Is Big Tech Overstating Earnings? The Depreciation Math

Microsoft earned $17.95 a share last year. The bears say the real number is lower. Both sides are arguing about the wrong thing — and this is the arithmetic, the history, and the way out.

By David Berkowitz · Berk on Value · July 30, 2026

PART ONE — What Microsoft actually reported

The quarter, and the year it finished

Microsoft closed its fiscal year on June 30, 2026, and reported on July 29. The June quarter brought in $90.0 billion of revenue, up 18%, with operating income of $40.6 billion and net income of $35.8 billion. Official earnings per share came in at $4.81, up 32%. The stock rose roughly 8% to 9%.

For the full year: revenue of $331.8 billion, operating income of $155.2 billion, and net income of $133.7 billion. Diluted earnings per share of $17.95. Microsoft Cloud passed $214 billion and Azure passed $100 billion.

Two numbers on the call are worth keeping straight. Chief financial officer Amy Hood said earnings per share were $4.74. The press release says $4.81. Both are correct. The $4.81 is the official figure under the accounting rules. The $4.74 is the adjusted figure, after backing out the effect of Microsoft’s investment in OpenAI, and it is what analysts had been forecasting against.

The four-quarter arc

One quarter tells you almost nothing. Four quarters side by side show what actually changed.

Figures from Microsoft’s releases for the first, second, third and fourth quarters.

What got better

Azure reaccelerated when it was supposed to slow down. At a $100 billion run rate, the law of large numbers says growth decays. Instead, it went 40%, 39%, 40%, 43%, and management guided the next quarter to about 45%. Microsoft beat its own Azure forecast three quarters in a row.

The pricing model changed and started producing revenue. Microsoft moved Microsoft 365 and GitHub from charging per seat to charging per seat plus usage. GitHub Copilot revenue accelerated more than 60% from the prior quarter after the June billing change. Paid Copilot seats went 15 million, then 20 million, then past 30 million, with net additions more than doubling in the June quarter. Revenue per customer stops being capped by headcount.

The contracted backlog reached $678 billion, up 84%. That is business customers have signed for, and Microsoft has not delivered yet.

What got worse

Margins fell in every single quarter. Company gross margin went 69, 68, 68, 67. Operating margin went 49, 47, 46, 45. Cloud gross margin fell almost exactly one point a quarter, from 68 to 65. Management now guides fiscal 2027 operating margins down again, which is the first guided decline of this cycle.

Free cash flow ran $25.7 billion, then $5.9 billion, then $15.8 billion, then $19.6 billion — while cash from operations grew every quarter. Capital spending grew faster. The June quarter alone absorbed $41.0 billion, and the next quarter is guided above $50 billion.

Where the profit actually came from

This is the part of the year that gets the least attention and deserves the most.

Microsoft’s official net income grew 31% for the fiscal year. Its adjusted net income grew 22%. The nine-point difference is not the software business. It is the change in value of Microsoft’s stakes in two private artificial intelligence companies.

Microsoft owns about 27% of OpenAI and a stake in Anthropic. When the estimated value of those holdings rises, the increase runs through the income statement as profit. No shares were sold. No cash arrived. Accountants call this an unrealized gain, and the rules require reporting it.

Quarter by quarter, those stakes threw the reported number around hard. The September quarter took a $3.1 billion loss from OpenAI. The December quarter took a $7.6 billion gain, which is why earnings per share printed $5.16 that quarter against $4.27 the next. The March quarter was almost nothing. The June quarter carried a $3.2 billion gain on Anthropic, part of 27 cents a share of one-time items that landed above what management had forecast in April.

Read that last one carefully. Microsoft beat the earnings expectation by about 50 cents a share in the June quarter, and 27 cents of the beat came from items that were not the software business. Roughly half. The headline said beat. The operating business delivered about half of what the headline implied.

Microsoft is not alone in this. Alphabet reported second-quarter 2026 net income of $112.1 billion, a headline that read as earnings up 294%. Strip out roughly $99 billion of paper gains on its stakes in Anthropic and SpaceX, and per-share earnings came in around $2.85 against a $2.88 estimate, with the core business growing an ordinary 30%. Same mechanism, larger number, same quarter.

This is legal and disclosed. The Securities and Exchange Commission requires companies to show the official figure just as prominently as the adjusted one, and Microsoft does. But an investor who reads only the 31% growth number is reading the value of a private stake, not the performance of a business.

One footnote on the accounting

Microsoft also announced on the call that it is extending the assumed life of its data center and office buildings from 15 to 25 years, and moving more future leases into a category that sits outside reported capital spending. Both changes shrink reported costs without changing cash. Neither is large enough to argue about — 25 years is ordinary for a building, and the IRS uses 39 years for tax purposes. Mentioned here so the record is complete. The number that matters comes next.

PART TWO — The bears’ argument, and what it would cost

The fight

In November 2025, the argument about depreciation stopped being technical and started being personal.

Michael Burry, who made his name on the housing bet in 2008, filed his quarterly holdings and disclosed put options on roughly 5 million Palantir shares and 1 million Nvidia shares — bets that both would fall. He argued that the large technology companies were stretching the assumed lives of hardware that goes obsolete in two or three years, and put the resulting understatement at $176 billion across 2026 to 2028, with Oracle overstating 2028 earnings by 26.9% and Meta by 20.8%. He called the practice “one of the more common frauds of the modern era.”

A note on how that trade was reported. The filing showed $912 million and $187 million of underlying stock. Those numbers went around the world as a billion-dollar bet. They were not. A quarterly holdings form reports the market value of the shares the options are tied to, not the money paid for the options. The amount actually at risk was a fraction of the headline. That distinction did not travel.

Days later, Palantir’s chief executive Alex Karp went on CNBC and used a word the network had to bleep to describe the short position, called the broader bet “egregious,” and promised he would be “dancing around when it’s proven wrong.”

Jim Chanos, the short seller who called Enron, put the technical case more plainly: the chips “basically depreciate over two years,” and some data center businesses are “equipment leasing dressed up as growth.”

Is Burry right? The short answer, before the arithmetic

Most people arriving at this argument want a verdict on a person, not a lesson in accounting. So here it is up front, in three parts.

On the direction, he is right. Reported profits at the large technology companies are flattered by depreciation assumptions; the gap between cash going out and cost recognized is real and large, and it is getting wider.

On the size, the evidence cuts both ways and the honest answer is unknown. More on that below.

On the word fraud, he has moved. In a July 8, 2026 note, he reframed the whole thing: “Depreciation is not a bet on when a chip stops working. A chip can rent and still depreciate very fast economically.” He now calls depreciation schedules “an economic lever, not a physical measurement.”

Read that reframe carefully, because it is the most important sentence anyone has written on this topic in a year. An economic lever, not a physical measurement. The loudest bear in the market has stopped arguing that the companies picked the wrong number and started arguing that the number is a lever in the first place. That is not a softer version of his case. It is a different and better one — and it is the argument this post ends on, from a completely different direction.

He also volunteered a fact that cuts against him, which few people in this argument do. CoreWeave has re-leased H100 chips at 95% of the original price at the end of contracts running five to seven years. He reads it as proof that the frontier moves faster than the schedules. It reads equally well as proof that the hardware holds its earning power.

When this actually was fraud

The word keeps getting used, so it is worth knowing what it looked like the last two times a regulator agreed.

Waste Management, 1992 to 1997

Dean Buntrock’s garbage trucks kept getting older. On the books, they kept getting more valuable. Buntrock was the founder, chairman, and chief executive. Phillip Rooney was president, James Koenig the chief financial officer, Thomas Hau the controller. Over five years, the six men running the company avoided depreciation expense on the fleet by assigning unsupported salvage values and stretching out useful lives, and went further by assigning arbitrary salvage values to assets that had never carried one.

The mechanism is the part worth studying. Each year they made what the regulator called top-level adjustments to bring reported results in line with earnings targets set in advance. Because the adjustment closed a gap rather than fixing the business, the next year needed a larger one. The company restated in February 1998 and misstated pre-tax earnings by roughly $1.7 billion — the largest restatement in corporate history at that point. The stock fell more than 33%. Shareholders lost over $6 billion.

Hertz, 2013 to 2015

In November 2013, chief executive Mark Frissora approved reaffirming Hertz’s earnings guidance. His own people had already run the numbers and come up short. He pressed his staff to find money by re-analyzing reserve accounts, and directed the company to hold rental cars in the fleet for longer periods — which lowered depreciation expense, and which Hertz did not properly disclose.

The company revised results in 2014 and restated in July 2015, cutting previously reported pretax income by $235 million. Hertz settled with the regulator for $16 million in December 2018. Frissora settled in 2020, repaying $1,982,654 in bonus and incentive compensation and paying a $200,000 civil penalty. Compare Hertz to hyperscalers.

Nobody has accused Microsoft, Alphabet, Amazon or Meta of anything resembling this, and the differences are large: those changes were disclosed, quantified, audited, and applied prospectively. But notice what Hertz proves. Holding an asset longer on paper is not an abstraction. It moved a chief executive’s pay, and the regulator made him give it back. That is the bridge between “it is only an estimate” and “estimates have consequences.”

The memo that answered a tweet

Then something unusual happened. In late November 2025, Nvidia sent Wall Street analysts a private memo. The first source document it set out to refute was a Twitter account. The account belonged to Michael Burry.

The memo made a physical argument rather than an accounting one. It said customers depreciate graphics chips over four to six years based on how the machines are actually used, and that A100 chips released in 2020 were still running at high rates with real economic value left in them. It corrected a figure Burry had cited, putting Nvidia’s buybacks since 2018 at $91 billion rather than $112.5 billion — his number, the memo said, appeared to include taxes on restricted stock. Burry called the reply disappointing, said it was a straw man, and said the right historical comparison was Cisco, not Enron.

Months later, Nvidia’s chief financial officer Colette Kress said it out loud on an analyst call. The A100 chips shipped six years earlier were still running at full utilization.

That sentence is the strongest evidence in the entire debate, and it is not an accounting claim. It is a claim about machines. If six-year-old chips really are running flat out, the four-year assumption was wrong and the companies extending lives are correcting an error rather than manufacturing profit. If utilization is measured generously, the bear case survives intact. Everything else is downstream of that one question, and outside investors cannot check it.

How big the gap is across the industry

Before narrowing to Microsoft, the scale. Lance Roberts published the most widely syndicated version of the bear case on July 24, 2026. His figures: the five largest cloud companies plan roughly $760 billion of capital spending in 2026 against only $211 billion of recognized depreciation — about $549 billion of cost sitting on balance sheets rather than running through income statements. He projects their combined free cash flow falling 91% to $16 billion this year while their combined net income rises 25% to $506 billion.

Two of those companies reported within a day of each other. Microsoft’s June quarter produced $41.0 billion of capital spending against $19.6 billion of free cash flow.

Meta reported the same week: capital spending of $31.08 billion against free cash flow of $784 million, with depreciation and amortization of $6.36 billion. Its 2026 spending guidance is $130 to $145 billion. Meta disclosed no change to any useful life.

Those two numbers belong side by side. Meta spent $31.08 billion on property in three months and finished the quarter with $784 million of free cash flow. Whatever anyone concludes about depreciation schedules, the cash statement is not an estimate. It already shows what the income statement has not caught up to.

Roberts is careful about the word, and so should everyone else be. He calls it “a depreciation timing phenomenon, not fraud.” That is the correct framing, and it is now shared by Roberts, Usvyatsky, and — as of July — Burry himself.

What a shorter life would cost Microsoft?

Nobody in this argument has run the numbers on Microsoft specifically. Here it is, with the method shown and the limits stated plainly.

Microsoft’s annual report breaks its property into categories. Servers, network equipment, and software stood at $215.9 billion at cost on June 30, 2026, up from $132.8 billion a year earlier — a 63% increase in twelve months. The stated policy is two to six years. Total depreciation for the year was $34.3 billion across every category.

Because the pool nearly doubled during the year, the honest basis for restating fiscal 2026 is the average balance, roughly $174.4 billion, not the year-end figure. Here is what happens if you change nothing except the assumed life, using a 19.4% tax rate and 7,453 million diluted shares — Microsoft’s own figures for the year.

At a three-year life, about $3.14 of Microsoft’s $17.95 goes away — roughly 17.5% of reported earnings per share, and about 19% of operating income. Set that against Burry’s own claims of 26.9% for Oracle and 20.8% for Meta and the method lands in the same neighborhood, which is a reasonable sign the arithmetic is not off by an order of magnitude.

The forward number is larger. The server pool ended the year 63% bigger than it started, so a fiscal 2027 run rate on the year-end balance produces about $3.89 a share at a three-year life rather than $3.14.

Five reasons to hold this loosely

• Microsoft does not publish accumulated depreciation by category. Only one combined figure of $118.7 billion appears. Nobody outside the company can compute the remaining book value of the server pool exactly.
• The policy is two to six years, not six. Some of the fleet already has shorter lives, so the starting point is below six, and the gap is smaller than the table implies.
• A real change would apply to future years only. Companies do not restate prior years for this. So this is a sensitivity, not a restatement.
• Assets placed in service partway through the year earn only part of a year’s charge. The average-balance approach approximates that; it does not replicate it.
• A six-year charge on the average pool comes to $29.1 billion against $34.3 billion of actual total depreciation across all categories. That is roughly 85%, which is plausible for a company whose spending is overwhelmingly servers — but it is a cross-check, not a proof.

The other side, which is stronger than the headlines suggest

Olga Usvyatsky is the accounting specialist worth reading here. Her work on useful lives for graphics chips takes apart the claim that extending a life is fraud. Under the rules, this is a change in estimate — a category the standard setters built on purpose, because estimates are supposed to be revised as experience accumulates. She has also written a technical version for legal readers.

Ed Yardeni comes down on the company side and says so directly: he sides with the hyperscalers rather than Michael Burry.

Amazon extended three times, then reversed. In the fourth quarter of 2022 it moved servers from four years to five and networking gear from five to six, cutting 2022 depreciation by about $3.6 billion. A year later it pushed servers out to six years, adding an expected $3.1 billion to 2024 operating income. Then in the first quarter of 2025 it shortened a subset of servers back to five years, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” That cost roughly $700 million of 2025 operating income, on top of $920 million of accelerated depreciation the prior quarter on equipment retired early.

Four changes in three years, in both directions, at one company, on the same class of machine. If stretching lives were free money, nobody would ever shorten one. Amazon did, and it is the same company that had stretched them twice before.

Microsoft has been here before and showed the number. In July 2022 Hood said “we are extending the depreciable useful life for server and network equipment assets in our cloud infrastructure from four to six years,” and the company disclosed in its annual report that the change would add $3.7 billion to the following year’s operating income. $1.1 billion landed in the first quarter alone. Analysts calculated it lifted quarterly earnings per share to $2.35 from $2.23.

Alphabet did the same and published it to the dollar. Moving servers to six years in January 2023, its first quarter showed depreciation down $988 million, net income up $770 million, and earnings per share up six cents.

Todd Castagno of Morgan Stanley makes a different point that gets lost: finance leases hide how capital-hungry these businesses really are, which means the reported spending figure has its own estimate problem sitting underneath it.

What the hardware evidence actually shows

The argument keeps getting fought with adjectives. There is real data, and it does not point one way.

For the companies: an H100 chip still fetches 60% to 83% of its value on the secondary market after 18 months. IBM Cloud and Google Cloud still rent 2016-era Tesla P100 chips — eight-year-old silicon. Nine-year-old M4000 chips run at near-total capacity and still earn revenue. Google reports seven- and eight-year-old tensor processing units at 100% utilization. And power limits now extend the value of older parts, because newer 700-watt chips exceed what many buildings can physically supply.

Against them: training Meta’s Llama 3 model saw 30.1% of all disruptions caused by chip failures over 54 days — roughly a 9% annualized failure rate. Usvyatsky notes H100 systems trade below half of their new price by year three. Both facts are about the same generation of hardware.

The people who actually rent the hardware out went on the record too. Nebius chief revenue officer Marc Boroditsky said older Hopper chips were selling immediately, often at better pricing than they had previously carried. CoreWeave chief executive Michael Intrator said an expiring H100 contract was recontracted within 5% of the original agreement. Their position is simple: if the machines keep earning, a six-year life is not a stretch; it is a measurement.

Chanos aims at those same companies from the other end. On CoreWeave, he sets $3.4 billion of annualized adjusted operating earnings against $1.2 billion of annual interest, under a ten-year depreciation schedule for the chips. That is not a depreciation argument. It is a question about whether the returns cover the cost of the money, which is where this post ends up.

One more comparison that rarely gets made. The specialist cloud companies, which have no incentive to flatter anyone else’s numbers, sit on both sides: Lambda Labs uses five years, Nebius four, CoreWeave six. The large companies at six years are at the top of that range, not outside it.

Read across all of it and the shape is clear. Nobody serious claims the accounting is illegal. The argument is whether the assumption matches the physics. And the answer is genuinely unknown, because the only people who can measure utilization are the people whose earnings depend on the answer.

PART THREE — Why none of it matters

Look at what both sides agree on

Burry says life is too short to justify six years. Nvidia says six years is about right. Microsoft says two to six. Amazon says five. Alphabet says six.

Every one of those positions accepts the same frame. There is a correct useful life, and somebody picked the wrong one. Win that argument, and you get a different earnings number for the same company, the same machines, and the same cash.

Look at the record inside thirty months. Microsoft revised upward in 2022 and added $3.7 billion to operating income. Alphabet revised upward in 2023 and added six cents a share in one quarter. Amazon revised downward for 2025 and gave back about $700 million. Three of the largest companies on earth, the same class of machine, three different answers — and every one of those filings was audited, legal, and disclosed.

If a number can move by billions because a committee revised an assumption, that number is not a measurement. It is a policy choice. The useful question is not which choice is right. It is why anyone builds a valuation on top of a figure that management sets and can reset.

Burry got there himself, from the other end. His July 2026 phrase — depreciation is “an economic lever, not a physical measurement” — is the same conclusion reached by a short seller working forward from the hardware. Jim Chanos arrives at a version of it too, when he stops talking about schedules and starts talking about “mid-to-low single-digit pretax returns on capital” at the specialist cloud companies. That is a capital-return argument, not a depreciation argument.

Three people with nothing else in common — a short seller, a forensic accountant, and a portfolio manager — keep sliding off the schedule and onto the capital. There is a reason for that, and it was written down fifty-two years ago.

The proof is in the disagreement

Here is the cleanest evidence that depreciation is a choice rather than a measurement, and it comes from the analysts who model these companies for a living. Consensus estimates for Meta’s 2028 depreciation and amortization carry a 24% standard deviation. For revenue, the figure is 4%.

Sit with that gap. The people paid to forecast this business agree closely on how much money customers will hand it two years from now — the hard part, the part that depends on competitors and pricing and demand. They cannot agree within a quarter of a trillion dollars on what it will cost to own the equipment it already bought.

Revenue is a fact that has not happened yet. Depreciation is an assumption about equipment already sitting in buildings that already exist. One of those should be easier to forecast than the other, and it is not the one you would expect.

And nobody official has weighed in

One more thing worth stating, because it is a verified absence rather than an oversight. There is no Securities and Exchange Commission comment letter on useful lives for artificial intelligence hardware. No Public Company Accounting Oversight Board inspection finding. No Financial Accounting Standards Board project. No on-point academic literature comparing economic life to book life for data center equipment.

The single largest capital program in corporate history rests on an assumption that no regulator has questioned in writing and no researcher has independently measured. That is not an accusation. It is the state of the evidence, and it is why this argument keeps getting settled by whoever is loudest rather than by anyone with data.

The strongest objection to everything above

A large part of the market reads all of this and shrugs. Depreciation is a non-cash charge. Change the schedule and you change a bookkeeping entry, not a dollar. Here is that case, made properly, because it is the best argument against spending ten pages on the subject — and it is mostly right.

The mechanic: in a cash flow model, it cancels out

This is arithmetic, not opinion. Build a discounted cash flow model and you start with net income, add depreciation back because no cash left the building, then subtract what the company actually spent on equipment.

Stretch the useful life and depreciation falls. The add-back falls by the identical amount. Capital spending is whatever the company actually spent, which no assumption can change. Free cash flow comes out the same to the dollar. Aswath Damodaran builds free cash flow exactly this way, and his line on Meta’s accounting applies here as well: the company spent the money, no matter how you categorize it.

The tax point: the cash tax bill does not move either

This one closes the last door, and most people arguing about this do not know it. In the United States, book depreciation and tax depreciation are two separate calculations. Companies depreciate for the tax authorities under statutory schedules, not under the useful life printed in the annual report. Stretching a book life does not shrink the deduction on the tax return, so it does not raise cash taxes paid.

So the one channel through which a depreciation assumption normally touches real money is shut. That is why the critics call the whole argument a distraction, and on their own terms they are right.

Management says the same thing, in the same words. Amy Hood, describing Microsoft’s July change: it affects only the timing of future depreciation.

Four places the objection breaks, and all four involve cash

• Pay. Compensation tied to earnings per share converts an accounting estimate into a payroll check. That is not theoretical. Mark Frissora repaid $1,982,654 of bonus and incentive compensation, and a company does not claw back money over a bookkeeping entry that made no difference.
• Price. Most investors never build a discounted cash flow model. They pay a multiple of reported earnings. The multiple lands on whichever number the schedule produced. Value and price are different things, and the schedule reaches one of them.
• Forecast. The useful life is management’s own estimate of how long the asset earns. Accept six-year servers, and if the machines get replaced in four, replacement spending arrives two years early. The cash flow model was wrong from the first cell — not because of the add-back, but because the timing assumption underneath it came from the same place as the depreciation schedule.
• Disclosure. Waste Management shareholders lost more than $6 billion when the stock fell 33% on the restatement. Whatever that is, it is not what “makes no difference” looks like.

The critics and the storytellers are answering two different questions and both are answering correctly. The critics ask: does this change what the business is worth? No. The stories ask: does this change what people pay, what managers earn, and what you find out later? Yes, all three. Keeping those questions apart is most of the clarity available on this subject.

Which leaves a practical problem. If the schedule cannot be trusted and cannot be verified from outside, and if it still moves the price and the pay, then an investor needs a way of measuring the business that the schedule cannot reach at all. That method exists. It is older than the argument.

This objection is fifty-two years old

1974: the man who said earnings per share do not count

Joel Stern was president of Chase Financial Policy, the corporate advisory arm of Chase Manhattan Bank, where he spent 18 years starting in 1964. In July 1974 he published an article in the Financial Analysts Journal and named it after the thing he wanted the profession to quit: “Earnings per Share Don’t Count”.

His argument was simple. Earnings per share survives because it is easy to compute, not because it explains prices. He set two companies side by side, both growing earnings 15% a year. The one that needed less capital to produce that growth was worth more. Investors, he wrote, discount earnings net of the capital required to produce them. In 1982 he left Chase and founded Stern Stewart & Co. with G. Bennett Stewart III. That firm turned the idea into EVA — economic value added.

The same decade: the man who attacked the other half

Bartley Madden co-founded Callard, Madden & Associates in Chicago in the early 1970s. Inflation had exposed a different problem. Plant sat on the balance sheet in old dollars while the cash coming out of it arrived in new ones. His answer was CFROI: charge the full gross asset base, capitalize intangibles, and restate everything in current dollars. HOLT Value Associates built it into a global database and Credit Suisse bought HOLT in 2002. Madden was still publishing on the same gap in 2025.

Stern attacked the numerator. Madden attacked the denominator. Same decade, no shared payroll, same verdict. And note what Madden reached for as the fix: gross assets, undepreciated. Hold that thought.

2019: the correction, written by the people who installed the original

Greg Milano was a partner at Stern Stewart before founding Fortuna Advisors. Installing EVA inside client companies, he kept watching managers do the same thing: hold the old machine one more year, because buying the new one would dent the number.

The math explains the behavior. Under EVA, the cost of owning an asset is depreciation plus a capital charge on what is left of its book value. That cost is heaviest in year one. It gets lighter every year after. It goes to zero once the asset is fully written off. Managers responded exactly as the formula was written. Milano named the habit sweating assets and found it carried a negative relationship with total shareholder return.

His update, published as “Beyond EVA” in 2019, is called Residual Cash Earnings. It adds depreciation back and charges a flat rate on gross assets. Owning something then costs the same in year seven as in year one. From 2012 to 2018, Amazon’s Residual Cash Earnings improved by more than $38 billion. Its EVA improved by less than $11 billion. Across 20 non-financial industries from 1999 through 2018, the newer measure tracked shareholder returns better than EVA in all 20.

Madden reached for undepreciated gross assets in the early 1970s. Milano reached for the same thing in 2019, from a completely different direction, after watching what managers actually did. When independent people keep arriving at the same fix, that is evidence rather than preference.

2017: proof that the schedule steers the company

In October 2017, Varian Medical Systems stopped paying its managers on earnings per share and stopped telling Wall Street what to expect each quarter. The replacement metric charged capital on undepreciated gross property, plant and equipment. It also capitalized research and development on an eight-year window. That turned a research dollar into an asset a manager had to earn a return on, rather than a cost to cut in December.

The chief financial officer at the time, Gary Bischoping, put it plainly: “This removes any incentive to cut R&D to meet a short-term goal, so it promotes investing in innovation.” The company bought Sirtex Medical for $1.3 billion in early 2018. Over the 18 months to March 2019, Varian returned 41.6% to shareholders against 15.7% for the S&P 500, as reported by the advisers who designed the plan.

This is the story that answers the reader thinking an accounting debate does not matter. Same company, same machines, same engineers. Change the capital charge and the research treatment, and the capital decisions change inside a year. The depreciation schedule does not merely describe a business. It steers one.

The way out

Go back to the table in Part Two. Every row is the same company. Same revenue, same customers, same servers, same electricity bill, same cash in the bank on June 30. The only thing that changes across those rows is a number a committee picked.

Run those same figures through an economic profit lens and almost nothing happens. The capital charge sits on what Microsoft actually spent on the fleet — $215.9 billion, a figure no assumption can move. The return is measured against the cash the business actually produced. Neither of those numbers cares what anyone decides about useful lives.

This is why the ValueAligned Portfolio is not valued on reported earnings. A capital charge is applied to what a company actually spent. The return is measured against the cash the business actually produces. Neither of those numbers moves when a management team revisits a depreciation table, which is the entire point of building the work that way.

The useful question about a graphics chip is not whether it lasts three years or six. It is whether the cash that chip produces clears the cost of the capital that bought it. Ask it that way and the Burry argument becomes interesting rather than decisive — because if the chips genuinely wear out in three years, the problem is not the accounting. The problem is that the returns were never there, and no depreciation schedule was ever going to hide that for long.

An investor who never let the schedule into the valuation has nothing to reprice while everyone else argues about the assumption.

What to watch

• Does Azure actually grow about 45%, the way management promised for the September quarter? That is the number the business case rests on, and it is checkable.
• Does the gap between official and adjusted earnings widen again? Nine points of the fiscal 2026 growth came from marking up private stakes. If that becomes a permanent feature, the reported growth rate stops describing the business.
• Does anyone shorten a server life in the next twelve months? Amazon already did. One more would turn the bear case from an argument into a trend.
• Does free cash flow recover as capital spending crosses $50 billion a quarter? Cash cannot be assumed into existence. It is the one line in this whole story that no estimate touches.

Big numbers are easy to report and hard to read. The work is in knowing which ones came from customers and which ones came from a spreadsheet.