The 25-Year Number Microsoft Never Filed

Introduction

On July 29, Microsoft CFO Amy Hood told analysts on the FY2026 fourth-quarter earnings call that the company is extending the estimated useful lives of its data centers and office buildings from 15 years to 25, starting with fiscal 2027. Microsoft filed its annual report the same day. Search that 10-K for "25 years" and you get nothing; the 8-K earnings release has nothing either, and neither carries a change-in-accounting-estimate note. Meta put its version of the same decision at $2.92 billion off depreciation expense and $2.59 billion onto net income, in writing, audited. Amazon and Alphabet quantified theirs the same way.

FASB, the private board that writes American accounting rules, opened a research file on "data infrastructure investments" on April 6, and hasn't acted on it since. Meanwhile, roughly 11 cents of every dollar tracking the S&P 500 sits in Microsoft, Amazon and Meta, per Slickcharts and StockAnalysis. If you own an index fund, three of your largest holdings just revised how long they say their AI buildout lasts, and one never had to say what its revision was worth.

Why One Assumption Moves Billions

Depreciation spreads the cost of a physical asset over the years it's expected to be useful. Stretch a $6 billion server fleet from a four-year life to six and the annual charge drops from $1.5 billion to $1 billion, reported profit rises by the difference, and no cash moves. Francine McKenna, an accounting writer, told Bloomberg in coverage republished by The Business Times: "It's the number one number that they can adjust back out because it's not a cash expense."

Four companies have run this in four consecutive years. Microsoft went first in July 2022, pushing server and network equipment from four years to six; Alphabet followed in January 2023, Amazon in January 2024, Meta in January 2025. Combined quarterly depreciation at Alphabet, Microsoft and Meta went from about $10 billion in late 2023 to nearly $22 billion by September 2025, per Bloomberg reporting carried by the Economic Times.

The Filings That Show Their Work

Meta's number sits in Note 1 of its FY2025 annual report: a January 2025 assessment raised the useful lives of most servers and network assets to 5.5 years, for "a reduction in depreciation expense of $2.92 billion and an increase in net income of $2.59 billion, or $1.00 per diluted share." On the January 29, 2025 call, CFO Susan Li listed the extension alongside workload optimization and Meta's custom silicon, then added: "This will deliver savings in annual capex and resulting depreciation expense, which is already included in our guidance."

The part that gets me sits a few pages away. Meta's MD&A names its critical accounting estimates as loss contingencies, income taxes and the valuation of non-marketable equity investments. Useful lives, the assumption that added a dollar per diluted share, isn't on that list.

Amazon ran it a year earlier, bigger on the expense line. Its FY2024 10-K reports server life going from five to six years effective January 1, 2024, producing "a reduction in depreciation and amortization expense of $3.2 billion and a benefit to net income of $2.5 billion," alongside roughly $920 million of accelerated depreciation in late 2024 for servers retired early. A year later 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," for a $1.0 billion hit (FY2025 10-K), and stretched heavy equipment from 10 to 13 years the same day.

In the same year, on the same hardware class, Meta told investors AI and non-AI servers alike could run longer while Amazon told investors AI's pace had shortened part of its own fleet's life, and both companies had auditors sign off.

Alphabet's is the biggest number of the group and the one almost nobody covered: a January 2023 assessment moved servers and certain network equipment to six years, for "a reduction in depreciation expense of $3.9 billion," per its FY2023 10-K, two years ahead of Meta.

Microsoft's own 2022 change left the cleanest receipt of any of them. In the 10-Q for that quarter, Microsoft Cloud gross margin "increased 2 points to 73%," followed immediately by: "Excluding the impact of the change in accounting estimate, Microsoft Cloud gross margin percentage decreased 1 point."

Depreciation isn't the only lever in these documents, and it isn't the only one that draws scrutiny when it's disclosed. Note 5 of Meta's same 10-K covers a Louisiana data center joint venture Meta holds at 20% and doesn't consolidate, with "maximum exposure to loss... $45.95 billion as of December 31, 2025." Ernst & Young flagged that consolidation judgment as a critical audit matter, the formal scrutiny Microsoft's unfiled buildings change never had a chance to draw.

Zero Hits in Microsoft's 10-K and 8-K

Here's where the July 2026 change stops fitting the pattern. Microsoft isn't the aggressor in it. Its pre-change building range was five to 15 years; Meta carries buildings at 25 to 30, Amazon at the lesser of 40 years or the building's remaining life, and Alphabet's FY2025 10-K at "seven to 40 years." Microsoft's new 25-year figure lands at or below all three, and Hood said the change "affects only the timing of future depreciation and is expected to have a minimal benefit to FY27 operating income."

The absence still stands. Because the change takes effect in fiscal 2027, Microsoft owed no quantification in the fiscal 2026 documents and filed none: not in the FY2026 10-K (accession 0001193125-26-323660), not in the 8-K earnings release (0001193125-26-323632). A decision to add a decade to the assumed life of every Microsoft data center exists in public as one spoken sentence and a line on an outlook slide.

What Hood did quantify was capital expenditure. She said the "greater impact is on capital expenditures as more of our future data center leases will shift from finance leases to operating leases," and moved calendar-2026 capex guidance to approximately $175 billion, from roughly $190 billion in April. Guided capex for the September quarter is above $50 billion, against $41 billion in the quarter just reported. The headline guidance number fell while the spending behind it rose, which is my reading and not hers. Roughly two-thirds of that quarter's capex went to short-lived CPUs and GPUs anyway, untouched by the buildings change.

Who Benefits

Start with the people whose pay tracks the stock. Satya Nadella's fiscal 2025 compensation was $96.5 million, per Microsoft's proxy filed October 21, 2025, with more than 95% of his target compensation opportunity performance-based, against a median employee at $200,972. Those awards key off operating income and shareholder return, so anything that softens reported operating expense feeds them. No filing connects a specific depreciation change to a specific payout, and I'm not going to pretend one does.

The bigger beneficiary is the buildout. Depreciation is where an AI capex boom finally shows up as a cost, and Bloomberg reported in November 2025 that the four biggest spenders were on track to raise combined capex about 40% over the following year, to $460 billion. Slowing that climb by revising an estimate buys room to keep spending without forcing the payback question out loud.

Microsoft's stock rose 8.13% after hours on July 29; Meta, which took the largest quantified net income benefit of the four, fell about 9% the same night, with free cash flow down 91%.

Michael Burry's $176 Billion, and the Estimates Below It

Michael Burry made the loudest version of this argument on X on November 10, 2025, in a post reproduced by Sherwood News: "Understating depreciation by extending useful life of assets artificially boosts earnings -one of the more common frauds of the modern era." He puts industry-wide understated depreciation at $176 billion between 2026 and 2028, with reported profits overstated 20.8% at Meta and 26.9% at Oracle by 2028.

Those are projections about a future year, not claims about current earnings, and his math sits behind a paywall.

Everyone else who has run the numbers lands nowhere near him. The Economist estimated in September 2025 that a three-year schedule would cut hyperscaler pre-tax profit by roughly 8%, about $26 billion a year (via Fortune). J.P. Morgan modeled a 6-8% hit to EPS and operating margin on gear added since 2022. Goldman Sachs Global Institute read A100 and H100 rental pricing as evidence that five-to-six-plus-year lives are defensible in April 2026, then hedged that this might just reflect today's "extreme capacity constraints."

The rental-price defense has a hole. J.P. Morgan's own research put GPU hourly rental rates down 20-25% over the year to late 2025, and Princeton's Center for Information Technology Policy put H100 rentals down 70% from peak in December 2025. Nvidia and CoreWeave publicly defend four-to-six-year schedules, and UNC accounting professor Stephen Glaeser told Bloomberg the question is "difficult to judge." The four estimates don't measure the same quantity, and the people holding the data that would settle it are the ones whose reported profit depends on the answer.

FASB Opened a File and Left It There

On April 6, 2026, FASB chair Rich Jones added a research-agenda project directing staff to analyze "data infrastructure investments and non-traditional lending," per Accounting Today. CFO Brew's coverage ten days later named the depreciation period for servers and networking equipment as one potential focus area. That is the entire institutional response: a private standard-setter under SEC oversight, not an enforcement body, opening staff work that may or may not become a proposal, on no timeline. Three and a half months later Microsoft added a decade to its building lives, prompted by nothing FASB had found.

What deserves attention is a timing rule. Under ASC 250, a company owes investors the dollar effect of an estimate change for the period in which that change hits results. Microsoft's hits fiscal 2027, so it can announce the direction of a change in July, begin booking it in July, and put no number in a filed document until the next reporting cycle. Investors get the sentence now and the arithmetic later, on the company's calendar. That's the rule working exactly as written, and it doesn't take a loophole to do it.

The Bottom Line

Microsoft's number is coming. Fiscal 2027 started July 1, and its first quarter closes September 30. The last time Microsoft did this, in 2022, the figures landed in exactly that filing: $1.1 billion of operating income in one quarter, with a $3.7 billion estimate for the year. Whatever the buildings change is worth will surface the same way, a quarter after anyone was asking.

FASB's file is still open. Three of these four companies published the math on their own, and Microsoft demonstrated nobody has to. Whether that file becomes a proposal decides whether the next company to revise this assumption says what it's worth when it announces the change, or a fiscal year later, the way Microsoft gets to.