The Depreciation Wall Is Real — But It's Sitting in Construction in Progress, Not in the Footnote

The Wall Is Real. It Is Just Not Where Everyone Is Looking.

Start with a line item almost nobody reads. In Alphabet's property and equipment note there is a row called "assets not yet in service." At the end of 2025 it stood at $78,592M. [5] Three months later, at 31 March 2026, it stood at $108,597M. [6] That is roughly thirty billion dollars of capital added to the balance sheet in a single quarter, and not one cent of it carries a depreciation charge today.

It will. Every dollar of it will begin to depreciate the moment it is placed in service, on a schedule that is already fixed — by the six-year useful life Alphabet uses right now, unchanged since 2023, with no change-in-estimate disclosure in either the 2025 10-K or the Q1 2026 10-Q. [5] [6]

Now add the neighbours. Meta's construction in progress: $61,017M. [15] Amazon's: $71,745M. [10] Oracle's: $39,973M. [17] Sum those four figures — and it is our sum, at mismatched period-ends, excluding Microsoft, which does not break out a construction-in-progress line in the filings we hold — and roughly $281bn of capital is sitting on four balance sheets that has not yet touched an income statement.

That is the wall. Not the footnote. The footnote is where everyone is looking, and it is the wrong place.

The bear thesis of the last eight months — Michael Burry's, and by now half the market's — rests on a single load-bearing clause: capex is not the story; the useful-life assumption is. Extend a server's life from four years to six and you defer depreciation, flatter earnings and build up a reckoning. The reckoning arrives when the assumption snaps back. Burry put a number on it: a "$176 billion understatement of depreciation projected between 2026-2028," with Oracle's earnings overstated by 26.9% and Meta's by 20.8% by 2028, premised on the view that "computing equipment typically has a 2-3 year product cycle." [26] He called the practice "one of the more common frauds of the modern era." [26]

The primary filings contradict the load-bearing clause on five separate fronts, and the first one is the most damning, because it is not a model. It is an experiment that already ran.

Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five, effective 1 January 2025, and it told the SEC exactly what that cost. [8] For the full year: an increase in depreciation and amortization expense of $1.4 billion, a reduction in net income of $1.0 billion, or $0.10 per basic share and $0.10 per diluted share. [10] Ten cents. Roughly 1.4% of Amazon's 2025 diluted EPS, in a year that EPS rose sharply.

The forecastable earnings revision the bear case promises has been measured, audited and published. It is the only realised data point in existence, and it is a rounding error.

What follows is the case for inverting the thesis: the depreciation shock is already locked in, at today's useful-life assumptions, and it is visible on the balance sheet rather than in the footnote. We give the bears their best arguments, name the one company where their case genuinely survives, and hand you a calculator so you can run the reversion yourself and see how small it is next to the backlog nobody is modelling.

Five Buyers, One Piece of Hardware, Five Answers

Begin with what the filings actually say, verbatim, and let the spread speak for itself.

Microsoft: "computer equipment, two to six years." [2] The company does not publish a discrete server life. The six-year server and network life it adopted in FY2023 — up from four years, disclosed as "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" — is embedded inside that range and has never been restated as a number since. [1] [2]

Alphabet: "We depreciate servers and network equipment generally over a period of six years." [5] Clean, single, unambiguous. It got there in January 2023, moving servers from four years to six and certain network equipment from five to six, worth "a reduction in depreciation expense of $ 3.9 billion and an increase in net income of $ 3.0 billion, or $ 0.24 per basic and $ 0.24 per diluted share." [4]

Amazon: "Servers and networking equipment Five to six years." [10] A ladder — four to five in January 2022, five to six in January 2024 (guided at "an anticipated impact to our 2024 operating income of $ 3.1 billion"), then six back down to five for a subset in January 2025. [7] [10]

Meta: "Servers and network assets Five to 5.5 years." [13] Not six. Several secondary outlets have reported Meta's life as six years; the filings say 5.5, twice.

Oracle: "Comprised primarily of servers and networking equipment with estimated useful life of six years." [17] It climbed there in two rungs — four to five in FY2023, five to six in FY2025 — making it the most-extended of the five. [16]

	Microsoft	Alphabet	Amazon	Meta	Oracle	CoreWeave
Current disclosed life (verbatim)[2][5][10][13][17][19]	"computer equipment, two to six years"	"servers and network equipment generally over a period of six years"	"Servers and networking equipment Five to six years"	"Servers and network assets Five to 5.5 years"	"servers and networking equipment with estimated useful life of six years"	"Technology equipment 6 years"
Last change in estimate[1][4][8][13][16][19]	4 → 6 years, effective FY2023	4 → 6 years (servers), 5 → 6 (certain network), Jan 2023	6 → 5 years on a SUBSET, effective 1 Jan 2025	extended to 5.5 years, effective 1 Jan 2025	4 → 5 (FY2023), then 5 → 6 (FY2025)	5 → 6 years, effective 1 Jan 2023
Disclosed per-share effect, year of change[1][4][10][13][16][19]	+$0.40 basic and diluted	+$0.24 basic and diluted	−$0.10 basic and diluted (the 2025 cut)	+$1.00 per diluted share	+$0.20 per diluted share	+$0.10 per share
Latest filing we checked[3][6][11][15][17][19]	Q3 FY2026 10-Q (quarter to 31 Mar 2026)	Q1 2026 10-Q	Q1 2026 10-Q	Q1 2026 10-Q	FY2026 10-K (year to 31 May 2026)	FY2025 10-K (filed 2 Mar 2026)
Has it shortened server lives?[3][6][10][13][17][19]	No	No	Yes — a "subset" only	No — it EXTENDED	No	No — it EXTENDED

Two things must be said alongside that table, and they pull in opposite directions.

The first is that the spread is narrower than the volume of the debate implies. Every audited answer in the set sits between five and six years. Nobody is at three. Nobody is at eight. Six filings, six independent audit committees, six different asset populations, and the answers cluster inside a twelve-month band.

The second is that GAAP licenses the divergence. Useful life is an entity-specific estimate over an entity-specific asset population. Two companies running GPU data centres can reasonably reach different conclusions depending on workload mix, data-centre design and thermal management — and any change must survive an external audit. Amazon changed "a subset of our servers and networking equipment." [10] Meta changed "most servers and network assets." [13] Those are not the same population, and the fact that they moved in opposite directions is therefore not, on its own, proof that the estimate is a lever rather than a measurement.

And then the disclosure gap that matters more than any of it: not one of the five discloses a separate useful life for GPUs or AI accelerators. They sit inside undifferentiated buckets — Microsoft's "computer equipment," Alphabet's "servers and network equipment," Amazon's "Servers and networking equipment," Meta's "Servers and network assets," Oracle's "Computer, network, machinery and equipment." [2] [5] [10] [13] [17] No investor can compute a GPU-specific depreciation charge from any 10-K in this group.

That matters because the asset mix is visibly shifting. Microsoft's own CFO, Amy Hood, on the Q3 FY2026 call: "this quarter, roughly two thirds of our capex was for short-lived assets, primarily GPUs and CPUs. The remaining spend was for long-lived assets that will support monetization over the next 15 years and beyond." [20] Two-thirds of the capex dollar is now going into the fast-obsolescing bucket — and the accounting policy that governs it has not moved. The mix shift is invisible in the footnote by construction.

That is the strongest thing the bears have, and we will come back to it. But it is a point about disclosure, not about arithmetic. To see why, you have to look at what the reversion actually costs.

The Reversion Already Ran. It Cost Ten Cents.

This is the heart of the inversion, and it deserves to be laid out slowly, because almost every retelling of it — including Burry's — gets some part of it wrong.

In Q4 2024, Amazon completed a useful life study. It concluded that a subset of its servers and networking equipment should move from six years to five, effective 1 January 2025. And it gave the reason, verbatim, in both the Q3 2025 10-Q and the 2025 10-K: "The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." [9] [10]

Read that again. It is the precise inverse of the reason every other company gave for lengthening. Amazon looked at AI, and concluded servers wear out faster. Everyone else looked at AI, and concluded servers last longer. Amazon is, on this question, the bears' own witness.

So what did the bears' own witness disclose?

Quarter by quarter, it is the only intra-year per-share depreciation trail anyone has published. Q3 2025 alone: an increase in depreciation and amortization expense of $392 million, a reduction in net income of $298 million, $0.03 per basic share and $0.03 per diluted share. [9] Nine months: an increase in depreciation and amortization expense of $889 million and a reduction in net income of $677 million, $0.06 per basic and diluted share. [9] Full year, in the 2025 10-K: an increase in D&A of $1.4 billion, a reduction in net income of $1.0 billion, $0.10 per basic and $0.10 per diluted share, "which primarily impacted our AWS segment." [10]

Ten cents. Roughly 1.4% of 2025 diluted EPS — our arithmetic — at the largest cloud provider on earth, in a year its diluted EPS rose sharply.

Two caveats make this honest, and both of them are the kind that get left out.

Amazon under-guided its own reversion by roughly half. The 2024 10-K anticipated "a decrease in 2025 operating income of approximately $ 0.7 billion." [8] The realised figure was $1.4 billion of extra D&A. [10] The gap is mechanical rather than sinister: the guided number was computed only on assets sitting in "Property and equipment, net" at 31 December 2024, while the realised number also captures every server Amazon bought during 2025. But the implication is important and it cuts against us: the cost of shortening scales with the capex ramp. Any forward estimate of a reversion built off a year-end asset base will understate it. In a year when capex is doubling, it will understate it badly.

And in the very same paragraph, Amazon extended something else. Heavy equipment went from ten years to thirteen, "an increase in 2025 operating income of approximately $ 0.9 billion, which will be recorded primarily in 'Fulfillment'." [8] That extension more than covers the $0.7 billion guided cost of the server shortening. Amazon's 2025 useful-life package, as guided, was a net positive to operating income.

Now take the maximal version of the bear case and run it. Suppose every one of these companies snapped its server life to four years tomorrow.

At Alphabet, using the two disclosed inputs — $203,679M of technical infrastructure in service, of which the 10-K footnote says "approximately 60 % of technical infrastructure assets were comprised of servers and network equipment" — the implied gross server base is roughly $122bn. That is our estimate, not a disclosed figure; Alphabet never states a gross server number. [5] Straight-lined at four years instead of six, that is roughly $10bn of incremental pre-tax annual depreciation. Against 2025 net income of $132,170M, it is a mid-single-digit percentage. [5]

At Meta, on $105,987M of gross servers and network assets at 31 March 2026, a snap from 5.5 years to four is roughly $7bn. [15] At Microsoft, on $132,836M of gross computer equipment and software — a bucket that is broader than servers and whose stated range is "two to six years," so treat it as an upper bound — roughly $11bn. [2]

These are our sums, crude straight-line arithmetic on gross bases, and they overstate the true first-year hit, because a change in accounting estimate is applied prospectively: to remaining carrying value over remaining life. No restatement. No catch-up charge. No surprise — it is disclosed at the start of the fiscal year in which it takes effect, as Meta's was, a full year in advance.

Mid-single-digit percentages of earnings, non-cash, telegraphed, applied forward only. That is not a wall. That is a haircut you can see in the mirror before it happens.

Run it yourself.

Alphabet (31 Mar 2026)Meta (31 Mar 2026)Oracle (31 May 2026)Reset

Gross server / network baseestimate122,000$M

Current useful lifefact6years

Reverted useful lifeassumption4years

Capital not yet in servicefact108,597$M

Annual depreciation at current life$20,333$M

Annual depreciation at reverted life$30,500$M

Incremental annual depreciation from the reversion$10,167$M

Annual depreciation the backlog adds once placed in service — at the UNCHANGED life$18,100$M

Backlog ramp ÷ reversion hit1.78×x

Illustrative straight-line arithmetic on the GROSS base, and it overstates the true first-year hit: a change in accounting estimate is applied PROSPECTIVELY, to remaining carrying value over remaining life — no restatement, no catch-up charge, and it is disclosed at the start of the fiscal year in which it takes effect. The Alphabet base is our estimate ($203,679M technical infrastructure × ~60%); Meta's ($105,987M gross servers and network assets) and Oracle's ($59,634M gross computer, network, machinery and equipment) are disclosed. The backlog figures are Alphabet's "assets not yet in service" and Meta's and Oracle's construction in progress. Note what the last row does: at the current six-year life, placing the existing backlog into service generates MORE incremental annual depreciation than a full snap from six years to four does on the entire existing server base.

Depreciation Is Already Exploding — With the Footnote Untouched

Depreciation is, to a first approximation, capital divided by life: D ≈ C/L. The bear case is a story about L. The filings say the entire move is in C.

Alphabet's depreciation of property and equipment went $11,946M in 2023, to $15,311M in 2024, to $21,136M in 2025. [5] The useful life did not change once across that period. Microsoft's depreciation expense went $11.0 billion, to $15.2 billion, to $22.0 billion across FY2023-25 — and the FY2023 figure was itself depressed by the extension, which is why depreciation actually fell in absolute terms that year, from $12.6 billion in FY2022. [1] [2] Oracle's went $3.1 billion, $3.9 billion, $7.6 billion across FY2024-26, on a six-year life it has held since FY2025. [17]

And the cleanest test of all is Meta, because Meta discloses the server line separately. Its servers and network assets depreciation expense was $7.32 billion in 2023, $11.34 billion in 2024, and $13.36 billion in 2025 — nearly doubling across exactly the period in which Meta was extending the lives of those very assets. [13] The extension slowed the growth rate. The line still ran away.

Set that against what is being spent. Microsoft: additions to property and equipment of $64,551M in FY2025, up from $44,477M and $28,107M — capex 2.3x in two years. [2] Alphabet: purchases of property and equipment of $91,447M in 2025, against $52,535M and $32,251M. [5] Amazon: $128,320M net of proceeds, up from $77,658M — and we use the net figure deliberately, because it is the one Amazon itself uses in its free-cash-flow reconciliation. [10] Oracle: capital expenditures of $55,663M, against $21,215M and $6,866M. [17]

Roughly $410bn of capex across the five — our sum, at fiscal years that do not align — against roughly $110bn of depreciation. Close to a four-times gap.

Source: Company Forms 10-K. Amazon capex is the NET figure ($128,320M, net of proceeds from sales and incentives) — the definition Amazon itself uses in its free-cash-flow reconciliation. Amazon depreciation includes amortisation of finance-leased property and equipment.

Oracle is the extreme: $55,663M of capex against $7.6 billion of depreciation, a ratio of about 7.3x. [17] Alphabet is at about 4.3x. And Alphabet's guided 2026 capex — "we are updating our full year 2026 CapEx guidance range to $180 to $190 billion" — is something like eight-and-a-half to nine times its actual 2025 depreciation. [21] [5]

The six-year life does not prevent any of that depreciation. It only defers when it arrives.

The companies are saying so themselves, in language that is not remotely evasive. Alphabet's 2025 10-K: "in 2026, we expect to significantly increase, relative to 2025, our investment in our technical infrastructure, including servers and network equipment, and data centers. The costs associated with operating our technical infrastructure - depreciation, energy, equipment, and network capacity - are expected to significantly increase." [5] CFO Anat Ashkenazi, unprompted, on the Q1 2026 call: "we're going to have the headwind associated with the depreciation coming with a higher CapEx level." [21] Meta's CFO ranks it first among all expense drivers for 2026: total expenses of $162-169 billion, and "the majority of expense growth will be driven by infrastructure costs, which includes third party cloud spend, higher depreciation, and higher infrastructure operating expenses." [22]

One housekeeping note, because it trips people up and honest arithmetic requires it. Cash-flow capex and management's capex are different numbers. Microsoft's Q3 FY2026 10-Q shows additions to property and equipment of $30,876M for the March quarter; Amy Hood said on the call that "capital expenditures were $31.9 billion," including "$4.7 billion" of finance leases, and separately that "cash paid for P, P, and E was $30.9 billion." [3] [20] Both are right. The same split applies to Meta. Always say which one you mean — and note that finance-leased datacentre assets are amortised over lease terms far longer than any server life, and no server-life change touches them at all.

The Real Wall: $281bn That Has Not Started Depreciating

Which brings us back to where we opened, and to the affirmative case.

Alphabet's "assets not yet in service" line reads: $50,597M, then $78,592M, then $108,597M. [5] [6] Zero depreciation charge against any of it. Meanwhile capex in the March 2026 quarter more than doubled year over year, from $17,197M to $35,674M, while quarterly depreciation of property and equipment went from $4,487M to $6,482M. [6] The gap between those two lines is not closing. It is opening.

Meta's construction in progress is $61,017M, up from $50,521M in a quarter, and its own footnote says it "includes costs mostly related to construction of data centers, network infrastructure and servers." [15] Amazon's is $71,745M, up from $46,636M in a year, against total accumulated depreciation and amortisation of only $177,073M on $534,098M of gross property and equipment. [10] Oracle's is $39,973M, up from $16,510M, against accumulated depreciation of just $22,694M on $122,651M of gross PP&E. [17] Oracle has barely begun to depreciate its AI fleet.

Microsoft does not break out a construction-in-progress line in the material we hold, so it is excluded from the $281bn — but its property and equipment, net went from $204,966M to $283,228M in nine months, on $80,146M of capex, while accumulated depreciation rose only from $93,653M to $111,723M. [3] The shape is identical.

And here is the thing that ought to end the argument about which lever matters. Take Alphabet's $108,597M of not-yet-in-service assets and place them into service at the unchanged six-year life. That single, entirely mechanical event generates more incremental annual depreciation than a full snap from six years to four does on the whole existing server base. The lever nobody is modelling is bigger than the lever everybody is.

The company that describes this mechanism most clearly is not one of the accused. It is CoreWeave, and its CFO says it out loud on a filed transcript: "when new capacity comes into service, data center lease costs, including power and depreciation expense commence, while customer revenue ramps over subsequent months... which means a corresponding increase in depreciation running ahead of associated revenue recognition." [24] Its construction in progress of "$9.4 billion... represents infrastructure not yet in service and not yet being depreciated." [24] Its 2026 capex is guided at "$30 billion to $35 billion" against revenue of "$12 billion to $13 billion" and adjusted operating income of only "$900 million to $1.1 billion." [24]

That is the wall, described in plain English, by a company nobody is accusing of hiding it. It has nothing to do with the useful-life assumption. It is a timing story about capital entering service.

Does the Hardware Actually Die? The Retirement Data Says No

Now test the premise underneath the bear case. If servers were economically dead at two or three years, the accounts would show it — because there are two lines that catch exactly that, and both go straight through the income statement, audited and disclosed.

The first is impairment. Meta's 2025 10-K: "During the years ended December 31, 2025, 2024, and 2023, total impairment losses for property and equipment were $ 237 million, $ 288 million and $ 738 million, respectively." [13] Falling — and falling on a base that grew from $164,663M of gross PP&E to $233,726M in a single year. [13] If useful lives were materially overstated, a writedown wave is the first thing you would expect. It is not there.

The second is accelerated depreciation on early retirement. There is exactly one real instance in the entire set: Amazon's 2024 10-K, "approximately $ 920 million of accelerated depreciation and related charges for the quarter ended December 31, 2024," with a further "approximately $ 0.6 billion in 2025." [8] Early retirement is the physical event that a useful-life extension implicitly denies — and when it happened, Amazon booked it, disclosed it and was audited on it. That is evidence the control works, not evidence that it has failed.

Then the counter-datum the piece cannot ignore, and which is, on its own terms, the single strongest fact on the companies' side of the ledger.

CoreWeave is the pure-play GPU landlord. Its entire business is renting out AI accelerators. If GPUs obsolesced in two to three years, CoreWeave would feel it first, hardest and in cash. CoreWeave's FY2025 10-K, filed 2 March 2026, states its policy in four words: "Technology equipment 6 years." [19] And it got there by extending: "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." [19] The disclosed effect 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." [19]

Small in dollars, because the base was small. But the direction is the whole point. The operator with the most to lose from GPU obsolescence uses a longer life than Amazon now uses for AWS servers, and has not shortened it. Either CoreWeave and its auditors are wrong, or the fast-obsolescence thesis is.

Its filed transcript goes further: "Average H100 pricing in Q4 was within 10% of where it started the year, while average A100 pricing increased in 2025." [24] The A100 launched in 2020. Its rental price went up in its fifth year. Management adds that it is signing older-generation SKUs "into new reserved instance contracts ahead of when it becomes available, firmly putting to bed concerns about demand for older generation SKUs." [24]

But that same transcript contains a line that points the other way, and we are not going to bury it.

One more datum the bear case ignores entirely. These companies depreciate servers straight-line to effectively zero salvage — yet a live resale market exists. Amazon's cash-flow statement books "Proceeds from property and equipment sales and incentives" of $3,704M in the twelve months to 31 March 2026. [11] Assets being depreciated toward nil are being sold for cash. The accounting, on that axis, is conservative.

And the serious counter-case to Burry is not "the chips still work." It is the value cascade, argued by the independent newsletter MBI Deep Dives: newest chips train, older chips do inference and general accelerated compute, and a hyperscaler — which "runs everything from cloud databases and video transcoding to scientific simulations" — can redeploy an idle training GPU where a pure-play AI lab cannot. [27] MBI's empirical claim is that observed Azure fleet lifetimes have run seven to nine years for the K80, P100 and P40 generations, and roughly 7.5 years for the V100 before its September 2025 retirement announcement — which would make the hyperscalers' five-to-six-year assumptions conservative, the exact inverse of Burry. [27] We flag this clearly: those are estimates asserted by an independent newsletter, unaudited and not in any filing. Attribute them to MBI Deep Dives. Do not state them as fact.

It also carries an important sting for the companies. If the six-year life is a claim about diversified demand rather than about silicon, then it is not a universal industry truth — it is a claim only the diversified operators are entitled to make. [27] Hold that thought. It is Oracle's problem.

The Best Case Against This Piece — Made Properly

The version of Burry that circulates on financial television is easy to beat. The version he actually published six days before this report is not, and it deserves to be quoted at length, because he explicitly disowns the strawman his critics attack:

"Rentability is not my claim, never has been. Depreciation is not a bet on when a chip stops working. It is the recovery of capitalized sunk cost over a defined window of earnings at the frontier, set against the terminal value. A chip can rent and still depreciate very fast economically. I never said the A100s would stop functioning." [25]

On that framing, every A100-still-busy datum in the section above is beside the point. Economic life is not physical life. A chip can be fully utilised and still be earning at a rate that fails to recover its capitalised cost over the window in which it sits at the frontier. Nothing in the retirement data settles that, and we should not pretend otherwise.

His evidentiary core is sharper still:

"Same hardware but opposite conclusions by two of the biggest. Depreciation cannot move two directions at once if it were a measurement of something. It is an economic lever, not a physical measurement." [25]

He is pointing at a real fact. Amazon's cut from six years to five was effective 1 January 2025. [8] Meta's extension to 5.5 years was completed "In January 2025... effective January 1, 2025." [13] Same hardware. Same month. Opposite directions. That is a fair inference. It is not a proof — the populations differ, "subset" versus "most," and GAAP permits entity-specific estimates over entity-specific assets — but it is fair, and anyone who waves it away is not arguing honestly.

Here are the bears' four best facts, all of which survive our verdict intact.

One: the extensions were real, and in aggregate they were large. Add the disclosed one-year net-income benefits — Microsoft $3.0 billion (FY2023), Alphabet $3.0 billion (2023), Meta $693 million (2022), Meta $2.59 billion (2025), Oracle $573 million (FY2025) — and you get roughly $9.9bn. [1] [4] [12] [13] [16] That is our sum; no filing states it, and it excludes Amazon's own extensions, which were disclosed at the operating-income line only ($3.1 billion for 2024). [7]

Two: the disclosure vanishes after year one. The benefit is disclosed once, in the year of change, and then disappears from the accounts — even though it recurs every year the asset lives. An investor reading only the current 10-K of Microsoft, Alphabet or Oracle sees no per-share depreciation benefit at all. The only companies with a live change-in-estimate paragraph in their latest annual report are Amazon (the cost of shortening) and Meta (the benefit of extending). [2] [5] [17]

Three: the asymmetry is grotesque, and it is the analytical core of the whole argument. Meta's half-year extension, to 5.5 years, was worth "$ 1.00 per diluted share" in 2025. [13] Amazon's full-year, one-year cut, on a subset of servers, cost "$ 0.10 per basic share and $ 0.10 per diluted share" in the same year. [10] Those two disclosures sit in filings one week apart. Ten to one. The credit taken for lengthening was an order of magnitude larger, per share, than the cost of the only shortening anyone has actually executed. (They are not like-for-like — Meta changed "most" servers, Amazon a "subset" — and we say so. It is still a ratio worth sitting with.)

Four: management is conceding the mix while denying the life. Amy Hood says two-thirds of the capex dollar is now short-lived GPUs and CPUs. [20] Anat Ashkenazi names the depreciation headwind out loud. [21] Neither has moved a useful life. The asset mix is shifting toward the fast-obsolescing bucket while the disclosed life stays put — and because no company discloses a GPU-specific life, that shift is invisible in the accounting policy by construction. That is the bears' argument, and it is being made by the CFOs.

Two things can be true at once, and an honest reader should hold both: the accounting is legal, disclosed, quantified, audited and defensible; and the aggregate effect on reported earnings is large and under-appreciated. What is not supported is the third thing.

Where the hypothesis genuinely survives: Oracle

We inverted the thesis. We did not annihilate it. There is one name in this set where the bear case is not neutralised by any of the counter-evidence above, and it is the name Burry singled out.

Oracle's gross "Computer, network, machinery and equipment" line doubled in a single year, from $30,345M to $59,634M, "comprised primarily of servers and networking equipment with estimated useful life of six years." [17] Against that, FY2026 depreciation of only $7.6 billion, and FY2026 capital expenditures of $55,663M — the widest capex-to-depreciation gap of the five, roughly 7.3x by our arithmetic. [17] Accumulated depreciation is $22,694M on $122,651M of gross PP&E. Net income is $17,087M. Free cash flow is negative $23,686M. [17]

Oracle has no consumer workload portfolio to cascade retired training chips into — no YouTube transcoding farm, no consumer search index, no advertising ranking fleet. The value-cascade defence is precisely the defence it cannot make. And it has told the SEC so, in a risk factor it wrote itself: "We typically depreciate these assets over their estimated useful lives, which could be shortened should our cloud strategies change, which could adversely affect our profitability." [17] No other company in the group frames the shortening of its own server life as a named risk.

Against GAAP earnings per share of $5.83 for FY2026, its FY2025 useful-life extension alone was worth $0.20 diluted — the highest ratio of the five by our arithmetic. [18] [16]

One structural twist must be netted off before anyone sizes the exposure. From the FY2026 results release: "Most of the RPO increase in both Q3 and Q4 were large scale AI contracts where the customer prepaid Oracle for the purchase of the GPUs, or the customer bought and supplied the GPUs to Oracle. The prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion." [18] A large slice of Oracle's AI GPU fleet may never enter Oracle's own depreciable base — and therefore generates no depreciation for Oracle at all.

Oracle is a named exception. It is not a thesis.

What Would Change Our Mind — and the Six Lines to Watch

State the verdict once more, narrowly, so it can be shot at.

The depreciation shock is already locked in, at today's useful-life assumptions. Roughly $281bn of capital sits on four balance sheets not yet depreciating, and it will begin to depreciate on schedules that are already fixed. The useful-life reversion the bears await is a second-order modifier on a first-order capex story — and the only time anyone actually executed it, it cost ten cents of Amazon EPS. Except at Oracle.

Here is what would prove that wrong, line by line, in documents you can pull yourself.

Line to watch	Where it lives	Latest reading	What would prove us wrong
Change-in-estimate footnote (contagion)	Note 1 / significant accounting policies, every 10-K and 10-Q	None — no new change at Microsoft (Q3 FY2026 10-Q), Alphabet (Q1 2026 10-Q), Meta (Q1 2026 10-Q) or Oracle (FY2026 10-K)	A second hyperscaler shortens. Then compare its disclosed per-share effect against Amazon's $0.10 benchmark before calling it material — if it lands at a multiple of that, our "reversion is survivable" conclusion is wrong.
Assets not yet in service / construction in progress	Property and equipment note, balance sheet	~$281bn across Alphabet ($108,597M), Amazon ($71,745M), Meta ($61,017M), Oracle ($39,973M) — our sume	If this line stops forecasting the depreciation ramp better than any useful-life change could, the assumption angle stops being a distraction. Until then it is the largest number in the story.
Depreciation expense, with the life unchanged	Property and equipment note; cash-flow statement	Alphabet $11,946M → $15,311M → $21,136M on an unchanged six-year life; Meta's server line $7.32bn → $11.34bn → $13.36bn while EXTENDING	If this stops compounding at 40-60% a year with no footnote change, then capex is not the driver after all and the useful-life assumption gets its status back.
Impairment losses on property and equipment	Property and equipment note	Meta: $738m (2023) → $288m (2024) → $237m (2025) — falling on a base that grew from $164,663M to $233,726M gross	A rising impairment line across the group. This is the direct, falsifiable evidence of early obsolescence, and it would confirm the bear case in the accounts rather than in a model.
Accelerated depreciation on early retirements	Property and equipment note; MD&A	Amazon: ~$920 million in Q4 2024, plus ~$0.6 billion through 2025 — the only real instance in the set	The same charge appearing at a second and third company. When assets genuinely fall out of use, the charge goes through the income statement and gets audited — the channel is not dormant, it is just quiet.
Free cash flow — the metric depreciation cannot touch	Cash-flow statement; company FCF reconciliations	Amazon $38,219M → $11,194M; Amazon TTM capex $151,003M vs operating cash flow $148,531M; Oracle NEGATIVE $23,686M against net income of $17,087M	Nothing — this is where we may be wrong in the largest possible way: not about the answer, but about the question. If the number that matters is cash rather than the depreciation footnote, then the whole argument is aimed at the wrong line of the accounts.

The most important row on that list is the last one, and it is the row on which we may be wrong in the largest possible way — not about the answer, but about the question.

Depreciation is non-cash. It is an allocation. It cannot, by construction, touch the one metric that is visibly and unambiguously breaking across this group right now.

Amazon's free cash flow fell from $38,219M in 2024 to $11,194M in 2025 — roughly two-thirds gone, while depreciation policy barely moved. [10] On a trailing-twelve-month basis to 31 March 2026, Amazon spent $151,003M on property and equipment against operating cash flow of $148,531M. [11] It is now spending slightly more on property and equipment than it generates from operations. Oracle's free cash flow is negative $23,686M against net income of $17,087M. [17] Microsoft's free cash flow in the March quarter was $15.8 billion against $31.9 billion of capital expenditures. [20]

Meanwhile the guidance keeps climbing. Microsoft: "For calendar year 2026, we expect to invest roughly $190 billion in capital expenditures which includes approximately $25 billion from the impact of higher component pricing." [20] Alphabet: $180 to $190 billion, with 2027 to "significantly increase" again. [21] Amazon: "We expect to invest about $200 billion in capital expenditures." [28] Meta raised its range $10bn in a single quarter, to $125-145 billion, explicitly because of "higher component pricing" — the same inflation Microsoft priced at $25 billion. [23] Component inflation is quietly swelling the depreciable base across the entire sector, and it will land in the same footnote everyone is arguing about, on the same six-year lives.

And Meta has made a promise that the arithmetic will test: "Despite the meaningful step up in infrastructure investment, in 2026 we expect to deliver operating income that is above 2025 operating income," repeated in Q1 2026. [22] [23] It is promising operating-income growth straight through the depreciation ramp — with quarterly depreciation on property and equipment already at $5.68 billion, up from $3.84 billion a year earlier. [15] Read its headline EPS with care while you watch: Meta's own CFO noted that Q1 2026 net income of $26.8 billion, or $10.44 per share, would have been $18.7 billion and $7.31 absent a tax benefit. [23] The $7.31 is the honest base against which any depreciation reversion should be measured.

To size the stakes at the outer limit: MBI Deep Dives cites The Economist's projection of a $2-4 trillion potential market-capitalisation reduction if useful lives were cut to one or two years. [27] That is an extreme scenario, a secondary citation of a third party's model, and we use it only to describe how loud this argument has become — not to settle it.

Because the argument, as posed, may simply be aimed at the wrong line of the accounts. The bears found a footnote and built a wall out of it. The wall was already there, four columns to the left, in construction in progress — and it is being poured faster than anyone is modelling, at useful-life assumptions nobody has touched.

Watch the capital enter service. That is where the depreciation is.