In his latest Substack article, Dr. Michael Burry has returned to a theme that defined his career: the gap between market narratives and fundamental reality. This time, his target isn’t subprime mortgages or overleveraged banks, but rather the infrastructure underpinning the artificial intelligence revolution. And in characteristic fashion, he’s not questioning whether AI is transformative—he’s questioning whether the economics of building it make sense.

Burry’s short position on Nvidia represents more than a bet against a single company. It’s a thesis about depreciation schedules, failure rates, accounting practices, and the uncomfortable mathematics of building infrastructure for a technology whose ultimate profitability remains theoretical. As always with Burry, the devil is in the details everyone else is ignoring.

The Electrification Concession

Burry begins with a notable admission: “AI really is like electrification.” This isn’t a capitulation to the bulls; it’s a clarification. The electrification of America was indeed transformative, fundamentally reshaping society and creating enormous value. But here’s what the comparison actually tells us:

When Burry compares AI to electrification, he’s not validating current valuations. He’s highlighting that transformative technology and profitable investment are not synonymous. Railroads transformed America too. Most railroad companies went bankrupt.

The GPU Failure Rate: Why 9% Matters

The centerpiece of Burry’s analysis is deceptively simple: Nvidia GPUs fail at a rate of approximately 9% in their first year. On its face, this seems manageable. One in ten chips failing doesn’t sound catastrophic—it’s a cost of doing business, easily factored into budgets and planning.

But Burry’s insight goes deeper. It’s not about the 9% in isolation; it’s about what that failure rate means in the context of how GPUs are actually used:

This is the critical point. Modern AI training doesn’t happen on individual chips. It happens across vast arrays of GPUs working in perfect synchronization. When you’re training a large language model across thousands of chips, a single failure doesn’t just cost you one chip—it can halt the entire training run.

The Synchronous Training Problem

Imagine an orchestra with 10,000 musicians. If one musician plays a wrong note, the entire performance is compromised. Now imagine that roughly 900 of those musicians will play a wrong note at some point during the first year. You can implement countermeasures:

• Redundancy (extra musicians standing by)
• Error correction (mechanisms to detect and isolate wrong notes)
• Regular maintenance (constantly checking each musician’s performance)
• Graceful degradation (designing performances that can survive some errors)

Each of these countermeasures works. Each of them also costs money, requires maintenance staff, reduces efficiency, and introduces complexity. These aren’t trivial costs that can be waved away. They’re structural costs that eat into the economics of AI training at scale.

The Depreciation Schedule Question

Here’s where Burry’s accounting background becomes crucial. The question isn’t just about failure rates; it’s about how hyperscalers are accounting for the infrastructure they’re building. Specifically:

If hyperscalers are depreciating GPU infrastructure over five or seven years, but actual useful economic life is shorter due to compounding failure rates and technological obsolescence, then current earnings are overstated. It’s not fraud in the criminal sense, but it is what Burry terms “blessed fraud”—aggressive accounting assumptions that markets accept during boom times.

Consider the cascade: 9% first-year failure rate, higher in year two, higher still in year three. Meanwhile, newer, more efficient chips are released, making older infrastructure relatively obsolete. The actual useful life of these assets may be far shorter than the depreciation schedules suggest.

Power: The Inventory Problem Nobody’s Solving

Burry mentions almost in passing that “Power is an inventory problem,” but this deserves expansion. The AI buildout isn’t constrained by capital or even by chip availability anymore. It’s increasingly constrained by electrical power.

Data centers training large AI models require staggering amounts of electricity. The infrastructure to deliver this power doesn’t exist in many locations, and building it takes years—not months. You can’t scale AI training faster than you can scale power generation and transmission. This is a hard physical constraint that no amount of capital can immediately overcome.

This creates an inventory problem in the following sense: hyperscalers are buying GPUs faster than they can deploy them productively. The chips sit waiting for power infrastructure, depreciating on balance sheets while generating no revenue. It’s like building a factory before you’ve secured the electricity to run it—economically questionable at best.

Why Short Nvidia?

Given all this, why specifically short Nvidia rather than the hyperscalers themselves? Burry’s reasoning, while not fully spelled out in the tweet, likely follows several paths:

1. The Picks and Shovels Myth

There’s a persistent belief that in a gold rush, you should sell picks and shovels rather than mine for gold. Nvidia is the pick-and-shovel seller in the AI rush. But this analogy breaks down when:

• The “mines” (hyperscalers) aren’t yet profitable from their mining
• The picks (GPUs) wear out faster than expected
• New, better picks are constantly being developed
• The miners have strong incentive to develop their own tools

2. Customer Concentration Risk

Nvidia’s revenue is heavily concentrated among a handful of hyperscalers. If any of these companies decide that the economics of AI infrastructure don’t work, or develop their own chips (as several are doing), Nvidia’s moat narrows considerably. The company’s valuation assumes both continued massive purchases and pricing power, but customer concentration threatens both.

3. The Accounting Cascade

If Burry is correct about hyperscalers overstating earnings through aggressive depreciation, the correction won’t be gradual. It will be sudden, as accounting revisions tend to be. When earnings are restated, capex plans are revised, and GPU orders dry up overnight. Nvidia would face a demand cliff, not a gentle slope.

4. The Cycle Turns

Nvidia has benefited from a perfect storm: massive capex budgets, limited competition, supply constraints that supported pricing, and a narrative that AI justifies any cost. But cycles turn. Supply constraints ease, competition emerges, and the question shifts from “how much AI can we build?” to “is this AI generating returns?”

The Fraud That Everyone Blesses

Burry’s use of “blessed fraud” is provocative but precise. He’s not alleging criminal activity. He’s describing a phenomenon where aggressive accounting assumptions become accepted practice during boom times. It’s “fraud” in the sense that it misrepresents economic reality; it’s “blessed” because markets, analysts, and regulators all choose to look the other way.

We’ve seen this pattern before. During the dot-com boom, pro forma earnings and eyeballs-as-metrics were blessed fraud. During the housing bubble, stated-income loans and optimistic default assumptions were blessed fraud. In each case, the fraud was obvious to anyone who looked carefully, but looking carefully was discouraged because the party was too much fun.

The current blessed fraud, per Burry, is assuming that AI infrastructure will maintain value and usefulness over depreciation schedules that don’t account for failure rates, technological obsolescence, or the possibility that AI won’t generate the returns needed to justify the buildout.

The Data We Don’t Have

One of Burry’s most important points is what we don’t know. We have first-year failure rates. We don’t have robust data on years two, three, four, and beyond for the specific Nvidia chips being deployed at scale today. We have semiconductor data in general, and that data suggests failure rates accelerate over time, but we lack the specific numbers.

This uncertainty cuts both ways. It’s possible that Nvidia’s chips prove more durable than semiconductor norms. It’s also possible they prove less durable, particularly given the extreme conditions under which they operate (high power draw, intensive computational loads, massive parallel arrays).

What we do know is that hyperscalers are making multi-billion dollar capital allocation decisions, and markets are valuing companies on the assumption that these decisions are sound, while key data points remain unknown. That’s not investing; that’s speculation with an accounting veneer.

Is Burry Right?

The honest answer is we don’t know yet, and that’s precisely Burry’s point. His track record suggests we should take the analysis seriously, even if the timing remains uncertain. Let’s consider the scenarios:

Scenario 1: Burry is Wrong

AI generates sufficient returns to justify current infrastructure spending. GPU failure rates prove manageable. Depreciation schedules are appropriate. Power constraints are overcome. The buildout was rational, and Nvidia’s valuation was justified. Burry loses on his short position, as he has on some trades before.

Scenario 2: Burry is Right but Early

The fundamental analysis is sound, but markets can remain irrational longer than short-sellers can remain solvent. The AI narrative persists for years before reality asserts itself. Burry may be right about the destination but wrong about the timeline, forcing him to cover at a loss or sustain years of mark-to-market pain.

Scenario 3: Burry is Right and Timely

The accounting chickens come home to roost sooner than markets expect. Hyperscalers announce earnings restatements. AI returns disappoint. Power constraints prove insurmountable. GPU demand collapses. Nvidia’s stock craters, and Burry is vindicated both analytically and financially.

Based on his history, Scenario 2 seems most likely if his analysis is correct. Burry has repeatedly been right about fundamental problems but early on timing. The question for other investors isn’t whether to copy his trade, but whether to take seriously his analysis of AI economics.

What This Means for Investors

Even if you don’t short Nvidia, Burry’s analysis demands engagement with several questions:

• Are you comfortable with the depreciation assumptions in hyperscaler earnings?
• Do you understand the true total cost of ownership for AI infrastructure?
• Have you factored in GPU failure rates and their cascade effects?
• Are power constraints limiting deployment more than markets recognize?
• What’s your thesis for when and how AI generates returns that justify current spending?

These aren’t rhetorical questions. They’re the homework that prudent investors should be doing before participating in or betting against the AI trade. Burry has done his homework. The question is whether the rest of the market has done theirs, or whether they’re simply riding narrative momentum.

The Pattern Repeats

What makes Burry’s latest thesis particularly interesting is how it echoes his past successes. In each case, he identified a gap between accounting/market narrative and physical/economic reality:

In each case, the gap was visible to anyone who looked carefully at the underlying data rather than accepting the prevailing narrative. In each case, markets considered the narrative more compelling than the data, until suddenly they didn’t.

The Maintenance Nobody Wants to Discuss

Return to Burry’s point about countermeasures and maintenance: “they all have a cost and involve maintenance, which is not free.” This seems obvious, yet it’s systematically underweighted in AI economics discussions.

Maintaining massive GPU clusters isn’t like maintaining traditional IT infrastructure. The scale is unprecedented, the complexity is extreme, and the synchronous nature of AI training means failures have cascading effects. Yet most analyses of AI infrastructure costs focus on acquisition and power, treating maintenance as an afterthought.

If Burry is correct that maintenance costs are significantly higher than markets are modeling, and if these costs are being capitalized rather than immediately expensed, then earnings quality is worse than it appears. This wouldn’t be the first time technology companies used accounting discretion to smooth earnings by capitalizing costs that should be expensed.