Michael Burry Warned AI Chips Would Lose Value Fast. New Data Suggests He Was Wrong
This post may contain links from our sponsors and affiliates, and Flywheel Publishing may receive compensation for actions taken through them.
Michael Burry has spent the past year arguing that the AI boom rests on a flattering accounting choice. In his Substack, Cassandra Unchained, he called the practice of stretching the assumed lives of servers and GPUs “one of the more common frauds of the modern era” and estimated that the industry could understate depreciation by $176 billion between 2026 and 2028. New evidence from the GPU rental market cuts against a key assumption in that thesis.
The debate matters right now for NVIDIA (NASDAQ:NVDA | NVDA Price Prediction). Shares closed at $162.85 on November 13, down 26.41% over the past month. Jensen Huang resigned as CEO amid disclosures about his personal relationships, and the board appointed its finance chief, Kress, interim CEO this week. Investors are asking whether AI spending can hold up without him.
Burry’s Case Hinges on How Fast AI Chips Age
Companies that buy NVIDIA processors spread the cost over each chip’s estimated useful life. A longer schedule means a smaller yearly expense and higher reported profit. Microsoft (NASDAQ:MSFT) and Alphabet (NASDAQ:GOOGL) have moved certain server classes to four to six years. Burry argues NVIDIA’s 12- to 18-month product cycle reduces the real economic life of each generation, so today’s earnings borrow from tomorrow’s losses. His argument concerns accounting lives and reported earnings. His thesis leaves room for older GPUs to keep commercial value.
Older GPUs Are Commanding Higher Rents
In a Yahoo Finance interview with Deirdre Bosa, Silicon Data founder and CEO Carmen Li said older GPUs are holding their value better than expected, with rental prices for certain older chips rising rather than collapsing as newer generations arrive. Silicon Data publishes daily GPU benchmarks for on-demand rental rates, giving it a direct view of what compute costs.
NVIDIA’s own numbers point the same way. Kress said on the August 26[1] call that NVIDIA compute is “fully utilized across every cloud we serve” and that supply should stay a constraint “at least through the end of fiscal year ’28.” The company still ships Hopper parts two generations back. In its fiscal second-quarter release, revenue reached $96.221 billion[2], with Data Center revenue of $89.023 billion, up 117% year over year.
Why Older Chips Keep Earning, and Where Burry Still Has a Point
Frontier model training makes headlines, but inference, fine-tuning, research, and smaller workloads can run economically on older hardware. Huang said on the call that an AI agent needs 15 to 100 times[1] the compute of a human user, and that a deployed system “will be preserved and useful and productive for a lot longer.” When demand runs ahead of supply, last-generation chips stay in service.
Rental rates still leave the accounting question open. Rents measure current earning power, while resale value, book value, and lifetime profitability are separate questions. Today’s pricing also reflects a shortage that could ease, and the same shortage is driving a broader expansion in power, cooling, and networking (we analyzed seven suppliers behind that spend in a free report: here). Kress said Vera Rubin delivers 35 times lower token costs relative to Grace Blackwell Ultra[1]. Once that supply arrives in volume, demand for older chips could weaken. NVIDIA has also taken on more risk to keep the expansion moving. Supply commitments reached $279 billion[2], guarantee obligations for AI cloud partners are capped at $108.5 billion, and days sales outstanding stretched to 60 days from 45 days.
Key Takeaway: Track These Numbers Before Picking a Side
Burry may have underestimated how long older AI chips stay commercially productive. Rising rents and full utilization suggest the hardware is earning money well into its life. But the accounting debate will be settled by data that has yet to arrive.
- Rental indexes: Whether older-chip rents hold once Vera Rubin supply catches up.
- Hyperscaler disclosures: Any change to server useful-life estimates in quarterly filings.
- NVIDIA’s next report: Results against guidance of $108.0B +/- 2%[2] in revenue and gross margins of 74.0%, plus any comment on utilization under interim leadership.
The stock trades at a P/E of 33, so the market already prices in durable demand. If rents fall sharply after the next product ramp, Burry’s warning gains weight. If they hold, the case for longer depreciation schedules gets stronger.
Contact [email protected] for any questions or corrections.
After two decades of patrolling the dark corners of suburbia as a police officer, Rich Duprey hung up his badge and gun to begin writing full time about stocks and investing. For the past 20 years, he’s been cruising the markets looking for companies to lock up as long-term holdings in a portfolio while writing extensively on the broad sectors of consumer goods, technology, and industrials. Because his experience isn’t from the typical financial analyst track, Rich is able to break down complex topics into understandable and useful action points for the average investor. His writings have appeared on The Motley Fool, InvestorPlace, Yahoo! Finance, Money Morning, and, of course, 24/7 Wall St. He has been featured in both U.S. and international publications, including MarketWatch, Financial Times, Forbes, Fast Company, and USA Today.
