In Case You’re in a Hurry

• Michael Burry, who predicted and profited from the 2008 housing crash, believes the AI boom is built on accounting tricks, circular money flows, and unsustainable financial engineering.
• Depreciation manipulation: Major tech companies extended server useful life assumptions from 4 to 6 years, artificially inflating profits by an estimated $176 billion between 2026-2028.
• Circular revenue: Much of AI’s reported growth comes from companies investing in each other and buying each other’s products, not actual end-user demand.
• Return on Invested Capital (ROIC) is falling fast across AI infrastructure companies, meaning they’re spending more to generate less return.
• Stranded asset risk: Data centers financed on 20-year timelines contain chips that become economically obsolete in 2-3 years due to rapid technological advancement.
• Stock compensation dilution: Nvidia spent $112.5 billion on buybacks since 2018 and still has 47 million more shares outstanding than when it started.
• China’s power advantage: China has 3x the electrical generation capacity of the U.S. and is accelerating while American grid development decelerates.
• Market valuations: The Shiller CAPE ratio sits around 40-41, higher than before the 1929 crash and second only to the 2000 dot-com peak.
• Burry’s timeline: He believes “the algebra will inevitably catch up to the narrative in 2026.”

---

The Man Who Saw It Coming

This article accompanies Episode 19 of The Gray Files podcast, where we explore the uncomfortable possibility that everything you’ve been told about the AI revolution, the generational wealth being created, the unstoppable march of progress… well, it might be built on accounting tricks, circular money flows, and the kind of financial engineering that collapsed markets in 2000 and 2008.

The man warning us saw it coming before. His name is Michael Burry. You might know him from the movie “The Big Short,” where Christian Bale played him as the eccentric hedge fund manager who bet against the housing market in 2007 when everyone else thought he was insane. Burry made hundreds of millions of dollars when the mortgage market imploded. He saw the fraud hiding in plain sight. He read the fine print nobody else would read. He did the math nobody else wanted to do. Now he’s doing that math again, and the numbers don’t look good.

So I went and collected every piece of text, articles and tweets Burry has been posting since October of 2025. I put them together to find a logical flow to everything and see the story, and what Burry is describing isn’t a simple thesis. It’s a web of interconnected problems, each one amplifying the others. And if he’s right, even partially right, the implications affect everyone who has money in the stock market, which is to say, almost everyone.

---

Understanding the Analyst Behind the Analysis

Before we dive into the weeds of depreciation schedules and stock compensation, it’s important to understand something about Michael Burry. This is a man who was diagnosed with Asperger’s syndrome as an adult. He processes the world differently. Where most investors follow narratives and momentum and what the crowd is doing, Burry obsesses over primary documents. He reads 10-K filings, those massive annual reports that public companies are required to file with the Securities and Exchange Commission, the way some people read novels. He builds spreadsheets the way some people build model boats in glass bottles.

Originally, he trained as a physician. He went to medical school at Vanderbilt, did his residency in neurology at Stanford. But he never practiced. Instead, he started writing about stocks on internet message boards in the late 1990s, and his analysis was so good that professional money managers started following his posts. He eventually started his own hedge fund, Scion Capital, with money from his family and a few early believers. That fund would go on to produce one of the most famous trades in financial history.

Most people, when they encounter information that contradicts what they want to believe, find reasons to dismiss it. Burry does the opposite. He digs deeper. He reads the footnotes. He follows the money until he finds where it actually goes, and it starts to make him very uncomfortable. That discomfort is what we’re going to explore.

Because Burry writes for an audience of professional investors, I want to make sure you can follow his analysis on your own. So along the way, I’ll explain some of the terminology he uses. Think of it as a mini masterclass in finance while following the man who called The Big Short. By the end of this article, you’ll be able to read his Substack, which he calls “Cassandra Unchained,” and understand exactly what he’s warning about.

---

The Escalator That Eats Your Margins

Let’s go back in time to a department store in 1960. Sears or Macy’s. Three floors of merchandise. Customers have to climb stairs to see everything. The store across the street just installed something new… an escalator. Moving stairs. The future. Now your customers are walking past your store to ride the escalator at your competitor’s. So you have a choice: install your own escalator, or lose market share.

You install the escalator and it costs you a fortune. And here’s the thing nobody talks about: your sales don’t go up. Your competitor’s sales didn’t go up either. You both spent massive amounts of capital, you both now have the same customers you had before, and now you both have maintenance costs and electricity bills you didn’t have before. The escalator didn’t create new shoppers, it just redistributed the existing ones. But every store had to have one. The value of the innovation flowed entirely to the customer, not to the businesses spending the capital.

This is what economists call a “sunk cost” situation. A sunk cost is money you’ve already spent that you cannot recover, regardless of what happens next. Once you build the escalator, the money is gone. You cannot unbuild it and get your capital back. All you can do is hope it keeps you competitive with everyone else who also built escalators. And Michael Burry believes artificial intelligence is the new escalator.

I’m not suggesting that AI is useless. The escalator wasn’t useless, it made shopping more convenient. And AI makes work more convenient. Coding is faster. Research is easier. Customer service chatbots actually work now… mostly. The productivity gains are real. The question Burry is asking is different: Who captures the value?

Right now, the world’s largest technology companies are spending trillions of dollars on chips and data centers and infrastructure. Microsoft. Google. Meta. Amazon. Oracle. They’re in a race. And in a race, you have to keep running just to stay in place. Competitive races often destroy value rather than create it. When one airline buys new planes, all the other airlines have to buy new planes too, or their customers will fly with the competition. The planes get better. The ticket prices stay roughly the same. The airlines collectively spend billions, and the value flows to the passenger.

Return on Invested Capital: The Metric That Matters

Burry likes to track something called Return on Invested Capital, or ROIC. It might be the most important metric in finance that most people have never heard of. Return on Invested Capital measures how much profit a company generates for every dollar it invests in its business. If you invest $100 in new equipment and that equipment helps you generate $15 in additional profit, your ROIC is 15 percent. Simple, but powerful.

ROIC tells you whether a company is creating value or destroying it. A company with high ROIC is turning investment into profit efficiently, so it’s compounding wealth. Each dollar invested generates more than a dollar of value. That’s how fortunes are built. A company with declining ROIC is spending more and more money to generate less and less return. It’s the financial equivalent of running faster and faster just to stay in the same place.

The best businesses in the world have high and sustainable ROIC. Think about a toll bridge: you build it once, and then cars pay you to cross it for decades. The upfront investment is substantial, but the returns are enormous because the ongoing costs are low and the barriers to competition are high. Nobody is going to build a second bridge right next to yours. The worst businesses have low ROIC. Think about restaurants: you spend a fortune on equipment and build-out, you need constant reinvestment, and competition is everywhere. Most restaurants fail because the economics are brutal.

And according to Burry’s analysis, the ROIC of the major AI infrastructure companies is falling—fast. These companies are spending more and more capital (what Wall Street calls CapEx, short for capital expenditure) and generating less and less return on that spending. They’re becoming what Burry calls “capital-intensive hardware companies” when they used to be high-margin software businesses.

The Inversion of Software Economics

The thing that made Microsoft and Google valuable for decades was that they could generate enormous profits without enormous capital expenditure. Software scales beautifully. You write Windows once, you sell it a billion times. The marginal cost of each additional copy is essentially zero. Google’s search algorithm serves a billion queries a day, but the cost of each query is fractions of a penny. That was the magic of software.

Artificial Intelligence doesn’t work that way. AI requires constant compute power. Every query to ChatGPT runs on physical hardware somewhere. Every image generated by Midjourney burns electricity. Every Claude conversation you have is consuming resources in a data center that cost billions of dollars to build. The economics are inverted: software companies are turning into hardware companies. And hardware companies have fundamentally different, and worse, economics than software companies.

---

The Bookkeeping of Delusion

Now we get into the part that makes accountants nervous. In November 2025, Michael Burry published an analysis on his Substack with a provocative claim: the world’s largest technology companies are understating their depreciation expenses by a projected $176 billion between 2026 and 2028. That’s not a rounding error, that’s larger than the entire annual profit of most Fortune 500 companies.

How Depreciation Works (And How It Can Be Manipulated)

When a company buys a piece of equipment, it doesn’t record the full cost as an expense in the year it buys it. Instead, it spreads that cost over the “useful life” of the equipment. Imagine you buy a server for $40,000 and you think you’ll use it for four years. You don’t say “that’s a $40,000 expense this year” and dump it all into year one. Instead, you spread it out, you record $10,000 as an expense each year for four years. That expense is called depreciation. The idea is simple: the server is helping you make money over several years, so the cost should also be spread over those same years.

Here’s the critical part: depreciation is subtracted from revenue to calculate profit. Higher depreciation means lower profit on paper. Lower depreciation means higher profit on paper.

Now here’s where it gets sneaky. If a company decides “you know what, our servers actually last six years, not four,” then the math changes. Instead of $10,000 per year, they now record about $6,667 per year. Nothing about the real world changed, the server is the same, the technology is the same, only the accounting assumption changed. But suddenly, profits on paper go up by more than $3,000 per server per year. Now imagine a company that owns millions of servers. Change that one little assumption, and boom, their reported profits look way better.

That’s why Burry is obsessed with this. He’s not just looking at how much money a company says it makes, he’s looking at how they got those numbers, and what assumptions they quietly changed to make the profits look prettier.

What the Hyperscalers Have Done

In July 2022, Microsoft announced it was extending the useful life of its servers from four years to six years. The company’s CFO, Amy Hood, told investors on the earnings call that this would add $3.7 billion to operating income in fiscal year 2023. That’s billions of dollars in additional reported profit from an accounting change, not from selling more products, not from cutting costs, not from innovation. From simply deciding that servers last longer than they used to.

Google did the exact same thing. So did Meta and Oracle. Amazon did too, though it’s worth noting that Amazon later reversed course. Burry calls this “one of the most common frauds of the modern era.”

Extending useful life assumptions isn’t technically illegal. Accounting standards give companies discretion to make estimates about how long their equipment will last. If a company genuinely believes its servers can operate productively for six years, it’s allowed to depreciate them over six years. But Burry argues these estimates are disconnected from economic reality, and he has receipts.

The Chip Obsolescence Problem

Nvidia comes out with new chip designs about every year and a half to two years. The H100 chips that were super fancy and crazy expensive in 2023 are already being beaten by Nvidia’s new design, called Blackwell. According to Nvidia’s own marketing, Blackwell can run large language models up to 25 times cheaper and with 25 times less energy than H100. Twenty-five times. That’s not a small upgrade, that’s a whole new level. If Nvidia is telling the truth, then using H100s instead of Blackwell is like paying 25 times more on your electric bill just to do the same work.

Burry brings up another chip, the A100. These are only a few years old, but they already look bad next to the H100. He says A100s now use about two to three times more power for the same amount of computing as H100s. So they’re not just slower—they’re more expensive to run. Every hour they’re turned on, they’re wasting money compared to newer chips.

He compares it to airlines. Airlines keep old planes for busy times like Thanksgiving and Christmas. But even then, those old planes barely make any profit, and they’re not worth very much. Same idea with servers: an old server that still turns on is not the same as an old server that is smart to use. Just because something works doesn’t mean it’s making money.

So when companies say their chips last six years, Burry basically says “no way.” He thinks the economic life of these chips, meaning the time they make financial sense to use, is more like two or three years, not six. The accounting rules that say “six years” are out of sync with how fast the technology is moving.

He did the math: by 2028, he estimates Oracle’s earnings will be overstated by about 26.9 percent because of these depreciation assumptions, Meta’s earnings will be overstated by about 20.8 percent, and Microsoft avoided about $3.7 billion of expense in one year just by changing how long they say their servers “last” on paper.

The Amazon Exception and the WorldCom Parallel

The one exception, and hats off to them, is Amazon. In early 2025, instead of pretending its servers would last longer, Amazon actually did the opposite. They changed their minds, said “yeah, this stuff is getting old faster than we thought,” and shortened the time they claim their hardware is useful. Because of that, they took a hit of $920 million. They explained it by pointing to how fast AI and machine learning are moving—the tech is evolving so quickly that their old hardware becomes outdated much sooner than before. Amazon looked at the numbers and said “okay, this is not realistic anymore.” The real question is: why didn’t the other big companies do the same?

And this is where Burry brings up WorldCom. WorldCom was a huge telecom company that blew up in 2002 in what was, at the time, the biggest bankruptcy in US history. Their main accounting trick was taking normal day-to-day expenses, things that should have been counted as costs right away, and instead treating them like long-term investments. Why did they do that? Because it made their profits look better in the short term. Eventually the truth came out. The CEO went to prison. The company collapsed. Tens of thousands of people lost their jobs.

Burry isn’t saying these modern tech giants are doing the exact same thing or committing clear, illegal fraud. What he’s saying is that the pattern feels eerily familiar. They’re using aggressive, very optimistic accounting assumptions to make their profits look better during a time when they’re spending insane amounts of money on new hardware and data centers. And at some point, those assumptions will have to be corrected to match reality. When that happens, the adjustments won’t be pretty.

---

The Stranded Asset Problem

There’s a phrase in energy economics called “stranded assets.” Burry uses this term often, and it captures a specific kind of risk that most investors ignore until it’s too late. A stranded asset is infrastructure that becomes worthless, or worth far less than expected, before it’s fully depreciated. The asset still physically exists. It might even still function. But the economic value has evaporated.

Coal plants that have to shut down because of environmental regulations are stranded assets. Oil wells that become uneconomic because prices dropped are stranded assets. Power stations made obsolete by new technology are stranded assets. The common thread is that someone financed these assets assuming they would generate returns for decades, banks extended loans, investors bought bonds, shareholders expected dividends, and then something changed, and the expected returns never materialized.

Burry believes the AI infrastructure buildout is creating trillions of dollars in potential stranded assets.

The Math That Should Make You Uneasy

Right now, companies are buying mountains of Nvidia chips and loading them into gigantic data centers. They’re not thinking short-term, they’re planning like this stuff will pay off for decades. They’re financing these data centers on twenty-year timelines. They’re signing building leases that last thirty years. They’re making huge, long-term deals with power companies that assume these data centers will be sucking down electricity nonstop for ages.

All of those decisions sit on top of financial models that basically say “these buildings and these machines are going to generate profits for a generation.” That’s the story on paper. Now let’s compare that to reality.

The technology inside those buildings doesn’t live in twenty-year time, it lives in “a few months and it’s already old” time. Chip performance is moving so fast that every new generation makes the previous one look clunky and expensive. History doesn’t just suggest this might continue, history is screaming that it will.

Imagine it’s 2027, and a brand-new chip architecture shows up. It’s way more efficient, does the same work, but faster, and uses far less power. Based on how things have been going, that’s not some crazy sci-fi scenario, that’s almost guaranteed. The second that happens, a lot of those trillion-dollar data centers become partially obsolete. Not totally useless, the buildings still stand, the air conditioning still hums, the power lines still feed electricity in. But the actual compute, the brains of the operation, suddenly looks like a gas guzzler next to an electric car. It works, but it costs far more to operate than whatever the competition is using.

So on one side, you have the spreadsheet fantasy, where these assets are useful and profitable for twenty or thirty years. On the other side, you have the brutal tech reality, where the stuff inside might only be truly competitive for two or three years before it starts dragging you down. That gap, between the story the accounting tells and the story the technology tells is exactly where the risk is hiding.

Burry compares this to having an airplane that still flies but costs three times as much per mile as your competitor’s planes. Technically functional. Economically crippled. You can keep flying the old planes, but every route you fly, you lose money against competitors with newer equipment. Eventually you either replace the fleet or go bankrupt.

---

The China Power Advantage

There’s a geopolitical angle to this that should concern American strategists. China now has approximately 3,300 gigawatts of total installed electrical generation capacity. As of December 2024, China’s total installed power capacity was 3,349 gigawatts, according to their National Energy Administration. The United States has just over 1,100 gigawatts. China has roughly three times the power generation capacity of the United States, and the gap is widening.

In 2024 alone, China added 429 gigawatts of new capacity to their grid; that’s more than the entire installed base of many countries, a 21 percent year-over-year increase. Burry published a chart showing this divergence. In 2000, US and Chinese capacity were roughly similar. By 2024, China has pulled dramatically ahead.

But here’s what Burry emphasizes: it’s not just the total that matters, it’s the slope. China isn’t just ahead. China is accelerating. The United States is not. US transmission grid development is actually decelerating due to permitting issues, while China is building transmission at will to match its power output. Every new solar farm, every new data center in China can be connected to the grid with relative ease. In the United States, new transmission projects face years of environmental review and litigation.

Nvidia’s development roadmap is essentially a power consumption roadmap. Each generation of chips is more powerful but also hungrier. The Blackwell architecture requires massive cooling and electrical infrastructure. If AI superiority depends on running the most powerful chips at the largest scale, the United States is structurally disadvantaged.

Burry’s conclusion is blunt: “US companies are plowing capital into a race it is structurally positioned to lose.” He goes further: “Not only is China way ahead, it will continue to press its advantage. The slope is the key. That is acceleration of power generation development.”

The United States needs to either massively accelerate its power grid development, which would require regulatory reform that seems politically unlikely, or find a different path to AI capability that doesn’t depend on raw compute power. Burry argues for the latter. He thinks the US should be investing heavily in ASICs (Application-Specific Integrated Circuits, chips designed for one particular task rather than general-purpose computing) and in small language models that can run efficiently on limited hardware. But instead, he sees the United States doubling down on the brute force approach.

He even makes a policy suggestion, which is unusual for him. He says if he had the ear of political leaders, he would ask them to “take a trillion dollars and bypass all the protests and regulations and dot the whole country with small nuclear reactors, while also building a brand-new, state-of-the-art grid for everyone.” He calls this “the only hope of getting enough power to keep up with China, and it is the only hope we have as a country to grow enough to ultimately pay off our debt and guarantee long-term security, by not letting power be a limiting factor on our innovation.”

---

The Circular Economy of Hype

Now we arrive at what I consider the most disturbing part of Burry’s analysis. Bloomberg published a visualization in late 2025 titled “How Nvidia and OpenAI Fuel the AI Money Machine.” If you can find it, I encourage you to look at it. It looks like a web, arrows going in every direction, money flowing in circles.

Burry’s comment on the chart was blunt: “Every company listed below has suspicious revenue recognition. The actual chart with all the give-and-take deals would be unreadable. The future will regard this as a picture of fraud, not a flywheel. True end demand is ridiculously small. Almost all customers are funded by their dealers.”

Vendor-Financed Demand Explained

Nvidia is one of the most valuable companies on the planet because it sells the chips that power AI. OpenAI, which makes ChatGPT and other AI models, is valued at around $500 billion, making it the most valuable startup in history. Why? Because it runs those AI models on Nvidia’s chips. Microsoft has poured a ton of money into OpenAI. Microsoft also uses OpenAI’s tech inside its own products. OpenAI itself runs on Microsoft’s cloud service, called Azure. And Microsoft buys Nvidia chips to power Azure.

So you have this loop: Nvidia sells chips. Microsoft buys chips to run OpenAI. OpenAI runs on Microsoft. Microsoft uses OpenAI. But it gets even more circular. Nvidia has also invested money back into OpenAI. Nvidia has invested in other AI startups too. Those startups then use that money to buy more Nvidia chips. So the money is kind of spinning around in a circle inside the same group of companies.

This is what Burry calls “vendor-financed demand.” Let me make that super simple with a hot dog example. Imagine I start a hot dog stand. You invest $100 into my hot dog stand. I take your $100 and use it to buy hot dogs from your hot dog supply company. Your company now records $100 in revenue. Great, right? Then you take that $100 of revenue and decide “wow, this hot dog stand is doing amazing, I’ll buy more shares.” So you invest more money into my hot dog stand. I then use that new money to buy more hot dogs from you again.

On paper, it looks fantastic. My hot dog stand shows growing investment and growing activity. Your supply company shows growing revenue. Investors look at both of us and say “wow, these businesses are booming.” But here’s the punchline: no actual customers have bought a single hot dog. All the action is happening in a little feedback loop between you and me.

Burry is saying that a lot of AI revenue might look exactly like this. Company A invests in Company B. Company B uses that money to buy products from Company A. Company A reports higher revenue. Its stock price goes up. More investors get excited and pile in. Then more companies do similar circular deals. He even compares this directly to the dot-com era in the early 2000s, where one company would invest in another, and that second company would then use the money to buy the first company’s products, making both of them look more successful than they really were.

Then he throws out a challenge: “If you can name OpenAI’s auditor in one hour, you win some pride.” For a company worth half a trillion dollars, at the center of the entire AI boom, you’d think that basic info would be easy to find. But it’s not. He’s not saying “OpenAI is definitely committing fraud.” His point is different: it’s kind of crazy how little we actually know about the financial health and honesty of the key players in this whole AI story.

The Cascading Failure Risk

The big risk Burry is worried about has a name. Economists call it “cascading failure.” Think of it like a chain reaction. Nvidia depends on huge tech companies, called hyperscalers, buying tons of chips. Those hyperscalers depend on AI startups to keep using their cloud and paying big bills. The AI startups depend on venture capital investors to keep giving them money. Those investors depend on AI stock prices staying high, so their earlier investments keep looking good on paper.

So you have this chain: Nvidia → cloud giants → AI startups → investors → back to Nvidia and the whole AI market. If any link in that chain snaps, the stress spreads through the entire system. For example, if Nvidia’s chip sales slow down, their stock price drops. If their stock drops, the value of their investments in AI startups also drops. Those startups now look weaker and have less money to spend on cloud services. So cloud providers make less revenue from AI customers. If they make less, they buy fewer chips from Nvidia. And now the cycle turns negative—the whole thing that was spiraling up starts spiraling down.

Burry’s provocative framing is that this is not a flywheel. A flywheel generates momentum that sustains itself through real economic activity. This is more like a circular firing squad, where everyone is pointing at everyone else, and if anyone fires, they all go down together.

---

The Tragic Algebra of Stock Compensation

Michael Burry has been on a mission to expose what he calls “the tragic algebra” of stock-based compensation (SBC). He credits a former aerospace engineer named Phil Clifton for building the math behind it, the framework that shows just how badly regular shareholders are being diluted. Burry has this theory about who makes the best financial analysts: he thinks engineers are better than traditional Wall Street types, because engineers care about hard math and reality, not “creative” stories and accounting spin.

What Stock-Based Compensation Really Costs

A lot of tech companies don’t just pay employees in cash, they also pay them with stock. That’s called stock-based compensation. The idea sounds nice: if employees own part of the company, then they’ll want the company to do well. If the stock price goes up, everyone wins. On the income statement, SBC is called a “non-cash expense.” That means the company isn’t actually handing over cash out of the bank—instead, it’s handing over shares of the company. Because of that, companies love to say “oh, that doesn’t really count,” and they leave SBC out of their “adjusted earnings” to make profits look bigger and cleaner.

Here’s the problem Burry is shouting about: when you pay people in stock, you’re creating new shares. Those new shares don’t come from nowhere—they come out of the total pie of ownership. Think of the company like a pizza. You have eight slices and eight people, so everyone gets one slice. If you suddenly add two more people to the group, but the pizza is still the same size, now there are ten people sharing eight slices. Everybody’s piece just got smaller. That’s what dilution is.

Imagine a company reports $10 billion in earnings. Sounds huge. But in that same year, it gave employees $3 billion worth of new stock. In reality, existing shareholders only truly gained about $7 billion in value. The other $3 billion went to employees through dilution.

The Buyback Illusion

Companies know investors hate watching the share count go up. The number of shares out there is called the “float.” When the float grows, each share becomes a smaller slice of the company. Wall Street analysts watch this number like hawks. A rising share count is a warning sign. So what do companies do? They buy back stock. They take the company’s cash and go into the market and buy their own shares. This makes it look like the total number of shares is staying flat or even shrinking.

Here’s Burry’s key insight, the thing most people miss: those buybacks aren’t some generous gift to shareholders. They’re not really “returning cash to investors.” Most of the time, they’re just covering up the dilution caused by stock-based compensation. The company earns money. Then it uses a big chunk of that money to buy back shares. The share count stays flat. From the outside, it looks like nothing changed.

But something absolutely changed. The cash that could have gone to shareholders as dividends, or been reinvested to grow the business, or used to pay down debt, was used instead to cancel out all the new shares given to employees. The employees got paid. The shareholders basically got zero from that part of the earnings.

The Nvidia Numbers

Let’s talk about Nvidia, because Burry actually ran the numbers. Since the start of 2018, Nvidia has earned about $205 billion in net income and $188 billion in free cash flow. Those are monster numbers, Nvidia has been one of the most profitable companies on Earth. Over that same period, stock-based compensation was $20.5 billion. That’s a lot, but compared to over $200 billion in earnings, it doesn’t look insane.

But here comes the twist. During that same time, Nvidia spent $112.5 billion on stock buybacks. That’s more than half of all the earnings they made, spent on buying back their own shares. And after all that, after burning through $112.5 billion on buybacks, there are still 47 million more shares outstanding than when they started.

Read that again: Nvidia earned about $200 billion. They spent $112.5 billion buying back stock. And they still ended up with more shares floating around than seven years ago. From Burry’s point of view, the real cost of that dilution wasn’t $20.5 billion in SBC. The real cost was $112.5 billion in buybacks, because that’s what it took just to try to keep the share count under control. He says that when you factor this in, what he calls “owner’s earnings”—the money that truly belongs to shareholders, gets cut by around 50 percent.

The Burry-Karp Feud

Burry specifically targets Palantir and its CEO Alex Karp as an example of this dynamic at its most extreme. Palantir is a company with approximately $4 billion in annual revenue. From that $4 billion in revenue, Palantir has created five billionaires from its executive ranks. The company trades at over 100 times sales and over 200 times forward earnings.

When Burry disclosed a $912 million put position against Palantir, betting the stock would fall. Karp went on CNBC and publicly called the trade “batshit crazy.” (A put option is a contract that gives you the right to sell a stock at a predetermined price within a certain time frame. If you own a put option and the stock price falls below that predetermined price, you profit. The bigger the fall, the bigger the profit.)

Karp’s response on CNBC was heated: “The two companies he’s shorting are the ones making all the money, which is super weird. The idea that chips and ontology is what you want to short is batshit crazy. He’s actually putting a short on AI.”

Burry’s response to Karp was characteristically understated. He published the algebra. He showed how Palantir’s reported earnings are inflated by stock compensation accounting. He didn’t argue, he just presented the math. He also made a sarcastic jab at Karp, noting that while Karp has a PhD, it’s not in a scientific field, the implication being that Karp might not fully understand the “physics of finance” that Burry and his rocket scientist analyst are presenting.

---

The View from 40x Earnings

Let’s step back from the individual companies and look at the broader market context. Because Burry’s concerns aren’t limited to AI stocks—they extend to the entire market structure.

Apollo, one of the giant private equity firms, put out a chart in January 2026 about something called the Shiller Cyclically Adjusted Price to Earnings ratio, or the Shiller CAPE ratio. This metric is usually used to look at the overall stock market, not one single company and not the whole economy. Most of the time, it’s applied to something like the S&P 500. It’s a tool for asking: “Are stocks, as a group, cheap or expensive compared to their long-term earnings?”

Understanding CAPE

The normal PE ratio compares a stock’s current price to how much money that company earned over the last year. But company earnings jump around a lot depending on where we are in the business cycle. In a boom, profits are high, so PE ratios can look low even if stocks are actually very expensive. In a recession, profits fall, so PE ratios can look super high even if stocks might actually be cheap.

The Shiller CAPE ratio, named after economist Robert Shiller who won a Nobel Prize partly for this work, tries to fix that. Instead of using just one year of earnings, CAPE uses the average earnings from the past ten years, adjusted for inflation. By smoothing out the ups and downs of the business cycle, it gives a cleaner, more stable picture of whether the overall market is cheap or expensive compared to its true earning power over time.

Right now, the Shiller CAPE ratio for the U.S. stock market is around 40 to 41. To understand what that means, we need to compare it to history. The only time in recorded history when the CAPE ratio was clearly higher was during the dot-com bubble around the year 2000, when it peaked at about 44. After that, tech stocks crashed—the NASDAQ fell roughly 80 percent from its peak, and trillions of dollars in paper wealth disappeared.

Today’s level is also higher than it was before the 1929 crash, when the CAPE was about 32.5. That crash led into the Great Depression. If you go all the way back to the 1870s, the long-term average CAPE ratio is around 17. Most of the time, it sits somewhere between 15 and 20.

So at a CAPE of around 40, the U.S. stock market is trading at roughly 135 percent above its historical average—in other words, more than double what has usually been considered “normal” pricing. The Shiller CAPE has a strong historical relationship with future returns. When the CAPE is low, the next 10 years of returns tend to be pretty good. When the CAPE is high, the next 10 years tend to be weak. At today’s levels, the historical data suggests that people buying broad stock market indexes right now should expect very low, possibly near zero, average yearly returns over the next decade.

Households More Exposed Than Ever

A common objection is that the market has changed. Today’s companies are more profitable. Technology is more valuable. The rules are different. Burry has heard this before. He heard it in 2000. He heard it in 2007. The specifics change. The argument is always the same: “This time it’s different.” Sometimes it is different. Mostly it’s not.

There’s another chart that Burry highlighted that I find haunting: for the first time in American history, households now have more of their net worth in stocks than in real estate. Think about what that represents. For generations, the family home was the foundation of middle-class wealth. You bought a house, you paid off the mortgage over 30 years, and by retirement you owned a valuable asset free and clear. That was the American dream. Now stocks are. The proportion of household wealth tied to equity market valuations is at an all-time high.

Burry puts this in context: “After nearly a decade of zero interest rates, trillions in pandemic helicopter cash money, the greatest inflation in 50 years, and a new paradigm of higher Treasury rates for the first time in 50 years, stocks have emerged victorious even over home prices that rose 50 percent.” He attributes this partly to “the gamification of stock trading, the nation’s gambling problem due to its own gamification, and a new AI paradigm backed by trillions of ongoing planned capital investment backed by our richest companies and the political establishment.”

Then he asks his signature question: “What could go wrong?”

The answer is: quite a lot. If the market drops 40 percent, which has happened multiple times in living memory, the wealth destruction will be unprecedented in its reach. More households are exposed to stock market risk than ever before. The losses won’t be concentrated among professional investors. They will hit retirement accounts, college savings, and household balance sheets across America.

---

The Passive Investing Time Bomb

Burry has written extensively about what he calls the “passive investing bubble.” This is a structural issue in how markets function that amplifies both gains and losses.

Traditional investing, what’s called “active” investing, involves analyzing individual companies and deciding which ones to buy based on their fundamentals. Is the company well-managed? Are the products competitive? Is the stock price reasonable relative to the earnings? Passive investing, by contrast, involves buying index funds that automatically own every stock in a particular index, like the S&P 500. You’re not analyzing companies, you’re just buying “the market.”

Passive investing has grown enormously over the past two decades. Index funds now control a massive portion of the stock market—by some estimates, passive strategies now account for more than half of all equity assets under management. The appeal is obvious: index funds are cheap, they’re simple, and historically, most active managers have failed to beat the index after fees. Why pay someone to pick stocks when you can just buy the whole market for almost nothing?

The Structural Problem

But Burry sees a structural problem. Index funds buy stocks mechanically, without regard to valuation, based solely on market capitalization. The bigger a company gets, the more weight it has in the index, the more the index funds have to buy it. This creates what Burry describes as a “liquidity trap” that works beautifully on the way up.

Money flows into index funds. The funds have to buy stocks in proportion to their index weight. The buying pushes prices up. Higher prices mean higher market caps. Higher market caps mean more index weight. More index weight means more buying when new money flows in. It’s a self-reinforcing cycle. The biggest stocks get bigger simply because they’re biggest.

On the way down, the mechanism reverses. Money flows out of index funds. The funds have to sell stocks in proportion to their index weight. The selling pushes prices down. Lower prices mean lower market caps. Lower market caps mean less index weight. When the next wave of selling comes, the impact is concentrated.

And here’s the key insight: because passive funds own such a large share of the market, there are fewer active buyers to step in and provide a floor. In traditional markets, when stocks fell to attractive valuations, value investors would step in and buy. Their buying provided support, it limited how far prices could fall. But if half the market is held by passive funds that sell mechanically regardless of valuation, there’s no one waiting to buy. The selling begets more selling. And the selling doesn’t stop until something external changes.

Burry wrote: “When the music stops in the age of passive investing, there will be no one there to hold up valuations.” He first raised this concern years ago, and so far it hasn’t played out in the catastrophic way he described. But the structural issue remains. The market has never experienced a prolonged downturn with passive strategies at their current scale. No one knows exactly how it will behave. Burry believes it will behave badly.

---

The Dominoes Start to Fall

Let me trace the potential sequence of events that keeps Michael Burry awake at night. This is not a prediction, but it’s a scenario grounded in the specific vulnerabilities he has identified.

First domino: The depreciation chickens come home to roost. Companies that extended useful life assumptions from four years to six years will eventually have to reconcile those assumptions with reality. When they replace their obsolete hardware, they can’t keep pretending the old hardware still has value. They’ll take massive accelerated depreciation charges. Reported earnings that looked robust will suddenly look hollow. Analyst downgrades follow. Stocks sell off.

Second domino: The revenue circularity becomes undeniable. If a meaningful portion of reported AI revenue is companies investing in each other and buying each other’s products with that investment, then “real” end-user demand—from actual customers paying actual money—is far smaller than headline numbers suggest. When growth rates decelerate and investors demand to see organic revenue, the narrative fractures.

Third domino: Return on Invested Capital forces a reckoning. Companies that have been celebrated for “investing in the future” will be judged on whether those investments generate returns. If ROIC continues to decline as Burry expects, the market will reprice these stocks from “growth” multiples to “value” multiples. That repricing is violent. A stock trading at 30 times earnings that gets repriced to 15 times earnings loses half its value even if profits stay flat.

Fourth domino: The stranded asset problem materializes. Data centers built with 20-year financing assumptions become economically impaired after five years. The private credit markets that financed this buildout face losses. Credit contracts. Capital becomes harder to obtain.

Fifth domino: The index funds become forced sellers. When the stocks at the center of the AI trade decline, their market cap declines. Their weight in the indices declines. Index funds have to sell to maintain their target allocations. But because so many assets are in passive strategies, the selling overwhelms the buying. The selling begets more selling.

Sixth domino: Household wealth collapses. With more net worth in stocks than ever before, a market decline of 30 or 40 percent would destroy trillions in household wealth. Consumer spending contracts. Retail sales fall. Corporate earnings fall. The recession becomes self-fulfilling.

Seventh domino: The banking system feels the stress. Burry has noted that US banks are already fragile. Bank reserves at the Federal Reserve have grown from $45 billion in 2007 to over $3 trillion today. He points out that the temporary bank support programs created during the 2023 banking crisis “became load-bearing.” If banks face loan losses from AI-related credit exposure, if commercial real estate values fall because tech companies scale back office space, if consumers default on credit cards because their stock portfolios collapsed, the stress propagates through the financial system.

This is the cascade that Burry sees as possible. Not certain. Possible. The same man who saw the housing collapse coming when everyone else thought he was crazy now sees similar patterns in artificial intelligence.

---

The Ghosts of Bubbles Past

On May 29, 1969, Warren Buffett wrote a letter to his investment partners. He was closing his partnership. He couldn’t find investments that met his criteria. Stocks were too expensive. He was returning everyone’s capital. “I am not attuned to this market environment, and I don’t want to spoil a decent record by trying to play a game I don’t understand just so I can go out a hero.”

A year later, the S&P 500 was down 30 percent. A decade later, adjusted for inflation, the market had lost 57 percent of its real purchasing power. Not nominal, but real. After accounting for inflation, investors who bought at the 1969 peak had less than half their money ten years later. Burry published the first page of that Buffett letter recently. He didn’t comment on it. He just posted it.

The parallel isn’t perfect, no parallel ever is. But Burry closed Scion Asset Management to outside capital in November 2025. He handed the operation to his protégé Phil Clifton, who is launching a new venture called Pomerium Capital. Burry has moved to a “purely advisory and commentary role” through his Substack. He’s not managing other people’s money in a market he doesn’t trust.

---

What The Man Who Called It Sees Now

I want to be clear about something. Michael Burry has been wrong before. He was heavily short the market in 2023 and missed a massive rally. He has made calls that didn’t pan out. Predicting market timing is notoriously difficult even for the best investors. He himself would tell you that being early is indistinguishable from being wrong, at least in the short term. Considering the cyclical nature of the market, many comments online poke fun at him saying eventually he will be right.

But his methodology deserves attention. While most analysts build models based on management guidance and growth projections, Burry reads the footnotes. He looks at depreciation schedules that most investors never examine. He calculates true shareholder dilution by tracking stock compensation and buybacks over years. He traces circular money flows through investment and revenue relationships. He compares current conditions to historical precedents that ended badly. He’s not predicting the future based on optimism or pessimism. He’s doing arithmetic.

And the arithmetic, as he sees it, suggests that much of what’s being reported as profit in the AI sector is not actually profit—it’s accounting choice. It’s depreciation manipulation, extending useful lives from four years to six when the economic reality is closer to two or three. It’s stock compensation excluded from adjusted earnings, creating the illusion of profitability while diluting shareholders. It’s circular revenue between related parties, money flowing in loops that inflate top-line growth without reflecting real end-user demand.

When you strip away the accounting flexibility, many of today’s AI companies might actually be losing money. The “Adjusted EBITDA” that companies love to highlight excludes all the costs that Burry thinks matter most. He draws an analogy to his early work in medicine: “When I was a medical student, I realized the business of healthcare is predictable when you follow the money.” The same principle applies to AI. Follow the money. See where it actually goes. Don’t be distracted by the narrative.

Burry’s 2026 Tipping Points

Burry believes “the algebra will inevitably catch up to the narrative in 2026.” That’s not a prediction in the sense of “the market will crash on this date.” It’s a warning. The gap between reported performance and economic reality is widening. Eventually, something closes that gap. Either the reality improves to match the reporting, or the reporting adjusts to match the reality.

He compares Nvidia to Cisco at the peak of the dot-com bubble. Cisco was the “picks and shovels” play of the internet revolution. Everyone needed routers. Everyone needed networking equipment. The thesis was unassailable, until it wasn’t. Cisco’s stock fell 80 percent from its 2000 peak and took 15 years to recover even to half that level. The comparison isn’t perfect. Nvidia is more profitable than Cisco was. The AI applications are arguably more substantial than the dot-com applications. But the structural point stands: being the dominant supplier during a capital spending frenzy doesn’t protect you when the frenzy ends.

In his analysis, Burry pointed to several potential tipping points for 2026. Financial tipping point: If AI revenue reaches $500 billion and proves to be real, durable revenue from actual end customers paying actual money, the bubble narrative might break. The fundamentals might catch up to the valuations. The bears would be proven wrong. If the revenue proves to be circular and inflated, the narrative breaks in the other direction.

Technical tipping point: If scaling hits a wall, if simply adding more compute no longer produces meaningful intelligence improvements, the entire investment thesis collapses. There are signs this might be happening. Despite massive investment, no single AI lab has maintained a runaway lead. Google, OpenAI, and Anthropic keep rotating on the podium. Someone pulls ahead, then the others catch up.

Infrastructure tipping point: If the market decides that application-specific chips and small language models are the future rather than giant GPU clusters, Nvidia’s dominance evaporates. Burry writes that he’s “surprised Nvidia has remained the dominant chip for inference this long.” He expects ASICs to eventually replace general GPUs for most production AI workloads.

---

Cassandra Unchained

The question for readers is not whether to bet with Burry or against him. I’m not giving investment advice. This is not a recommendation to short AI stocks or buy puts on Nvidia. The question is whether you’ve actually looked at the numbers yourself. Whether you understand what you own and why you own it. Whether you’re investing based on analysis or based on the assumption that prices will keep going up because they’ve been going up.

Because that’s not analysis. That’s faith. And faith is a wonderful thing, but it’s a terrible investment thesis.

Burry named his Substack “Cassandra Unchained.” In Greek mythology, Cassandra was the daughter of Priam, King of Troy. Apollo gave her the gift of prophecy. But when she rejected his advances, he cursed her. She could see the future accurately, but no one would ever believe her. Cassandra warned the Trojans about the wooden horse. They didn’t listen. Troy burned.

Michael Burry has been playing Cassandra for two decades now. Sometimes he’s early. Sometimes he’s wrong entirely. But when he’s right, he’s catastrophically right. The people who dismissed him on housing learned that lesson the hard way.

The trillions of dollars flowing into AI infrastructure, the accounting tricks inflating reported earnings, the circular money flows masking anemic end demand, the stranded asset risk hiding in 20-year financing assumptions, the stock compensation diluting shareholders while buybacks create the illusion of capital return, all of it is documented. All of it is verifiable. All of it is being ignored by a market that has decided artificial intelligence is the future and the future cannot fail.

Maybe it can’t fail. Maybe this time is different. Maybe the algebra doesn’t apply.

You now have the tools to follow Michael Burry’s analysis yourself. You understand ROIC and why it matters. You understand depreciation and how companies manipulate it. You understand stock-based compensation and the “tragic algebra” of dilution. You understand the Shiller CAPE and what it says about market valuations. You understand vendor-financed demand and circular revenue flows.

Whether you use those tools is up to you. Whether you agree with Burry’s conclusions is up to you. But at least you can read his work and evaluate it for yourself.

The files remain gray because the truth is never simple. But sometimes, if you do the math, it becomes clearer than anyone wants it to be.

---

Listen to the full podcast episode for additional context and audio experience. Subscribe to The Gray Files for more deep dives into technology, economics, data science, and the forces shaping our world.