This is BMInsider PRO content
You're reading a Deep Dive — our most in-depth research. Unlock full access to all analyses, Smart Money breakdowns & more.
Unlock BMInsider PRO →One-time payment · instant access · no account needed
In the second quarter of 2026, Meta grew revenue 28 percent to $60.8 billion. Net income fell 14 percent. Nothing sits between those two numbers except an assumption: the number of years over which a data center is written off. It is not measured, it is estimated, it appears in small print in the notes to the accounts – and since 2022 it has manufactured more reported profit than most products ever have.
No single line item in modern financial reporting has moved as much earnings with as little scrutiny as the estimated useful life of servers and data centers. It explains why the margins of the largest technology companies expanded between 2022 and 2025 while their capital spending exploded. It also explains why those margins are contracting now. What follows takes the mechanism apart, lays out every disclosed change of the past four years with its stated dollar effect, gives the counterarguments the weight they deserve – and ends with four checks any investor can run without taking anybody at their word.
Two Numbers From the Same Quarter
Meta’s July 2026 report is the cleanest specimen this cycle has produced. Revenue rose 28 percent to $60.801 billion. Operating income came in at $18.775 billion, a 31 percent margin against 43 percent a year earlier. Net income fell to $15.848 billion and diluted earnings to $6.18 per share. Twelve margin points surrendered in four quarters, on revenue growth approaching thirty percent, is not an accident. It is a change in the shape of the cost base.
The cause sits two lines below. Depreciation and amortization reached $6.356 billion in the quarter, up 46 percent year over year, while capital expenditure ran at $31.08 billion. Operating cash flow of $31.9 billion was almost entirely consumed by that spending; free cash flow finished at $784 million against reported net income of $15.8 billion. A company reporting twenty times more profit than cash is telling two different stories about the same three months.
In the same summer those numbers appeared, Microsoft extended the estimated useful life of its data centers and office buildings from 15 to 25 years, effective at the start of fiscal 2027. The two events belong together. At Meta the bill is arriving. At Microsoft it is being spread further out.
What a Useful Life Actually Is
Buy a server for a thousand dollars and you have spent a thousand dollars. The income statement will not show it. The amount goes onto the balance sheet and is fed into earnings in slices across the estimated useful life. Cash is indifferent to that estimate – the money left on the day of purchase. Reported profit is anything but indifferent.
The scale becomes visible on a single billion dollars of hardware. The gap between a three-year and a six-year assumption is $167 million of annual expense per billion. Applied to the hundred-billion-dollar annual build-outs now underway at individual companies, it becomes an eleven-figure number.
| Assumed useful life | Annual depreciation per $1bn | Difference vs. six-year assumption |
|---|---|---|
| 3 years | $333m | $167m more expense |
| 4 years | $250m | $83m more expense |
| 5 years | $200m | $33m more expense |
| 5.5 years | $182m | $15m more expense |
| 6 years | $167m | reference |
What matters is the character of the number. It is not a measurement but a management assertion about how long equipment will remain economically useful. An auditor tests it for reasonableness, not for accuracy – accuracy is not available for statements about the future. It flows instantly into every metric built on earnings: margin, earnings per share, price-to-earnings, return on equity. It never touches cash.
A Chronicle of Extensions
Between 2020 and 2026 the largest buyers of data center equipment moved their assumptions almost uniformly in one direction. Every change was properly disclosed, with a dollar figure attached. Placed side by side, they produce a picture no individual filing conveys.
| Company | Effective | Change | Effect as stated by the company |
|---|---|---|---|
| Microsoft | fiscal 2023 (July 2022) | server and network equipment from 4 to 6 years | expected roughly $3.7bn of additional operating income for the fiscal year; through the first nine months, actual operating income up $2.9bn and net income up $2.3bn |
| Alphabet | January 2023 | servers from 4 to 6 years, certain network equipment from 5 to 6 years | depreciation lower by $3.9bn, net income higher by $3.0bn, or $0.24 per share, in 2023 |
| Amazon | January 2024 | servers from 5 to 6 years | roughly $900m of operating income benefit in the first quarter of 2024 alone |
| Amazon | January 2025 | a subset of servers and networking equipment back from 6 to 5 years | approximately $700m lower operating income in 2025, plus a $920m early-retirement charge in the fourth quarter of 2024 and a further $600m drag in 2025 |
| Meta | January 2025 | most servers and network assets to 5.5 years | full-year 2025 depreciation expense lower by approximately $2.9bn |
| Microsoft | fiscal 2027 | data centers and office buildings from 15 to 25 years | described by the CFO as only a minimal benefit to operating income; reported calendar-2026 capital expenditure falls from about $190bn to about $175bn |
The last entry is the most interesting, precisely because it is absent from the public argument. Since November 2025 the entire debate has been about graphics processors and whether they last three years or six. The freshest extension is not about chips at all. It concerns the shells – and it nearly doubles their depreciation horizon at the exact moment those shells are being erected at unprecedented speed.
The Month Two Companies Read the Same Hardware in Opposite Directions
The most revealing episode in this story takes place in January 2025. Meta extended the useful life of most servers and network assets to 5.5 years, reducing that year’s depreciation expense by roughly $2.9 billion. In the same month, Amazon shortened the useful life of a subset of its servers and networking equipment from six years to five – having extended from five to six only twelve months earlier. The reversal followed the company’s own useful life study, which observed an increased pace of technology development, explicitly in artificial intelligence and machine learning.
Two companies, the same class of equipment, the same month, the same available evidence, opposite conclusions. Anyone looking for proof that useful life is a judgment with an earnings consequence rather than an engineering fact will find it here. Amazon went further, retiring hardware early and recording approximately $920 million of accelerated depreciation and related charges in the fourth quarter of 2024. That is the most honest number in the entire chronicle, because it points the wrong way.
It Is Not Fraud – and It Still Matters
In November 2025 investor Michael Burry publicly called the extension of useful lives one of the more common frauds of the modern era, and put the understated depreciation of the five largest providers for 2026 through 2028 at roughly $176 billion. By 2028, on his numbers, Oracle would overstate earnings by close to 27 percent and Meta by around 21 percent. He backed the view with money, shorting Oracle and Nebius among others.
The framing does not survive contact with the rulebook. A change in useful life is a change in accounting estimate, not an error. Under US generally accepted accounting principles it is applied prospectively, with disclosure of the effect on income from continuing operations and earnings per share in the period of change. Nothing is restated backwards, and nothing is concealed either. The disclosure requirement exists precisely because judgment is involved. Every figure in the table above comes from the companies’ own filings. A fraud whose full effect the perpetrator prints to the million in his own annual report is not an especially accomplished fraud.
Two objections do survive. First, prospective treatment is asymmetric. Concluding today that yesterday’s assumption was generous does not correct yesterday’s earnings; the depreciation not taken in the past does not disappear, it is pushed forward and served later in thicker slices. Second, the direction is conspicuous. From 2020 through 2024 the assumptions moved almost exclusively upward, during a period in which the hardware innovation cycle was accelerating rather than slowing.
The Other Side Has a Real Case
It would be convenient to stop there. In fact the defense of long useful lives is better evidenced than the outrage suggests. Nvidia finance chief Colette Kress has said that the A100 accelerators shipped six years ago are still running at full utilization today, credited to a vastly improved software stack. At CoreWeave, the 2020-vintage A100 fleet is fully booked; when a 2022 H100 contract expired, the freed capacity was immediately rebooked at 95 percent of the original price. On Microsoft’s Azure platform, P100 accelerators stayed in production from 2016 to 2023 – seven years.
Behind this sits a mechanism classical depreciation does not model: the cascade. The newest accelerator takes the largest training runs. The generation it displaces moves to inference, where latency and throughput matter more than peak performance. The generation before that handles batch work, research and smaller customers. Equipment loses its top rank, not its purpose. As long as total demand for compute grows faster than the performance embedded in each new device, the cascade holds.
There is also a composition argument that most commentary skips. Accelerators are only part of the capital budget. Land, shells, power infrastructure, cooling and network connectivity account for a substantial share and genuinely do last decades. A blanket three-year assumption for the entire spend would be as wrong as a blanket six-year one – only in the opposite direction.
The Difference Between Broken and Worthless
This is where the two camps talk past each other, because they are using different definitions of life. Physical life is how long a device runs before it fails. Economic life is how long it earns more than it costs. Competitive life is how long it can be rented at the original price. The three rarely coincide, and straight-line depreciation models only the first.
An accelerator that works perfectly after five years but commands half its original price per compute hour has lost more than half its economic value, while the books keep shaving off even slices as though nothing had happened. The reverse also holds: fully depreciated equipment can earn for years while appearing in no cost line at all, flattering margins artificially. Accounting has exactly one instrument for the first case, the impairment charge, and it fires only when an event triggers it. It is a smoke alarm, not a thermometer.
Which yields the variable actually worth watching, and it is not a financial statement item at all: the achievable price per compute hour for older hardware generations. While it holds, the six-year assumption is defensible. If it falls materially while the books stay put, the gap Burry alleges opens up for real – not as fraud, but as a delayed adjustment.
2026: The Bill Arrives
This year’s numbers explain the urgency. Alphabet spent roughly $44.9 billion in the second quarter of 2026, more than double the year-earlier figure, and reported free cash flow of negative $5.9 billion against positive $10.4 billion a year before; 2026 capital expenditure guidance was raised through the year from $175bn to $185bn up to $195bn to $205bn. Amazon reported $53.1 billion of cash capital expenditure for the same quarter and lifted full-year guidance to about $220 billion. Meta narrowed its range to $130bn to $145bn. Microsoft spent $41 billion in the June quarter while Azure grew 43 percent and crossed $100 billion in annual revenue for the first time.
Add the four companies’ own guidance together and 2026 capital spending lands in the region of $720bn to $745bn. The share of that running through this year’s income statements is a fraction. Meta’s own ratio makes it tangible: $31.1 billion of capital expenditure against $6.4 billion of depreciation in a single quarter, close to five to one. The income statement currently describes a company about one fifth the size of the one being built.
The defining property of that gap is inertia. Depreciation is the one cost that cannot be cut in a downturn, because it originates in money already spent. Freeze capital spending tomorrow and the expense would still climb for years as recently completed assets grow into a full annual charge. Meta’s slide from a 43 percent to a 31 percent operating margin is the first visible instalment of that process, not the last.
Microsoft’s Second Lever: Lease Classification
Microsoft’s summer 2026 announcement deserves separate attention, because it demonstrates that capital expenditure is itself a reported figure. CFO Amy Hood justified extending the useful life of data centers and office buildings from 15 to 25 years by reference to the company’s operating history and expected use of the assets, and described the operating income benefit in fiscal 2027 as minimal.
The larger consequence lies elsewhere. Because the longer assumed life shifts how future data center leases are classified – from finance leases toward operating leases – reported capital expenditure for calendar 2026 falls from roughly $190 billion to roughly $175 billion. Hood stated in the same breath that outside this effect, calendar-2026 investment expectations were unchanged. Not one dollar less is being spent. A different dollar figure is being reported.
Anyone tracking industry capital intensity as a time series needs to know about that break. A $15 billion drop in reported spending looks like discipline in any chart. Here it is a change of definition.
The Tax Mirror Image
US investors have a second set of books to compare against, and it points the other way. The One Big Beautiful Bill Act, signed on 4 July 2025, permanently restored 100 percent bonus depreciation for qualifying property placed in service after 19 January 2025, covering tangible property with a class life of twenty years or less – which includes servers, and many of the specialized electrical and cooling systems inside a data center. Servers otherwise fall under five-year MACRS property; the building shell is 39-year non-residential real property.
The result is a striking spread. For tax purposes an operator may deduct the full cost of a server in year one. For reporting purposes the same machine is spread across six. The difference does not vanish; it accumulates as a deferred tax liability on the balance sheet and shows up as a widening gap between the reported effective tax rate and cash taxes actually paid. That gap is one of the more reliable independent readings on how aggressive a company’s book depreciation is, and it requires no assumption of your own – it is disclosed in the tax footnote.
The practical point for a US taxable investor is narrower but worth stating: none of this changes anything about the cash the business generates. It changes which year the accounting recognizes it in. Positions taken on reported earnings multiples during a period when those earnings are being reshaped by estimate changes are positions taken on the estimate, whether or not the investor knows it.
Four Checks Anyone Can Run
First, the implied useful life. Divide gross property, plant and equipment by the year’s depreciation expense. The result is the average life the company is actually living, across all asset classes. As an absolute level it says little, because buildings and servers are mixed together. As a five-year time series it says a great deal: if it rises without an explainable shift in asset mix, an assumption has been changed.
Second, the capex-to-depreciation ratio. A reading near one describes a business in replacement mode. A reading of five, which is where Meta currently sits, describes a business whose reported profit barely reflects today’s capital deployment at all. The higher the ratio, the more future expense has already been purchased.
Third, your own sensitivity calculation. Take the server and equipment balance from the notes and recompute depreciation on an assumption one year shorter. On $100 billion of hardware, the difference between six years and five is about $3.3 billion of annual expense. Set that against net income and you know precisely how much reported profit rests on a judgment call.
Fourth, the divergence between earnings and cash. Depreciation is the single largest item driving the two apart. When net income grows while free cash flow shrinks or turns negative, the useful life question has stopped being academic. Meta’s second quarter of 2026: $15.8 billion of profit, $784 million of free cash flow.
Three Scenarios for the Coming Quarters
| Scenario | Trigger | Effect on earnings | Leading indicator |
|---|---|---|---|
| The assumption holds | the cascade works, older accelerators stay utilized and hold price | no adjustment, but depreciation still climbs sharply as the asset base grows | rental rates and utilization on older generations, secondary market prices |
| Convergence on five years | further useful life studies follow Amazon’s 2025 finding | a one-time prospective adjustment, mid single-digit percentage hit to earnings per share, no restatement of the past | useful life study language appearing in quarterly filings, individual first movers |
| Shortening plus impairments | demand for older generations breaks, capacity sits idle | double-digit earnings damage of the order Burry alleges, plus write-downs | early retirements like Amazon’s $920m charge, falling price per compute hour |
The middle scenario is both the most likely and the least dramatic. It is also the one the 2026 share price reactions already point toward: what is being punished is not the accounting, but the collision of rising depreciation with shrinking free cash flow.
What Follows From This
The extension of useful lives between 2022 and 2025 was not a bookkeeping trick. It was a series of disclosed, audited, reasoned changes in estimate. It raised the reported profits of the world’s largest technology companies by tens of billions of dollars without moving a cent of cash. And it did not eliminate the expense, it relocated it. Meta’s second quarter of 2026 shows what the relocation looks like on arrival: revenue up 28 percent, profit down 14 percent, depreciation up 46 percent.
Four questions are enough to judge any company on this dimension. How long does it depreciate, and has that assumption changed in the past five years? How much expense would a one-year shorter assumption create, measured against net income? How wide is the gap between reported profit and free cash flow, and is it widening? And finally: does the hardware still carried at a book value in the accounts still command, out in the market, the price that book value implies?
An investor who can answer those four needs neither the fraud allegation nor faith in the cascade. They have a number instead. And the recognition that the line in the notes nobody reads currently says more about where the earnings of the largest companies on earth are heading than most of the headlines written about them.