Artificial intelligence infrastructure spending has become the engine of the American economy. In the first half of 2025, AI-related capital expenditures contributed more to U.S. GDP growth than consumer spending—a first in economic history. Deutsche Bank has argued that without this investment surge, the United States might already be in recession.
This context makes Michael Burry’s return to public commentary all the more striking. The investor who famously predicted the 2008 housing collapse has turned his attention to AI, launching a Substack newsletter and disclosing bearish positions against Nvidia and Palantir. His thesis isn’t that artificial intelligence is fake or useless. It’s that the economics of the current buildout don’t add up—and that accounting practices are obscuring the problem.
The question for investors isn’t whether AI will matter. It’s whether the current spending spree represents a sound industrial investment or a financial mania that will leave shareholders holding depreciated assets and inflated expectations.
Accounting concerns: depreciation and earnings
At the center of Burry’s argument is a technical but consequential issue: how long do AI chips actually remain economically useful?
When companies like Meta, Microsoft, and Oracle buy Nvidia GPUs, they depreciate the assets over their estimated useful life, spreading the cost across multiple years. The longer the depreciation schedule, the lower the annual expense—and the higher the reported earnings.
Burry argues that hyperscalers have stretched these schedules beyond what the technology supports. Meta’s filings confirm that effective January 2025, the company extended server useful lives from four or five years to five and a half years. Other major cloud providers have adopted similar six-year schedules.
The problem: Nvidia now releases new chip generations annually, each offering substantial improvements in performance and energy efficiency. A six-year depreciation schedule assumes a chip purchased today will remain economically productive through 2031. Burry estimates this mismatch could understate depreciation by roughly $176 billion across major hyperscalers between 2026 and 2028, inflating reported earnings at companies like Meta and Oracle by 20% or more.
Nvidia has pushed back. On its November 2025 earnings call, CFO Colette Kress emphasized that the company’s CUDA software platform extends the productive life of older hardware. “Thanks to CUDA, the A100 GPUs we shipped six years ago are still running at full utilization today,” she said.
Burry’s response was pointed: physical utilization and economic value are not the same thing. He compared the situation to airlines keeping aging aircraft in service during peak seasons. The planes fly, but they’re marginally profitable at best.
Do old chips still have value?
The technical reality is more nuanced than either side suggests. AI workloads fall into two categories: training and inference. Training involves building new models and requires cutting-edge hardware. Inference—running trained models to generate outputs—is less demanding and can run on older equipment.
This distinction matters. CoreWeave’s CEO noted in November 2025 that the company’s Nvidia A100 chips—first announced in 2020—are fully booked. When H100 chips came off contract, they were re-leased at 95% of their original price. “All of the data points I’m getting are telling me that the infrastructure retains value,” he said.
But Burry raises a valid counterpoint. Older chips consume significantly more power than their successors. For organizations running chips around the clock, electricity costs compound quickly. There’s also the question of physical infrastructure—newer data centers are built with cooling and electrical configurations optimized for specific chip generations. An older GPU may not slot easily into a facility designed for its successors.
The honest conclusion: older AI chips retain some residual value, particularly for inference. But that value declines faster than a six-year depreciation schedule implies. The precise useful life probably falls somewhere between Burry’s estimate of two to three years and the hyperscalers’ assumption of five to six. That gap matters enormously for reported earnings.
Financial bubble or industrial buildout?
Not all bubbles are equal. Financial bubbles destroy capital with little lasting benefit. Industrial bubbles overbuild infrastructure that eventually finds productive use, even if the original investors lose money.
The telecom boom of the late 1990s offers an instructive parallel. Between 1996 and 2001, telecommunications companies invested over $500 billion laying more than 80 million miles of fiber optic cable. Then the bubble burst. Companies collapsed, stock prices fell 80%, and by 2002 only 3% to 5% of installed capacity was in use.
But the fiber remained in the ground. Within four years, bandwidth costs fell 90%. That excess capacity enabled YouTube, Netflix, and the smartphone ecosystem to scale without building infrastructure from scratch. The investors lost; society gained.
Burry explicitly invokes this history, calling Nvidia “Cisco, not Enron”—a real company at the center of a real technology transition, but one whose valuation may not survive contact with reality. Cisco’s stock rose 3,800% between 1995 and 2000, then fell over 80% and took nearly two decades to recover.
Will AI data centers follow the same pattern? There are reasons for skepticism. Fiber optic cables are passive infrastructure requiring minimal maintenance. GPUs consume power, generate heat, and face both physical degradation and technological obsolescence. A data center built around 2024-era chips may need substantial retrofitting for 2028-era hardware.
AI capex likely sits somewhere between a pure financial bubble and a lasting industrial buildout. Some value will persist, but the assumption that today’s spending will generate returns for five or six years faces real headwinds.
An economy leaning on AI spending
The macroeconomic picture adds another layer of risk. AI-related capital expenditure contributed an estimated one percentage point or more to U.S. GDP growth in 2025—potentially half of total growth. Bank of America estimates that Microsoft, Amazon, Alphabet, and Meta will spend over $340 billion on capital expenditures this year.
Outside of AI, the picture is less encouraging. Private business investment excluding AI has essentially flatlined since 2019. Commercial construction outside data centers is declining.
This concentration creates asymmetric risk. If confidence wavers—whether due to disappointing returns, tightening credit, or shifting sentiment—the effects could cascade quickly. Peter Berezin of BCA Research put it bluntly: “If you take a fragile labor market and you kick it with a capex bust, you’re probably going to get a recession.”
What This Means for Investors
No one knows whether we’re in 1995 or 1999—early in a transformative buildout or late in an unsustainable mania. What seems clear is that portfolios heavily concentrated in AI beneficiaries face asymmetric risk.
Balanced positioning matters. We’ve maintained successful portfolios with measured AI exposure—not betting against the technology, but not assuming current prices reflect conservative assumptions about depreciation, demand, and returns. How you structure a portfolio for uncertainty matters as much as the individual positions you hold.
Burry’s framework is worth remembering: he’s not calling this Enron. He’s calling it Cisco. The infrastructure may prove valuable over time. The stocks trading at current valuations may not.
If you’re wondering how your portfolio is positioned for what comes next, we can help. Start with four quick questions to see if we’re a good fit.
