificial intelligence and machine learning.2
Both adjustments are in the filings, but they cover different asset populations. Asset scope and use may affect useful-life estimates, yet these disclosures lack matched fleet data to establish whether such differences explain the opposite adjustments. The depreciation schedules alone cannot establish which GPU fleet lasts longer.
An earlier short commentary looked at an equipment purchase from the supplier’s and platform’s sides. A cash purchase records the payment, then allocates cost over the service period. Replacement depends on what the equipment can still do. Depreciation estimates address that service period. Cash recovery and technological obsolescence require further evidence.
When a cloud company buys a server for cash, qualifying equipment is generally recorded as an asset and the payment appears in investing cash flow. Once the equipment is ready for its intended use and enters service, its cost is allocated to earnings through depreciation. Periods benefiting from the equipment share the cost, which requires management to estimate how long it will serve.31
The following illustration compares two assumptions made at initial purchase. An asset costs 100 accounting units, has no residual value, enters service at the start of the year and uses straight-line depreciation. All other conditions are held constant.
| Initial estimated life | Annual depreciation | Cash already paid at purchase |
|---|---|---|
| Four years | 25 | 100 |
| Six years | Approximately 16.67 | 100 |
A longer life produces a smaller annual charge, with no change in equipment price or sales. The same cost is spread over more periods. Revising the life of equipment already in use involves a different calculation: its remaining carrying amount is depreciated prospectively over the revised remaining life. Dividing the original 100 by six does not allow earlier accounts to be rewritten. Meta’s accounting policy expressly uses prospective treatment.1
Under the indirect method, operating cash flow starts with net income, adds back noncash charges such as depreciation and adjusts for working capital and income-tax items. Lower depreciation raises pretax profit, while book tax expense may also change. Even with cash taxes unchanged, higher net income and a smaller depreciation add-back need not offset on their own; deferred-tax and other relevant adjustments must enter the reconciliation. Holding cash taxes and other cash flows constant, with the relevant tax items fully reconciled, explains why total operating cash flow does not change with this noncash charge. Book and tax depreciation can differ. Reconciliation does not bring back the equipment payment.31
That distinction changes how the opening earnings figures should be read. Meta’s $2.59 billion increase in net income cannot be treated as extra cash collected by AI services. Amazon’s $1.0 billion reduction does not show that AWS paid equivalent extra cash because equipment became obsolete. The figures measure annual earnings sensitivity to the estimate changes; recovery of the cost of individual equipment batches remains unproven.
Amazon’s cash flow statement supplies a separate set of figures. Consolidated operating cash flow in 2025 was $139.514 billion. Deducting net cash purchases of property and equipment of $128.320 billion produced reported free cash flow of $11.194 billion. The purchase measure is net of sales proceeds and incentives. This company-defined non-GAAP measure does not deduct all lease principal payments or describe the cash left after every funding requirement.2
The scope is Amazon as a whole, so AWS or AI returns cannot be isolated from it. These figures compare operating cash generation with cash invested in property and equipment and the remainder under the company’s definition. Separately, a capitalized cash purchase is expensed over its service life, so payment and expense recognition may fall in different periods. The cash flow bridge above does not measure that timing difference. When capacity enters service, how much revenue it produces and the cash required to sustain it determine whether the expenditure can be recovered.
The annual difference between capital expenditure and depreciation also mixes equipment from several periods. Capital expenditure includes new capacity, replacement equipment and long-lived buildings, while depreciation includes assets bought in different years. Measuring the fleet’s economic consumption still requires their ages and business uses.
Microsoft’s annual report for the year ended June 30, 2026 lists estimated useful lives of two to six years for servers and network equipment. This is a category range; the table does not identify each GPU generation.4
On the July 29 earnings call, CFO Amy Hood announced that data center and office building lives would be extended from fifteen to twenty-five years at the start of fiscal 2027. Microsoft’s fiscal 2027 began on July 1, 2026. Twenty-five years applies to buildings. The accelerators inside the racks belong to a different asset category.5
Hood’s explanation concerned capital expenditure presentation more than the depreciation charge. Microsoft expected only a minimal direct benefit to fiscal 2027 operating income, but more future data center leases would move from finance to operating classification. The company’s capital expenditure measure includes finance leases and excludes operating leases, so classification affects the number. She also said that calendar 2026 investment expectations were unchanged outside the useful-life update’s effect.5
The reported capital expenditure figure could therefore shrink without a corresponding reduction in physical construction. Acquiring fewer assets and changing lease classifications have different implications for equipment demand. This disclosure describes the latter.
Lease classification can also affect cash flow presentation, which prevents direct application of the simple depreciation illustration above. Building, buying or leasing a facility changes payment schedules, accounting classifications and risk allocation. Cash purchases, finance leases and operating leases need to be considered together to reconstruct the economic investment.
CoreWeave’s 2025 annual report assigns technology equipment a six-year estimated life. Its second-quarter 2026 risk disclosures acknowledge that revised life assumptions, or an inability to redeploy equipment after its original contracted life, could affect financial results. Six years remains an estimate that subsequent use must support.67
In financing presentation materials publicly furnished to the SEC on September 17, CoreWeave offered an example: an NVIDIA A100 contract extends into 2029, with the customer using the equipment for scientific research, producing and refining datasets and fine-tuning models. NVIDIA announced that A100 was in full production and shipping in May 2020.89
The older model still has a role in this contract. Its launch year, however, does not establish when those machines were bought or entered service. A contract extending into 2029 does not prove that a particular GPU has operated continuously for nine years.
The example comes from an equipment operator seeking financing and presents a successful renewal. This exhibit omits the full contract, fleet size and complete pricing. It offers a company-disclosed case of older-model demand, without providing utilization for an entire generation or validating a six-year useful life for every asset. The life is a cost-allocation estimate; recovering the original expenditure requires operating records.
Installed, energized equipment already adapted to software may produce positive cash contribution more readily by continuing to serve familiar workloads than by being shut down while the facility is retrofitted. Customer migration also costs money, and scientific computing, fine-tuning and some inference workloads need not use the newest system. These conditions give older equipment a business role. Its contribution still depends on service pricing, maintenance and electricity needed to keep it running. Whether the original purchase cost has been fully recovered is a separate calculation. Buying the next fleet is another expenditure again.
More efficient new systems change that comparison. With power and facility capacity scarce, a new system serving more paying demand from the same power allocation raises the opportunity cost of keeping older machines. The public evidence does not settle how that cost compares with the contribution from continued use. Specific workload and cash-cost data are missing.
The filings of Meta, Amazon, Microsoft and CoreWeave supply accounting policies and operating examples, yet servers and network assets remain largely aggregated categories. Complete generation-level data on GPU age, idle time, realized rental prices, maintenance costs and retrofit spending are generally absent. CoreWeave’s renewal example also applies directly to a rental business. Equipment used internally needs to be tested against its contribution to the platform’s own business, rather than a common GPU renewal test.
The notes can signal a changing estimate. Amazon, Meta and Alphabet describe ongoing or regular useful-life evaluation: an unchanged number does not show that reassessment stopped, and a shorter estimate need not mean the machines have already been retired.2110 Actual records of widespread early retirement or idling would directly weaken an extended-life assumption. For a rental fleet, renewal revenue covering the cash expenses needed to keep it operating tests the value of retention. Internally used equipment requires attributable business benefits and costs. Neither test automatically establishes recovery of the original purchase price.
A six-year depreciation period can be reassessed as operating evidence develops. One successful contract or an annual report without material impairment charges cannot carry that validation. Continued usefulness, positive cash contribution from retention and eventual recovery of the full expenditure require the operating accounts of specific equipment batches.
This article represents KSINQ’s personal market research and opinion, for informational purposes only, and does not constitute investment advice. Companies are discussed as analytical examples, without recommendations to buy, sell or hold their securities. Historical financial results do not indicate future performance. Company life estimates, individual contracts and the article’s conditional inferences do not guarantee future cash returns. Research cutoff: October 4, 2026.
Meta, 2025 Form 10-K, filed January 29, 2026, fiscal year ended December 31, 2025. Note 1, Use of Estimates, printed p.94; Property and Equipment, pp.98–99. The $2.92 billion depreciation reduction and $2.59 billion net income increase are annual actual effects based on relevant assets in service at December 31, 2024. The server and network asset range in the policy table is five to 5.5 years. ↩ ↩2 ↩3 ↩4 ↩5
Amazon, 2025 Form 10-K, filed February 6, 2026, fiscal year ended December 31, 2025. Note 1, Use of Estimates, printed p.41; MD&A, Free Cash Flow, p.28. The life reduction applies to a subset. Free cash flow is the company’s non-GAAP operating cash flow less net cash purchases of property and equipment, rather than an AI-only cash flow measure. ↩ ↩2 ↩3
SEC, Beginners’ Guide to Financial Statement, February 4, 2007, Cash Flow Statements / Operating Activities and Investing Activities. The table is an original KSINQ accounting illustration in hypothetical units, rather than company data or a prediction of equipment life. ↩ ↩2
Microsoft, FY2026 Form 10-K, filed July 29, 2026, fiscal year ended June 30, 2026. Note 1, Accounting Policies, Property and Equipment. The two-to-six-year category does not separately identify GPUs. ↩
Microsoft, FY2026 Q4 earnings conference call, July 29, 2026, Amy Hood’s prepared remarks before the fiscal 2027 outlook. Building lives, operating-income effects and lease classification are management’s contemporaneous explanation, rather than realized fiscal 2027 results. ↩ ↩2
CoreWeave, 2025 Form 10-K, filed March 2, 2026, fiscal year ended December 31, 2025. Note 1, Property and Equipment, printed p.94; Impairment of Long-Lived Assets, p.96. Six years is the company’s technology-equipment accounting estimate. ↩
CoreWeave, 2026Q2 Form 10-Q, filed August 12, 2026, quarter ended June 30, 2026. Part II, Item 1A, Risks Related to Our Business and Industry, discussing replacement, estimated lives and equipment redeployment. ↩
CoreWeave, SEC 8-K Exhibit 99.2, publicly furnished September 17, 2026, slide 10. The article paraphrases a single company-disclosed renewal example. The complete contract was not obtained and renewal economics were not independently audited. The case is not an industry pricing or whole-fleet useful-life estimate. ↩
NVIDIA, A100 full-production and shipping announcement, May 14, 2020. Model launch does not establish the date a particular asset entered service. ↩
Alphabet, 2025 Form 10-K, filed February 5, 2026, fiscal year ended December 31, 2025. Note 1, Property and Equipment, describes regular evaluation of technological obsolescence, planned use and utilization. ↩