Michael Burry May Be Early — But Big Tech’s AI Accounting Problem Is Now Unavoidable

The market is still rewarding the AI buildout. The harder question is whether it is properly pricing the economics underneath it.

By Drew Stegman | April 22, 2026

Executive Summary

Michael Burry’s latest AI critique is more serious than a generic “bubble” call. His argument is that hyperscalers are spending so aggressively on AI infrastructure that the accounting assumptions behind those assets now matter as much as the revenue narrative itself. That matters because Microsoft, Alphabet, Meta, Oracle, and Amazon have all made useful-life decisions on servers, chips, and networking gear that materially affect depreciation expense and reported earnings, while at the same time pushing AI capex to historic levels. The bull case is real: demand is real, contracts are real, and monetization is improving. But the accounting question is real too — and it is now too large for investors to dismiss.

The AI boom has reached the point where broad enthusiasm is no longer enough. Investors now need to separate two different questions: whether AI is transformative, and whether today’s reported earnings accurately reflect the economic wear-and-tear of the infrastructure required to power it. Michael Burry’s latest argument is aimed squarely at that second question. In Reuters reporting on his new Substack, Burry compared today’s AI frenzy to the late-1990s dot-com era and wrote that the five public “horsemen” of the boom — Microsoft, Google, Meta, Amazon, and Oracle — are promising nearly $3 trillion of AI infrastructure spending over three years.

That framing is provocative, but the underlying issue is not fringe. The filings show that useful-life assumptions for servers and networking gear have already moved in ways that can materially affect reported profit. Microsoft disclosed that it increased the useful lives of server and network equipment from four years to six years, effective in fiscal 2023. Alphabet changed the useful lives of servers and certain network equipment to six years, and said that change reduced depreciation expense by $988 million in the first quarter of 2023 alone. Meta said in January 2025 that it increased the estimated useful life of certain servers and network assets to 5.5 years, a move expected to reduce 2025 depreciation expense by about $2.9 billion. Oracle likewise said it increased the useful lives of servers and networking equipment from five years to six years, which reduced operating expenses by $567 million and increased net income by $442 million in the first nine months of fiscal 2025.

This is the heart of the story. Depreciation is not a side note anymore. Once AI capex becomes large enough, small changes to useful-life assumptions become large changes to earnings presentation. That does not automatically make those estimates wrong, and it certainly does not prove fraud. But it does mean the market has to scrutinize these assumptions much more aggressively than it did when infrastructure investment was smaller and less central to the valuation story. Burry’s critique lands because the numbers are now simply too big to wave away as accounting trivia.

The strongest evidence that this issue is real — and not just a one-directional talking point — actually comes from Amazon. Amazon disclosed that effective January 1, 2025, it shortened the useful lives of a subset of servers and networking equipment from six years to five years because of the faster pace of technological development, particularly in AI and machine learning. In its third-quarter filing, Amazon said that change increased depreciation and amortization expense by $889 million for the first nine months of 2025 and reduced net income by $677 million, primarily affecting AWS. In other words, Amazon is showing investors the exact other side of the trade: when useful lives move the other way, earnings get hit.

That is why the right conclusion is not “Big Tech is faking AI profits.” The right conclusion is more disciplined: the accounting treatment of AI infrastructure is now material enough that investors must underwrite it as part of the thesis. If AI chips and networking equipment really have a shorter economic life than the market is assuming, then future depreciation pressure could be meaningfully higher than investors expect today. Reuters reported in November that Burry estimated these accounting choices could understate depreciation by about $176 billion between 2026 and 2028. That figure is Burry’s estimate, not a settled fact, but it captures the scale of the debate now entering the market.

The bulls, however, are not arguing from thin air. They have real numbers too. Amazon CEO Andy Jassy wrote this month that AWS’s AI revenue run rate is now over $15 billion in Q1 2026 and that Amazon’s chips business is running at over $20 billion annually. He also said Amazon is not investing approximately $200 billion in 2026 capex “on a hunch,” adding that AWS already has customer commitments for a substantial portion of the capacity it expects to spend against in 2026. Reuters separately reported this week that Anthropic committed to spend more than $100 billion on Amazon’s cloud technologies over the next decade.

Microsoft’s disclosures also give the bull case real weight. In fiscal Q2 2026, Microsoft said capital expenditures were $37.5 billion, with roughly two-thirds of that spend on short-lived assets such as GPUs and CPUs. Microsoft also disclosed commercial remaining performance obligation of $625 billion, up 110% year over year, with a weighted average duration of about 2.5 years, and said approximately 45% of that balance is from OpenAI. On the earnings call, investors explicitly pressed management on the mismatch between six-year server lives and shorter contract durations. Microsoft’s answer was revealing: it said much of the GPU capacity being bought today is already contracted for most or all of its useful life. That does not settle the debate, but it does show management is prepared to argue that the monetization pipeline is real, not hypothetical.

The same is true at Alphabet and Meta. Alphabet spent $91.4 billion on capital expenditures in 2025, recorded $21.1 billion of depreciation on property and equipment, and guided to $175 billion to $185 billion of 2026 capex. Meta spent $72.22 billion on capex in 2025 and guided to $115 billion to $135 billion for 2026, while warning that expense growth would be driven by infrastructure costs, including higher depreciation. These are not niche numbers at the margin of the income statement. They are now central to the investment case.

This is where the story becomes larger than accounting. Reuters reported this week that hyperscalers Microsoft, Amazon, Alphabet, and Meta alone are expected to invest about $635 billion in AI infrastructure this year. Reuters also cited Morgan Stanley estimates that U.S. data-center power demand could rise to 80 gigawatts by 2028, with a possible 55-gigawatt shortfall. In other words, the market is not just betting on AI software adoption; it is betting that the physical layer — chips, power, land, cooling, networking, and financing — can scale without impairing margins or changing how investors think about returns on capital.

That physical-layer pressure is exactly why Burry’s warning has traction, even if he is early. When enthusiasm is strong, investors tend to focus on bookings, run rates, and TAM expansion. When conditions tighten, they start looking harder at depreciation schedules, power costs, free cash flow conversion, and the true duration of customer commitments. Reuters reported today that Google is now putting AI agents at the center of its enterprise monetization push, highlighting a broader race among major players to prove that AI can evolve from massive capex into durable, enterprise-grade revenue streams. The market is no longer paying only for promise. It is beginning to demand proof.

The most important takeaway for investors is straightforward. The AI boom can still be real and the accounting question can still matter. These are not mutually exclusive ideas. In fact, the more real the boom becomes, the more important it is to understand how the economics are being recognized through the income statement. That is the real institutional question now: not whether AI matters, but whether current valuations fully reflect the capital intensity, depreciation risk, and power burden required to sustain it.

Burry may be early. He has been early before. But the market no longer has the luxury of dismissing his argument as pure macro theatrics. The accounting question is now unavoidable because the spending is now unavoidable. And when capital spending reaches this scale, the distance between reported profit and economic reality becomes one of the few questions that truly matters.

Ready to Analyze Your Next Investment?

Get a free AI-powered fair value analysis on any stock. See intrinsic value, margin of safety, and institutional-grade risk metrics in seconds. No credit card required.

Want full access to our institutional research tools? Explore Invest Daily Pro.