Michael Burry's AI Depreciation Thesis: Accounting Distortions and the Looming Overinvestment Risk in Tech

The tech sector's valuation has long been a subject of fascination for investors, but in 2025, a new debate has emerged: the potential accounting distortions created by the rapid depreciation of AI hardware. Michael Burry, the investor known for his 2008 housing crisis short, has turned his attention to the AI infrastructure boom, arguing that major hyperscalers like Microsoft, Amazon, and Google are artificially inflating earnings by extending the useful life of their computing assets. This thesis, if validated, could trigger a reevaluation of tech valuations and expose systemic risks in the AI capital expenditure (capex) frenzy.

The Core of Burry's Argument: Depreciation Mismatch

Burry's central claim is that hyperscalers are depreciating AI chips-such as Nvidia's GPUs-over 5–6 years, despite their actual economic lifespan being only 2–3 years due to rapid technological obsolescence and physical wear according to analysis. For example, Microsoft's $17 billion in GPU purchases are being depreciated over six years instead of three, resulting in an annual earnings overstatement of approximately $2.9 billion, or 3.3% of its 2024 net income according to data. If this pattern persists, the cumulative overstatement could reach $10–12 billion annually, representing 11–14% of reported earnings according to estimates.

This discrepancy creates a "$4 trillion accounting puzzle," as companies spread costs over longer periods than the actual useful life of the hardware. By deferring depreciation expenses, hyperscalers artificially lower early-stage costs, gaining a competitive edge as they lock in customers during the formative AI market. The financial implications are staggering: Burry estimates that correcting Microsoft's depreciation assumptions could reduce its fair value by 37% according to analysis.

Industry Practices and the "Value Cascade"

The debate over AI hardware's economic lifespan is far from settled. While Burry and skeptics argue for a 2–3 year lifespan, companies like CoreWeave and industry analysts defend a 6-year depreciation cycle. CoreWeave cites real-world data showing that older Nvidia A100 chips retain 95% of their original price in expired contracts, suggesting a longer useful life according to data. Others propose a middle ground: a 5-year depreciation cycle, which accounts for repurposing older chips for inference and analytics workloads according to industry analysis.

This "value cascade" allows hyperscalers to maximize revenue from aging hardware, but it also masks the true cost of AI infrastructure. For instance, the Nvidia H100, released in 2022, became economically obsolete with the 2025 launch of the Blackwell chip according to reports. Such rapid obsolescence challenges the logic of 6-year depreciation schedules, particularly as AI workloads intensify thermal and electrical stress on hardware according to research.

Overinvestment Risks and the AI Bubble

The AI capex boom has already reshaped the tech landscape, with hyperscalers projecting combined capex increases of 40% in 2025. However, this surge raises concerns about overinvestment and sustainability. Free cash flow for major cloud providers has turned negative, signaling valuation risks according to Morgan Stanley. Meanwhile, AI venture funding hit $80 billion globally in Q1 2025, with 46% of all VC dollars flowing into the sector despite AI startups representing only 18% of companies raising capital according to analysis.

The parallels to the dot-com bubble are hard to ignore. As stated by industry analysts, "the business case for AI remains untested," with many early corporate AI initiatives failing to deliver meaningful returns according to reports. Vendor financing and off-balance-sheet debt further complicate the picture, as companies like Nvidia and OpenAI engage in deals that obscure revenue quality according to analysis.

The Ripple Effect on Nvidia and the Market

If Burry's thesis gains traction, the consequences could extend beyond hyperscalers. A shift in investor sentiment might reduce demand for AI chips, destabilizing Nvidia's market position. The company's Blackwell chip, launched in 2025, relies on sustained hyperscaler spending, which could falter if depreciation assumptions are revised according to analysis. Additionally, the $2.9 trillion in AI-related data center spending projected from 2025 to 2028-half of which requires external financing-could face scrutiny as lenders reassess risk according to economic data.

Conclusion: A Call for Caution

Burry's thesis highlights a critical blind spot in tech valuations: the interplay between accounting practices and technological obsolescence. While hyperscalers argue that their depreciation schedules align with industry standards, the rapid pace of AI innovation suggests these assumptions may be outdated. For investors, the key takeaway is clear: the AI infrastructure boom is not immune to overinvestment risks. As data centers and GPUs become the new "dot-com" assets, the market must grapple with whether these valuations are built on sustainable fundamentals or speculative accounting.

The coming months will test the resilience of this sector. If hyperscalers fail to demonstrate meaningful returns on their AI investments, the ripple effects could extend far beyond earnings reports-reshaping the entire tech landscape.