> ## Content Index
> Fetch the complete content index at: https://www.edgewisely.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Footnote Quietly Inflating AI Earnings
- URL: https://www.edgewisely.com/the-footnote-quietly-inflating-ai-earnings/
- Published: 2026-09-07T08:34:28.000Z
- Updated: 2026-09-07T08:35:52.000Z
- Description: How a dispute over how long a GPU stays useful could be worth $176 billion in overstated hyperscaler profit.
- Author: John Karpentar
- Tags: Finance, Opinion

**Depreciation policy, not demand, may be the softest number in AI's biggest balance sheets.**

*This is an analysis piece: the numbers below are drawn from public filings, disclosed accounting policies, and named research, but the interpretation and conclusions are Edgewisely's own.*

Every hyperscaler's earnings report now contains a number almost nobody outside an accounting department reads closely: the assumed useful life of its servers. It sits in a footnote, expressed in years, and it quietly decides how much profit the company reports each quarter. In 2026, that footnote became one of the most contested numbers in tech.

## The number that started the argument

Michael Burry, the investor who famously bet against subprime mortgages before 2008, spent much of the second half of 2026 arguing that hyperscalers are using overly generous depreciation schedules to inflate reported earnings. His claim, as reported by [CNBC](https://www.cnbc.com/2025/11/11/big-short-investor-michael-burry-accuses-ai-hyperscalers-of-artificially-boosting-earnings.html?ref=edgewisely.com): major cloud providers depreciate Nvidia-based data-center hardware over five to six years, even though Nvidia's roughly 12-to-18-month architecture cycle means the real economic life of that hardware is closer to two or three years. Burry estimated the cumulative overstatement of profits across the industry could exceed $176 billion between 2026 and 2028, and singled out Oracle and Meta specifically, projecting their profits could be inflated by roughly 27% and 21% respectively by 2028, according to coverage of his analysis picked up by [Yahoo Finance](https://finance.yahoo.com/technology/ai/articles/michael-burry-sounds-alarm-again-144154784.html?ref=edgewisely.com).

Depreciation isn't a cash expense — it's an accounting estimate of how fast an asset loses value. Stretch that estimate out, and quarterly profit looks bigger than the underlying cash economics justify. Shrink it, and profit shrinks with it, even though nothing about the actual business changed. That's the mechanism Burry is pointing at, and it's one every AI infrastructure investor should understand before trusting a hyperscaler's reported margins at face value.

## Why the useful-life assumption isn't fixed

Between 2020 and 2024, most large tech companies steadily extended the useful lives they assigned to servers and networking gear — a trend that diverged sharply in 2025, when Amazon shortened the useful life of a subset of its servers while Meta extended its own estimate further, according to industry analysis tracking the trend. There is no single correct answer here; useful life depends on workload mix, cooling design, and how aggressively a company plans to keep running older chips for less demanding jobs even after newer ones arrive. But the divergence between Amazon and Meta in the same year, moving in opposite directions, is itself evidence that these estimates carry real management discretion — and real incentive to lean optimistic when the market rewards earnings growth.

## For hyperscalers

The pressure here is asymmetric. Extending useful life boosts reported earnings today at the cost of a bigger writedown risk later if hardware really does need replacing sooner than assumed. Shortening it depresses today's numbers but reduces future surprise. Given how much of 2026's market enthusiasm for AI infrastructure spending rests on hyperscalers showing healthy margins alongside massive capex, the incentive to lean toward longer useful lives is obvious — and that's precisely why analysts are now scrutinizing the footnote instead of skipping past it.

## For investors

If Burry's numbers are directionally right, the AI capex boom's reported profitability is somewhat inflated, and the correction — when accounting assumptions catch up to actual hardware replacement cycles — could hit reported earnings without any change in the underlying business. That's a very different risk than a demand shortfall; it's a risk that shows up purely in how a real business's results get reported. Edgewisely has tracked the debt and capex side of this story separately in [Broadcom's debt-fueled AI bet](https://edgewisely.com/broadcoms-debt-fueled-ai-bet/?ref=edgewisely.com) and [AI's $100 billion debt habit](https://edgewisely.com/ais-100b-debt-habit/?ref=edgewisely.com), both of which compound the same underlying question: how much of today's AI infrastructure profitability is real versus assumed. Investors pricing AI infrastructure stocks on trailing or forward earnings multiples should treat the useful-life assumption as a lever that can move those earnings by double-digit percentages without a single customer canceling a contract.

## Why this doesn't mean the AI boom is fake

It's worth separating two different claims that get conflated in this debate. One is that AI demand is real and growing — evidenced by contracted backlogs at neocloud operators, Nvidia's continued order book, and enterprise adoption data. The other is that reported earnings from that demand are being measured generously. Both can be true simultaneously. Depreciation policy is an accounting choice layered on top of a real business, not a verdict on whether the business itself is sound.

## The takeaway

Analysts tracking this debate believe hyperscalers will likely converge toward something closer to a five-year depreciation cycle — shorter than today's prevailing five-to-six-year assumption, but still longer than the two-to-three-year hardware-replacement pace Burry argues is the real economic reality. Watch this footnote in the next round of 10-Ks and 10-Qs as closely as the headline capex and revenue numbers. If a hyperscaler quietly revises its useful-life assumption downward, that's a tell about how confident its own accountants are in the numbers Wall Street has been cheering.

*The AI boom's balance sheets are real. Whether its income statements are measuring the boom accurately is a separate question — and it's the one nobody wants to answer in an earnings call.*

## Frequently Asked Questions

### What is the GPU depreciation debate in the AI industry actually about?

It's a dispute over how many years hyperscalers should assume their AI chips remain useful for accounting purposes. Most cloud providers currently depreciate GPUs over five to six years, but critics argue Nvidia's roughly 12-to-18-month chip refresh cycle means the real economic useful life is closer to two to three years, which would make current reported profits look larger than the underlying hardware economics justify.

### Who is Michael Burry and why does his view matter here?

Michael Burry is the investor known for correctly betting against the U.S. subprime mortgage market before the 2008 financial crisis, a bet later depicted in "The Big Short." In 2026, he argued publicly that hyperscalers were understating depreciation to inflate earnings, estimating a cumulative industry-wide overstatement that could exceed $176 billion between 2026 and 2028, according to CNBC's reporting on his analysis.

### Does a shorter depreciation schedule mean the AI boom isn't real?

No. Depreciation policy is an accounting estimate applied on top of real demand, contracts, and hardware purchases; it doesn't measure whether AI adoption itself is genuine. A company can have real, growing AI revenue while still reporting profit margins that look better than they would under a more conservative depreciation assumption.

### What should investors watch for going forward?

Investors should watch whether hyperscalers revise their disclosed useful-life assumptions downward in future 10-K and 10-Q filings. Analysts tracking the debate expect the industry to converge toward roughly a five-year cycle. A downward revision at any single company is a signal that its own accounting team sees less confidence in the longer assumption than its reported earnings currently reflect.

Editor's note — sources: CNBC, Yahoo Finance, industry analysis of hyperscaler depreciation-policy trends (2020–2025), public company 10-K/10-Q useful-life disclosures.