Change One Estimate, Earn Billions More: How to Spot Legal Window Dressing Yourself
2026-11-06 · @Mythogen Engine
AI in the Financial Statements, Part 2
When people hear "window dressing" in accounts, most think of fraud: fake orders, forged invoices, a bribed auditor. That happens, but rarely, because the price is too high.
The more common method is entirely legal. A company changes one accounting estimate, and the same business produces billions more in profit. The change is written in the annual report, the auditor signs off, the regulator files the announcement. Procedurally, it is flawless.
This kind of practice is common, and not confined to a few companies.
This article takes apart three public cases: Intel, Microsoft, and Taiwan's Macronix. All three changed the same number: how many years their equipment can be used. Then it shows how to find this yourself, using financial statements anyone can get.
Why now. Part 1 of this series covered the expansion of Taiwan's AI supply chain; the capital spent today becomes depreciation over the following years. Bank of America estimates that TSMC's depreciation expense could grow at a compound annual rate of 20% from 2026 to 2028 (Economic Daily News); Capital Securities estimates ASE's depreciation this year will rise 12% to 15% (China Times). The larger the depreciation, the larger the effect of the "useful life" estimate on profit.
One machine, two ways to count
A factory spends NT$10 billion on a batch of equipment. The NT$10 billion is not booked as an expense all at once; it is spread over the "estimated useful life", one slice a year. That slice is depreciation.
| Estimated useful life | Annual depreciation | Annual profit (all else equal) |
|---|---|---|
| 5 years | NT$2 billion | Baseline |
| 8 years | NT$1.25 billion | NT$750 million higher |
The machines have not changed, the products have not changed, the customers have not changed. Management simply judged that "this batch can be used for three more years", and profit rises by NT$750 million a year.
Useful life is an estimate, and accounting standards allow companies to adjust it to reflect actual use (Deloitte Taiwan). A reasonable estimate reflects reality more accurately; an over-optimistic one pushes today's costs into the future. In the current period's accounts, the two look exactly the same.
Case 1: Intel, 2023
From January 2023, Intel extended the estimated useful life of some fab production equipment from 5 years to 8. By its own estimate, the change cut that year's depreciation by US$4.2 billion, raised gross margin by about US$2.5 billion and lowered R&D expense by US$400 million (Intel 2023 annual report).
The reason the company gave: the same tools can be reused across several process generations. That reason may well hold.
But the same year had another entry. In Q1 2023, Intel offered incentives to some customers to respond to competition and accelerate their purchases; these contributed about US$900 million of revenue that quarter (Intel Q1 2023 quarterly report).
On one side, costs spread further into the future; on the other, revenue pulled from the future into today. Both are legal, and both are in the filings.
The same annual report also contains another sentence: if the company's manufacturing strategy does not succeed, the related plant and equipment may need to be impaired. Extending the useful life assumes the machines really can be used for the full eight years.
Case 2: Microsoft, fiscal 2023
In July 2022, Microsoft extended the estimated useful life of servers and network equipment from 4 years to 6, effective from fiscal 2023. The change raised operating income for that year by US$3.7 billion, net income by US$3.0 billion and earnings per share by US$0.40 (Microsoft 2023 annual report). It was not the first time: it had already extended them once in July 2020 (Microsoft 2022 annual report).
The company's reason was that software had improved how efficiently the equipment runs, together with technological progress. In the same period, Amazon and Oracle also extended the useful life of their servers (Bedrock). This may be a real change in the industry, not necessarily deliberate dressing-up.
But the same annual report contains another number. In January 2023, Microsoft announced it would cut about 10,000 jobs and took a US$1.2 billion charge for it (Microsoft 2023 annual report).
Same fiscal year: one accounting estimate added US$3.7 billion; the fate of ten thousand people cost US$1.2 billion. The first sat in one explanatory paragraph of the annual report; the second made news around the world.
Case 3: Macronix, 2016
In 2010, Macronix bought ProMOS's 12-inch fab in Hsinchu Science Park. Heavy depreciation plus price-cutting in the market left it with 18 consecutive quarterly losses from Q1 2012, nearly NT$24 billion in all (TechNews).
Then came 2016. I recalculated that period with my own financial-statement database:
| Q4 2015 | Q1 2016 | |
|---|---|---|
| Quarterly depreciation | NT$1.41 billion | NT$490 million |
| Property, plant and equipment (net) | NT$16.6 billion | NT$16.3 billion |
| Depreciation as % of revenue | 25.2% | 9.6% |
| Revenue (QoQ) | — | −9% |
| Gross margin | 8.3% | 15.5% |
Assets barely changed and revenue fell, yet depreciation dropped by two-thirds in one quarter and the gross margin doubled. Full-year depreciation was NT$5.72 billion in 2015 and about NT$2 billion in 2016.
The reason was in the news: Macronix returned to profit in Q3 2016 by adjusting the economic useful life of buildings and equipment to lower depreciation, together with cost control and growing demand (TechNews). That quarter its gross margin was 31% (Liberty Times), matching the 30.7% my database calculates.
The demand recovery was real: flash prices did rise over those two years, and Macronix's gross margin reached 43% in Q4 2017 (TechNews). But at the moment it turned profitable, the decisive entry was accounting: had depreciation stayed at about NT$1.4 billion a quarter, that quarter's net profit of NT$633 million (Liberty Times) would have been a loss of about NT$300 million (a rough estimate before tax effects). After returning to profit, Macronix announced in 2017 a NT$18.65 billion capital reduction to offset losses, bringing net asset value per share back above NT$10 (TechNews).
What the three cases share
One: all legal, all public. Intel and Microsoft wrote it in their annual reports. When a Taiwan-listed company changes useful lives, it is a "change in accounting estimate" and must be announced as material information on MOPS (for example, Chang Heng Precision's 2025 announcement). Nobody hid anything; very few people read it.
Two: the timing always falls when the numbers most need help. Intel, in the year revenue fell and it was investing heavily in new fabs; Microsoft, in the same fiscal year as its layoffs; Macronix, after its eighteenth losing quarter. Accounting estimates can be adjusted at any time, but the moment they are actually adjusted is often not random.
Three: the cost does not disappear; it moves into the future. Extending the useful life does not reduce the total cost of a batch of machines; it only spreads it thinner and longer. Depreciation not booked today still has to be booked tomorrow; and if the machines cannot in fact last the new life, an impairment must be taken all at once. Intel wrote that risk into its own annual report.
In other words, extending depreciation lives is borrowing profit from your own future. Borrowed sensibly, it is good financial management; borrowed too heavily, the bill arrives at the worst possible time.
How to check it yourself
No inside information and no paid database are needed. The quarterly reports on MOPS and some simple division are enough.
| Signal | How to calculate | What looks suspicious | Could also just be |
|---|---|---|---|
| Depreciation suddenly falls | Quarterly depreciation ÷ net property, plant and equipment | Assets unchanged, ratio drops 20–30% or more in a quarter | Construction in progress not yet depreciating during an expansion; old machines fully depreciated |
| Collecting cash more slowly | Average receivables ÷ quarterly revenue × 91 days | Days keep rising, above the company's own historical median | Customer mix changed; longer terms for big customers |
| Goods piling up | Average inventory ÷ quarterly cost of sales × 91 days | Days keep rising while revenue grows | Stocking up ahead of price rises; higher prices inflating inventory book value |
| Selling to your own people | Related-party receivables ÷ total receivables | Share suddenly jumps from near zero | Normal division of work within a group |
Every signal has a normal explanation. So after spotting one, there are four more steps:
- Quantify. What share of revenue is the difference? How many percentage points of gross margin does it move?
- Confirm. For Taiwanese companies, search MOPS material announcements for "change in accounting estimate" or "useful life"; for US companies, look for "Change in Accounting Estimate" in the annual report.
- Ask about timing. Did the change come when the company most needed good-looking numbers?
- Look up and down the chain. Put companies in the same supply chain side by side. If one company's inventory days fall while another's rise at the same time, the problem may only have changed address.
A counterexample
The first step is the one most often skipped and the most important. In Q4 2025, Nanya Technology's quarterly depreciation fell from NT$3.55 billion to NT$2.9 billion while its assets did not shrink, and my database flagged it.
But by Q2 2026, Nanya's quarterly revenue was already NT$82.5 billion, and depreciation was only 3.5% of revenue. Even if the whole NT$650 million drop came from an accounting change, its effect on gross margin would be less than 1 percentage point. Its 79.5% gross margin was pushed up by memory prices, not made by accounting. Its Q3 gross margin rose again to 85.7%, which the company attributed to average selling prices rising more than 60% quarter on quarter (Liberty Times).
For Macronix in 2016, the same signal moved the margin by about sixteen percentage points; for Nanya this time, by less than one. A signal is only a starting point; the size of the effect decides whether it counts as window dressing.
Public is not the same as seen
In the three cases, nobody lied. The numbers were right, the procedures were right, the disclosures were there.
The problem is where the disclosure sits. Stock charts, financial news headlines and earnings-call slides say "margin improvement", "return to profit", "record earnings per share". Why it improved is written in a paragraph on accounting estimates near the back of the annual report, or in one announcement among many in the material-information feed. Everyone reads the first; almost nobody reads the second.
Legal window dressing works not because it is well hidden, but because it does not need to be hidden at all. Most people look only at the result, not the formula.
Depreciation is just the easiest kind to see. Harder to see is putting the problem on someone else's books.
A company's inventory days falling may mean goods are selling well, or may only mean the goods moved into a supplier's or a downstream customer's warehouse. Looked at alone, the company's statements are clean and attractive; the problem is on someone else's books. If that someone else is not listed and has no public accounts, it is even harder to trace. Taiwan's memory industry has this shape this year: the chipmakers' inventory days plunged, while two module makers' inventory days rose to thirteen-year highs (partly because higher prices inflated the book value of inventory).
Problems like this do not go away. The chain of cause and effect always runs to the end: goods have to be sold, costs have to be fully depreciated, and the bill lands in someone's hands. Nobody knows how long it will take.
So I am not a prophet. I do not know the day the bill will arrive, only that it will. This data is not useless; very few people look at it. Not looking means not seeing, and not seeing is itself the biggest risk.
Depreciation lives are only one method, and the easiest to check, because there is an announcement and a formula. The ways of dressing up numbers are endless, and most do not leave such a clear trail.
The hardest to check is shifting the problem outward. The car industry has public data on this: from 2003 to 2005, US carmakers bore about 90% of the industry's warranty spending; they then shifted warranty costs onto parts suppliers step by step, and by mid-2012 the carmakers' share had fallen to 82.7%, while the suppliers' share had risen from about 10% to about 17% (Warranty Week). In the summer of 2019, General Motors added new terms to its purchasing contracts: if a part is later recalled for safety reasons, the supplier bears the responsibility, even if the part was built to GM's specifications (FenderBender). The consultancy Stout notes that carmakers' warranty cost-sharing programmes in practice often do not depend on who was at fault, and that the final say on how much is allocated rests with the carmaker (Stout). On paper, the big company's quality costs are low and its margins look good; the price is booked on the supplier's side — compensation, returns, rework, or simply lost orders.
This pattern can run for a long time, because suppliers can be replaced; when one cannot hold on, the next is brought in. What it leaves behind is a string of hollowed-out suppliers. The contract terms usually only come to light when something goes wrong. In 2014, when Apple's sapphire supplier GT Advanced filed for bankruptcy, court documents revealed the contract: it could not sell sapphire to other consumer-electronics customers, and it could not change equipment, specifications or processes without Apple's consent, while Apple could change the terms at any time (MacRumors); its plants closed and 890 people lost their jobs (Cult of Mac).
How much can outsiders see? Only the result, not the process. If the supplier is listed, its annual report usually names major customers that account for more than 10% of revenue; squeezed margins, longer receivable days and rising compensation costs all leave traces in the numbers. Put it next to its big customer and the outline may come together. If the supplier is not listed, outsiders will only find out on the day it goes under.
So reading financial statements is not about finding every problem; that cannot be done. What it can do is keep you from being fooled by the numbers you can see, in the places you can see them.
Data and method
Intel's and Microsoft's figures all come from the annual and quarterly reports the two companies filed with the US Securities and Exchange Commission. Macronix's and Nanya's figures come from my own financial-statement database: the raw data are quarterly reports on MOPS from 2013 to Q2 2026, with the original files kept, calculated quarter by quarter.
- Quarterly depreciation: "Depreciation expense" in the cash flow statement. Cash flow statements are year-to-date totals; single quarters are derived by subtracting adjacent quarters.
- Gross margin: gross profit (loss), net ÷ total operating revenue.
- Cross-check: the database's figures for Macronix's 2015 full-year depreciation (NT$5.72 billion) and Q3 2016 gross margin (30.7%) match outside figures of NT$5.722 billion (CMoney) and 31% (Liberty Times).
AI disclosure
The database architecture and specifications were settled in discussion between me and Claude; the code was implemented to spec by Antigravity (Gemini 3.8 Flash), with every step reviewed by me. The data analysis, outside fact-checking and writing of this article were assisted by Claude Opus 5.5; the arguments, choices and conclusions are my own.
Disclaimer
The changes in accounting estimates cited here were all disclosed publicly by the companies concerned, as required by law. This article does not accuse any company of breaking the law; it only explains how legal accounting choices affect how financial statements look. Nor is it investment advice, or a reason to buy or sell any stock.
This site has no advertising budget and no paid promotion. If this article was useful to you, one share is one vote.
AI in the Financial Statements
- One Industry Code, Two Opposite Moves
- Change One Estimate, Earn Billions More (this article)
- Where Did the Costs Go?
- The Payer Starts Collecting
- Margins Still at Record Highs, and the First Turning Signal Has Appeared