Over the past two years, Amazon (NASDAQ:AMZN), Alphabet (NASDAQ:GOOG)(NASDAQ:GOOGL), and Microsoft (NASDAQ:MSFT) have collectively committed nearly $3 trillion to AI infrastructure. Microsoft alone plans $80 billion in fiscal 2025 capex, mostly data centers; Alphabet raised its 2025 guidance to $75 billion; and Amazon’s AWS is on pace for $100+ billion annually by 2026.
The stated goal: lock in dominance before rivals do. Wall Street cheers every upward revision, sending these stocks to repeated all-time highs. Earnings calls overflow with phrases like “once-in-a-lifetime opportunity” and “land-grab phase.” Analysts now bake in 20% to 30% cloud and AI revenue growth for the next half-decade.
Yet beneath the euphoria is a quiet accounting choice that is making those rosy numbers possible — and one investor says it won’t end well.
Michael Burry, the man who correctly called the subprime mortgage crisis in 2008, has taken aim at AI hype. Not just targeting Nvidia (NASDAQ:NVDA), which he says is supplying an AI build out that doesn’t have the end user demand to support it, but also the hyperscalers, accusing them of systematically under-depreciating the very GPUs and servers they brag about buying.
In his new Substack, “Cassandra Unchained,” Burry calls it “one of the more common frauds of the modern era,” not outright Enron-level fabrication, but aggressive life extensions that inflate current profits at the expense of future writedowns.
He’s doing the hard work most investors won’t: digging into the SEC filings to find nuggets of concern companies don’t want you to see. Depreciation of assets is one of them.
Depreciation is the accounting process of spreading the cost of a physical asset (like servers or GPUs) over its estimated useful life. Instead of expensing the full purchase price in year one, a company recognizes a portion of the cost as an expense each year, reducing reported profits gradually while matching the expense to the periods the asset generates revenue.
Burry says the industry could be understating depreciation by $176 billion between 2026 and 2028, enough to inflate combined earnings by roughly 20%. By extending the assumed life of these assets, companies are lowering their annual depreciation expense and boosting near-term earnings. If they shorten them — which Burry argues they should — it has the opposite effect.
The timing is deliberate, as most of these chips won’t hit the back half of their assumed lives until 2026 to 2027, conveniently after the current hype cycle peaks.
Burry’s central claim is simple: Nvidia designs its GPUs for planned obsolescence on a 12- to 18-month cadence. For example, Grace Blackwell began rolling out in 2024 and the next-generation Rubin chips are due out in the second half of 2026. CEO Jensen Huang has said because demand for the chips is growing so fast, it is necessary for Nvidia to accelerate the pace of its design cycle to improve performance to match. Standing still is not an option.
That means that GPUs that had a useful life of four years are now rendered obsolete in half the time. Yet the hyperscalers are depreciating these assets over longer periods, in many cases four to six years, sometimes longer.
Microsoft, for example, discloses a four-year life for “high-performance servers,” but recently stretched it to six years for certain classes. Alphabet similarly moved from a three- to four-year schedule to four to six years over the past three years.
Amazon is an exception: it actually shortened lives from six to five years in 2024, citing “changes in estimated useful lives of servers,” a move Burry views as the canary acknowledging reality.
When useful lives are extended while technological obsolescence accelerates, every extra year delays roughly $30 billion to $40 billion in annual expense recognition across the sector. By Burry’s back-of-the-envelope calculation, if realistic economic lives are two to three years instead of five to six, catch-up depreciation plus future expense will wipe out most of the incremental AI profit Wall Street expects from 2026 onward. Margins that look 500 to 700 basis points wider today because of slow depreciation will snap back hard tomorrow.
So, despite Nvidia saying Burry is wrong on the useful life of GPUs, the chipmaker is operating on the opposite principle. It wants to have it both ways: a long lifespan for its products, but also an “insane” annual release cycle.
A forced normalization to two- to three-year depreciation life would trigger immediate impairment charges and permanently impose higher ongoing expenses — potentially cutting 2026 to 2028 EPS at Microsoft and Alphabet by 15% to 25%. Multiples would compress from growth to virtual stagnation overnight. Amazon’s shorter depreciation life gives it a relative head start, but Burry says AWS margins still face 500 to 800 bps of pressure.
Investors pricing in perpetual 25% AI growth are effectively going long on an accounting tailwind that expires in 12 to 24 months. Volatility will spike when the first hyperscaler takes the write-down, and history suggests they move in packs. The dominoes will rapidly fall.
The bottom line is, while the AI capex boom is real, a large chunk of the boost to profits hyperscalers and others are enjoying is really borrowed from the future. And when the music finally stops, it won’t be pretty.
You may think retirement is about picking the best stocks or ETFs, but you’d be wrong. See even great investments can be a liability in retirement. The difference comes down to a simple: accumulation vs distribution. The difference is causing millions to rethink their plans.
The good news? After answering three quick questions many Americans are finding they can retire earlier than expected. If you’re thinking about retiring or know someone who is, take 5 minutes to learn more here.
