The accounting
tell.

Amazon shortened its server depreciation life this year and took a $700 million hit. Meta lengthened its server life and booked a $2.9 billion gain. Same workloads. Same chips. Opposite calls. One of them is wrong.

→  Good morning

On February 7, 2025, in a sentence buried in its annual report, Amazon revealed that it had shortened the useful life of its servers from six years back to five. The company took a $920 million accelerated depreciation charge in the fourth quarter for equipment it was retiring early, and warned that the new estimate would reduce 2025 operating income by another $1.3 billion. The reason given was specific: AI workloads are wearing out the hardware faster than expected.

Six weeks earlier, Meta did the opposite. The company extended the assumed useful life of its servers from four-and-a-half years to five-and-a-half. The disclosure was a paragraph in the Q4 2024 10-K. The benefit to operating income, also disclosed, was about $2.9 billion in 2025 alone.

Two of the largest hyperscalers in the world, running similar AI workloads on similar Nvidia chips inside similar data centers, looked at the same physical evidence and made opposite accounting decisions in the same calendar quarter. The total swing between them, in reported earnings, was around four billion dollars. That is not a small detail.

How we got here.

For most of the cloud era, hyperscalers depreciated servers over three years. Then, beginning in 2020, the assumption began to drift higher. Microsoft moved from three years to four, then to six. Alphabet went from three to four to six. Amazon went from three to four, then five, then six. Each step was disclosed. Each step was justified.

Every extension produced a one-time boost to reported operating income. Microsoft's 2022 change from four years to six added approximately $3.7 billion to fiscal 2023 operating income. Amazon's 2024 change from five years to six contributed about $900 million per quarter. Across the four largest hyperscalers, the cumulative reduction in annual depreciation expense from these decisions ran into the tens of billions of dollars by 2024.

The disclosed justification was always the same. Software optimization. Better cooling. More efficient utilization. Servers were lasting longer in practice than the original three-year estimate had assumed. The argument was reasonable, and it produced beautiful operating leverage in cloud businesses where infrastructure is the dominant line item.

Then the AI buildout arrived, and the math changed.

What Sloan figured out in 1996.

Richard Sloan was a young accounting professor at the University of Pennsylvania when he published a paper in 1996 in The Accounting Review that quietly changed how serious investors read financial statements. The question Sloan asked was this: When a company reports earnings, how much of those earnings comes from cash that actually moved, and how much comes from accounting estimates?

Cash earnings are unambiguous. The customer paid and the cash arrived. Accruals, however, are everything else: revenue recognized but not yet received, expenses recognized but not yet paid, and most importantly, the depreciation policies that determine how quickly capital spending flows through the income statement.

Sloan ran the comparison across roughly 30 years of U.S. corporate filings. His finding has been replicated dozens of times since. Earnings produced by cash flows persist into the future. Earnings produced by accruals do not. A long-short strategy that bought low-accrual firms and shorted high-accrual firms generated abnormal returns of roughly 10 percent per year in his sample, with the effect concentrated in the year after the financial statements were filed.

The mechanism Sloan identified is when a company reports an earnings number that depends heavily on optimistic estimates, the optimism eventually has to be paid back, either through future writedowns, reduced future earnings, or both. The market, in his data, was slow to recognize the difference between high-quality cash earnings and lower-quality accrual earnings. Investors who could read these details captured the alpha.

Q1 2025  ·  Same Quarter, Opposite Calls

+$2.9B / -$1.3B

Meta extended server life by 12 months. Amazon shortened by 12 months. Same chips. Same workloads.

Why this divergence is unusual.

The interesting feature of the Amazon-Meta split is not that two companies disagreed about something. The interesting feature is that for fifteen years before this, they did not. Microsoft, Alphabet, Amazon, and Meta extended server lives in roughly synchronized fashion through 2022, 2023, and 2024. When one company moved, the others followed within twelve months. The peer pressure of comparable disclosure produced a long, steady drift in one direction.

The Amazon Q1 2025 reversal broke that pattern for the first time. The company explained the change like this: AI workloads stress hardware in ways general cloud workloads do not. GPUs running near full utilization for continuous model training fail more often. Networking equipment routing AI traffic at scale wears differently. The company's internal “useful life study,” referenced in the disclosure, had concluded that six years was too generous.

Meta, doing roughly the same physical work, came to the opposite conclusion in the same window. The company extended its assumed useful life by another year. Both companies cannot be right.

The honest read is that Amazon's accounting now reflects what the engineers are seeing in the data centers, and Meta's accounting reflects what the income statement needs to absorb the largest capex cycle in corporate history. Meta's capex-to-revenue ratio is approaching 36 percent, the highest among the hyperscalers. The longer the company can stretch the depreciation, the smaller the headline drag from that capex appears in any single year. The benefit is non-cash. The economic reality of the hardware does not change. Only the timing of when the cost shows up in earnings does.

How to read this data.

Every 10-K and 10-Q filing of a company with material capital spending discloses depreciable useful lives, usually in the “Property and Equipment” footnote. For most companies, the assumption holds steady for years and the disclosure is boilerplate. When it changes, the disclosure becomes informative, and the dollar impact is almost always quantified in the same paragraph.

Three patterns are worth watching for. Lengthening a useful life produces a one-time tailwind to operating income that decays over the new, longer life of the asset. The benefit is largest in the year of the change and smallest in the final year. Repeated lengthenings within a few years are a flag. Meta has now extended useful lives three times in three years. Each extension occurred precisely as AI capex ramped further.

Shortening a useful life produces an immediate hit, often combined with an accelerated depreciation charge for assets being retired early. The hit is painful in the quarter of the change but signals that management is taking a more conservative posture about the underlying economics. It is, by Sloan's framework, the higher-quality earnings outcome.

Divergence within a peer group is the most informative pattern. When two companies operating in the same business with the same hardware disagree about useful life by a year or more, the more conservative number is usually closer to the truth. The other company is using the wider assumption to flatter earnings, and the eventual catch-up will arrive as a future write-down or a future operating income shortfall. The market may not price in the difference for years. By Sloan's data, that is exactly the lag in which the conservative-accruals firm outperforms the aggressive-accruals firm.

→  The takeaway

Tuesday's memo was about the language inside filings. Today's is about the numbers. The market reads the headline. The information sits in the footnote.

Amazon and Meta are both reasonable companies run by capable management teams. They cannot both be reading the physical evidence correctly. One of the two will be writing down hardware in 2027 or 2028 that the other already wrote down in 2025. The footnote is where you find out which.

Have a good Wednesday. Tomorrow we shift from how earnings are constructed to how rare it is for any company to compound them at all.

— Marques

About the author

Marques Blank runs Blank Capital, a fractional CFO and FP&A advisory. Previously Northrop Grumman and Citibank. CMA, MBA, Series 65.

Sources

Sloan, R.G. (1996). Do stock prices fully reflect information in accruals and cash flows about future earnings? The Accounting Review, 71(3), 289–315.  ·  Penman, S.H., & Zhang, X.J. (2002). Accounting conservatism, the quality of earnings, and stock returns. The Accounting Review, 77(2).  ·  Amazon.com Inc. Form 10-K (Feb 7, 2025) and Form 10-Q for Q1 2025; Meta Platforms Inc. Form 10-K for FY 2024; Microsoft Corporation FY 2022 and FY 2023 earnings disclosures; Alphabet Inc. depreciation policy disclosures, 2021–2024.

For informational purposes only. Not investment advice. Specific securities mentioned are case studies, not recommendations. Dollar impact figures are taken directly from company filings. Past performance does not predict future results.