Same GPUs, same month, opposite depreciation

TL;DR. In early 2025 Amazon shortened the estimated useful life of a subset of its servers from six years to five, citing the increased pace of technology development in AI, and took an accelerated depreciation charge reported at around $920 million. In the same period Meta extended most of its server and network lives to 5.5 years, booking a roughly $2.9 billion reduction in depreciation expense. Same Nvidia hardware. Both audited. Opposite conclusions, weeks apart. Across four US hyperscalers, $433.9 billion of property and equipment was purchased in the four quarters to March 2026 against about $149 billion of reported depreciation. Michael Burry argued in November 2025 that the sector would understate depreciation by $176 billion across 2026 to 2028; he held disclosed put positions while saying so. The useful life of a GPU is not a measurement. It is an assumption, and it lands directly on reported earnings.

---

Status: established facts, contested interpretation. The accounting changes, charges and capex figures are from company filings and earnings releases. The $176 billion estimate is one investor's projection made while holding disclosed short positions. Analyst counter-evidence is included. This article is descriptive and is not investment advice; nothing here evaluates any security.

---

The two decisions

Amazon, in 2025, shortened the estimated useful life of a subset of servers and networking equipment from six years to five, citing the increased pace of technology development, particularly in artificial intelligence and machine learning. Reported figures for the resulting accelerated depreciation charge vary between roughly $700 million and $920 million depending on which measure is quoted.

Meta, in the same window, extended the useful lives of most servers and network assets to 5.5 years, which reduced depreciation expense by approximately $2.9 billion.

The hardware is substantially the same. Both companies buy Nvidia data centre GPUs, run them in purpose-built facilities, and both had their assumptions reviewed by auditors.

One concluded the equipment wears out faster than it thought. The other concluded it lasts longer.

Why this is not fraud

Worth settling before going further, because the strongest public framing of this claim overstates it.

Michael Burry described the practice in November 2025 as understating depreciation by extending useful life, calling it one of the more common frauds of the modern era. That characterisation is wrong in a specific way and the specificity matters.

Useful life is an accounting estimate, not a fact to be discovered. Under GAAP, estimates are revised when new information arrives, and a revision is treated as a change in accounting estimate applied going forward. A restatement would imply the earlier figure was wrong when made. Amazon's change was the former, and its earlier assumption had been reviewed annually by its auditor.

And the underlying question is genuinely uncertain. A GPU's economic life depends on workload mix, data centre design, power and cooling costs, maintenance, and what the next generation offers. Older accelerators may be uneconomic for frontier training while remaining productive for inference for years. Two companies with different workloads can reasonably reach different answers, which is exactly what happened.

The disagreement is evidence of genuine uncertainty, not of misconduct. What it is not is a technicality, because of what follows.

The scale makes the assumption load-bearing

The four US hyperscalers purchased $433.9 billion of property and equipment in the four quarters to March 2026, against roughly $149 billion of reported depreciation over the same span. Quarterly combined capex reached $129.8 billion in Q1 2026, up around 80% year on year, with 2026 guidance summing to roughly $700 billion.

Alphabet spent $91.4 billion on capital expenditure in 2025, up from $52.5 billion, and has guided $175 to $185 billion for 2026.

Microsoft's property and equipment reached $298.6 billion at cost as of 30 June 2025, from $212.0 billion a year earlier, with depreciation expense rising $11.0 billion, $15.2 billion, $22.0 billion across fiscal 2023 to 2025.

Because depreciation recognises spend over five to six year server schedules and 25 to 40 year building schedules, today's income statements carry only a fraction of today's build-out. The gap is structural, and the recognition arrives later regardless of what AI revenue does.

That is what makes a one-year change in an assumption worth billions.

What the numbers do under a different assumption

J.P. Morgan's stress test found that applying a three-year depreciation life to AI hardware would reduce EPS and operating margins by roughly 6 to 8% for most hyperscalers, with Oracle a larger exception. Material, and not collapse-level, which is a more useful framing than either extreme.

Burry's estimate was $176 billion of understated depreciation across 2026 to 2028, with Oracle's earnings overstated by around 27% and Meta's by around 21% by 2028. One independent analysis using a four-year counterfactual across a broader asset base arrived at roughly $228 billion, in the same neighbourhood.

Burry disclosed put positions on Nvidia and Palantir while making the argument. That does not make it wrong, and it is the kind of disclosure a reader should weight, in the same way an interested party's own validation is weighted.

The counter-evidence

Goldman Sachs noted in April 2026 that A100 and H100 accelerators still command rental prices consistent with five to six year useful lives, which suggests the secondary market is not pricing rapid obsolescence.

That is the strongest counter available, because rental price is a market signal about remaining economic value rather than an accounting judgement, and it comes from participants with money at stake on both sides.

And Meta's own disclosure cuts against the simple story. Its 10-K reports server and network depreciation of $7.32 billion in 2023, $11.34 billion in 2024 and $13.36 billion in 2025, up 83% in two years despite the extension that deferred about $2.9 billion. Absolute depreciation is rising steeply even with the longer schedule, which is not what a company hiding costs would produce.

Three things this establishes

A number that determines reported profit is a judgement with a wide legitimate range. Two audited companies, same hardware, opposite conclusions, weeks apart. Anyone treating a depreciation line as a measurement has misread what it is.

The divergence is the most informative single data point. Not because either is wrong, but because it establishes the range is wide enough to contain both, which is the thing neither company's filing says on its own.

And the recognition gap is structural rather than a choice. $433.9 billion of purchases against $149 billion of depreciation is arithmetic about timing, not an accounting position. The expense arrives later whatever anyone assumes, and the assumption only determines how much arrives when.

What it does not establish

That any company has done anything improper. Every change described was disclosed, auditor-reviewed and treated under the applicable standard.

That the longer schedules are wrong. Goldman's rental-price observation is real evidence in the other direction, and inference workloads can keep older accelerators productive long after they stop being competitive for training.

That earnings will be restated. A change in estimate applies prospectively. Nothing in this points at a correction of past reporting.

And nothing about any security. This article describes an accounting assumption and its sensitivity. It makes no claim about valuation and no reader should treat it as guidance.

What is unresolved

What a GPU's economic life actually is. Nobody knows, because it depends on what the next generation offers and on whether inference demand keeps older hardware productive.

Whether the divergence narrows. If schedules converge, one of the two positions moved, and which one moves will be informative.

What happens when the wave lands. A single year's cohort at current scale adds substantial annual depreciation once fully in service, and more so on shorter lives. The timing depends on in-service dates and mix, neither of which is disclosed in enough detail to model from outside.

And whether disclosure improves. Microsoft's policy gives computer equipment a range of two to six years, which is wide enough to convey very little.

The counter-argument

Focusing on depreciation misreads where the risk sits. Capital expenditure is a cash outflow that has already happened; depreciation is its accounting echo. Investors can see the cash statement directly, and treating a non-cash allocation as the hidden problem inverts which number is real.

The Amazon and Meta divergence may be less meaningful than it looks. Their workload mixes differ, their data centre designs differ, and the assets covered by each change are not the same population. Two companies reaching different conclusions about different asset bases under different usage is ordinary, not remarkable.

Burry's framing invites the wrong reading. Calling a disclosed, audited, prospectively-applied estimate change a fraud makes the argument easier to dismiss than the underlying point deserves, and the underlying point, that the assumption is load-bearing, does not need the accusation.

And a three-year life may be too aggressive in the other direction. It benchmarks against frontier training economics, which is one use of the hardware. If most accelerator-hours end up serving inference, the accounting life may be closer to right than the critique assumes.

The short version

In early 2025 Amazon shortened server useful life from six years to five, citing the accelerating pace of AI development, and took an accelerated depreciation charge reported between roughly $700 million and $920 million. In the same window Meta extended most server and network lives to 5.5 years, reducing depreciation expense by about $2.9 billion.

Same hardware. Both audited. Opposite conclusions.

This is not fraud. Useful life is an accounting estimate revised prospectively when new information arrives, not a fact to be discovered, and the underlying economic life genuinely depends on workload, design and what the next chip generation offers. The disagreement is evidence that the range is wide, which is the most informative thing either filing tells you.

And the assumption is load-bearing. Four hyperscalers bought $433.9 billion of property and equipment in the four quarters to March 2026 against about $149 billion of reported depreciation, with 2026 guidance near $700 billion. J.P. Morgan found a three-year life would cut EPS and operating margins by roughly 6 to 8%. Michael Burry put the sector's understated depreciation at $176 billion across 2026 to 2028, while holding disclosed put positions.

The counter-evidence is real too. Goldman Sachs observed in April 2026 that A100s and H100s still rent at prices consistent with five to six year lives, and Meta's own server depreciation rose 83% in two years despite the extension.

What is certain is the timing. Purchases exceed recognised depreciation by a wide margin, so the expense arrives later regardless of what AI revenue does. The assumption decides how much arrives when, and two audited companies looking at the same chips could not agree on it.

Common questions

What actually happened in early 2025? Amazon shortened the estimated useful life of a subset of its servers and networking equipment from six years to five, citing the increased pace of technology development particularly in artificial intelligence and machine learning, and took an accelerated depreciation charge reported at between roughly $700 million and $920 million depending on the measure quoted. In the same window Meta extended the useful lives of most of its servers and network assets to 5.5 years, reducing depreciation expense by approximately $2.9 billion. The hardware in both cases is substantially the same class of Nvidia data centre accelerator.

Is this accounting fraud? No, and describing it that way is the weakest version of the argument. Useful life is an accounting estimate, not a fact to be discovered, and under GAAP estimates are revised prospectively when new information arrives. That is a change in accounting estimate rather than a restatement, which would imply the earlier figure was wrong when made. Amazon's earlier assumption had been reviewed annually by its auditor. The genuine point is that the estimate is load-bearing, and that point does not need the accusation.

Why does a depreciation assumption matter so much? Because of scale and timing. The four US hyperscalers purchased $433.9 billion of property and equipment in the four quarters to March 2026 against roughly $149 billion of reported depreciation, with 2026 guidance summing to about $700 billion. Depreciation recognises that spend over five to six year server schedules, so current income statements carry only a fraction of the current build-out. A one-year change in the assumption moves billions of expense between periods.

What did Michael Burry claim? He argued in November 2025 that hyperscalers were understating depreciation by extending useful lives beyond what two to three year product cycles justify, estimating $176 billion of understatement across 2026 to 2028, with Oracle's earnings overstated by around 27% and Meta's by around 21% by 2028. He disclosed put positions on Nvidia and Palantir while making the argument, which is the kind of interest a reader should weight without treating it as refutation.

What is the strongest evidence against the claim? Goldman Sachs noted in April 2026 that A100 and H100 accelerators still command rental prices consistent with five to six year useful lives, suggesting the secondary market is not pricing rapid obsolescence. Rental price is a market signal about remaining economic value rather than an accounting judgement, made by participants with money at stake. Separately, Meta's own 10-K shows server and network depreciation rising from $7.32 billion in 2023 to $13.36 billion in 2025, up 83% in two years despite the extension.

How much would earnings move under a shorter life? J.P. Morgan's stress test found that applying a three-year depreciation life to AI hardware would reduce EPS and operating margins by roughly 6 to 8% for most hyperscalers, with Oracle a larger exception. That is material without being catastrophic, which is a more useful characterisation than either the dismissal or the alarm.

Why do two companies disagree about identical hardware? Because economic life depends on more than the chip. Workload mix, data centre design, power and cooling costs, maintenance practice and what the next generation offers all bear on it, and an accelerator that is uneconomic for frontier training can remain productive for inference for years. Two companies with different workloads can reasonably reach different estimates, and the fact that they did is the clearest available evidence that the legitimate range is wide.

What should someone watch next? Whether the schedules converge, and in which direction. If the longer schedules shorten, the market has revised its view of economic life; if Amazon's lengthens back, the opposite. Also worth watching is disclosure quality: Microsoft's policy gives computer equipment a range of two to six years, which is wide enough to convey very little about what is actually assumed.