Useful Lives Under Scrutiny: The AI Capex Debate

Introduction

On January 1, 2025, two of the largest buyers of artificial intelligence (AI) computing hardware changed how long they expect their servers to last, in opposite directions. Meta lengthened the estimated life of most of its servers and network equipment to 5.5 years, which cut its 2025 depreciation by $2.92 billion and added $2.59 billion, or $1.00 a share, to net income. Amazon shortened the life of a subset of its servers from six years to five, citing the pace of development in AI, which added $1.4 billion to its depreciation and took $1.0 billion off its net income. Both changes are disclosed in the companies' 2025 annual reports, including Meta's.

The two decisions sit at the center of one of the most consequential accounting debates of the AI investment cycle. The largest technology companies are spending on data centers at a scale few companies have ever matched, and the life they assume for that equipment decides how quickly the cost reaches their earnings.

Capital spending by the largest technology companies has risen dramatically, by roughly 60% to 90% in each company's latest year alone:

Capital expenditure here is purchases of property and equipment in each company's latest annual cash flow statement, and it includes buildings and land as well as servers. Much of the spending is on servers, chips and networking equipment, the categories whose lives are in question. Because depreciation follows spending with a lag, the depreciation from today's investment will build up in the income statement over the next several years, and the assumed life sets how fast.

The gap between spending and depreciation is already wide. Meta's depreciation and amortization was $18.6 billion in 2025, about a quarter of its capital expenditure, and it carried $50.5 billion of construction in progress, assets not yet depreciated at all because they were not yet ready for use. Depreciation begins only when equipment is placed in service, and the cost of building the data centers, including capitalized interest, enters the asset base under the rules described in capitalization and cost basis of property, plant and equipment. The pattern of the charge itself, usually straight line for servers, follows the methods covered in depreciation methods.

Leased Capacity

Not all AI capacity is owned. The largest companies also lease data centers and computing capacity, and their commitments under leases that have not yet started are disclosed in the notes. Microsoft reported $329.1 billion of leases, primarily for data centers, that had not yet commenced at June 30, 2026, starting between its 2027 and 2033 fiscal years, and Meta reported about $103.8 billion at the end of 2025. When those leases begin, they will add right-of-use assets and lease liabilities to the balance sheets, with costs presented as depreciation and interest for finance leases or as lease expense for operating leases, the distinction explained in operating vs finance lease classification. For leased capacity, the lease term rather than a chosen useful life drives the expense.

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.

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.

Economic obsolescence

The loss of an asset's usefulness because newer technology, changes in demand or the cost of operating it make it uneconomic, even though it remains physically capable of working. For depreciation purposes, the useful life is the period over which an asset is expected to be used economically, so obsolescence can justify a shorter life than physical durability alone would suggest.

Critics have tried to size the effect. The investor Michael Burry argued in November 2025 that longer depreciation lives at the largest cloud and AI infrastructure providers would understate depreciation by about $176 billion between 2026 and 2028. 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. Whatever the right figure, the direction of the argument is clear: if AI servers have shorter economic lives than assumed, reported earnings in the coming years are higher than the economics justify, and impairments may follow.

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.

Two other rules come into play if lives prove too long. Assets are tested for impairment when events suggest their carrying amounts may not be recoverable, and an AI data center whose equipment is overtaken by new generations could face write-downs. And when a company decides to retire equipment early, it records accelerated depreciation over the shortened remaining life, so the cost reaches earnings sooner even without an impairment. Amazon did this at the end of 2024, recording about $920 million of accelerated depreciation and related charges in the fourth quarter after deciding to retire certain servers and networking equipment early. Either way, the full cost of the equipment is eventually expensed; the useful life decides when.

Analysts deal with the debate in several ways. They normalize depreciation across peers by applying a common useful life to each company's server and network equipment, using the asset balances and lives disclosed in the property note, so that differences in policy do not masquerade as differences in profitability. They compare capital expenditure with depreciation: when a company spends far more than it depreciates, as the largest technology companies currently do, reported earnings will be under growing pressure from depreciation in later years. And they focus on free cash flow, which deducts capital expenditure as it is spent and is unaffected by useful life choices. Returns on capital tell the same story from another angle: the assets of Amazon Web Services (AWS), Amazon's cloud segment, rose by more than $96 billion in 2025 while its operating income rose by under $6 billion, so its return on segment assets fell, the analysis set out in segment reporting. Whether that return recovers depends on revenue growth from the new capacity and on how long the equipment keeps earning.

The question will be settled by experience: how long AI hardware remains in productive use, and whether the companies keep or retire it. Until then, useful lives are one of the clearest examples of an accounting estimate that moves the earnings of the world's largest companies by billions of dollars, and one of the first items to check when reading their results.