Michael Burry says Big Tech stretched the life of its AI chips. Here's the accounting behind the fight.

In November 2025, Michael Burry — the investor who bet against the housing market in The Big Short — pointed at one line on Big Tech's income statement and called it a polish job. His claim is that the largest AI spenders are quietly inflating reported profits, and the tool is depreciation.

The mechanism he described is something every CFA Level I candidate meets early in financial reporting. When a company buys a long-lived asset, such as a building, a machine, or a rack of AI servers, it does not expense the whole cost at once. It capitalizes the cost and spreads it across the asset's useful life as depreciation. The most common method, straight-line, is a single division:

The useful life sits in the denominator. Make it bigger and each year's depreciation gets smaller. That means a smaller expense, and a larger reported profit, from the very same hardware.

That is precisely the change Burry flagged. According to company filings, the hyperscalers have been lengthening the assumed lives of their servers: Alphabet moved from four years to six, Microsoft from four to six, Oracle from five to six, and Meta to about five and a half. Burry's argument is that this stretches the book life of chips whose real product cycle is two to three years. He estimates the practice could understate depreciation by roughly $176 billion across 2026 through 2028 and, on his numbers, leave companies like Oracle and Meta overstating earnings by something like 27% and 21%.

Why one assumption moves so much

Walk a single purchase through it. Suppose a company buys $6 billion of servers with no salvage value.

• At a four-year life, annual depreciation is $6B / 4 = $1.5 billion.
• At a six-year life, it is $6B / 6 = $1.0 billion.

The cash is already spent and the machines are the same, but there is $500 million less expense in the first year, and so $500 million more pre-tax profit. Nothing in the bank account changed; only the timing of the expense did. Over the asset's whole life the total depreciation is identical. The assumption simply decides how much lands in this year versus later years.

And the effect does not stop at net income. A lower depreciation charge leaves a higher asset value on the balance sheet, which raises return on assets but lowers asset turnover. The curriculum's directional table is exactly this: a less aggressive depreciation schedule produces higher early-year earnings, higher assets, and higher ROA.

Aggressive, or just correct?

Extending a useful life is not, by itself, wrong. If older chips really do keep working, repurposed from training the newest models to running cheaper inference, a longer life can be the more honest assumption. Depreciation is meant to match an asset's cost to the periods it earns revenue. The whole dispute is whether a six-year book life is a fair read of how long an AI server earns, or an aggressive one chosen because it flatters earnings.

This is reporting-quality analysis in miniature. Financial statements are built on estimates such as useful lives, salvage values, and impairment triggers, and each estimate is a lever. The analyst's job is to find the assumptions that matter, judge whether they are conservative or aggressive, and adjust before comparing one company with another. Two firms with identical servers and identical cash flows can report different profits purely because they set this one assumption differently.

What it means for the exam

Long-lived assets are high-density on Level I — two to three questions a cycle — and they reward exactly this instinct. Know that straight-line spreads cost evenly while accelerated methods front-load it; know that over the full life the total expense is the same regardless of method or assumed life; and know that a change in useful life flows straight through to net income, assets, and ROA.

Whether Burry is right about Big Tech is a question for the next few years of filings. That a single estimate can swing reported profit by double digits is not in doubt, and spotting the estimate that matters is much of what financial-reporting analysis involves.