Big short Burry's latest in-depth article: Why I am bearish on everything

As the saying goes, the end of an investor is an internet celebrity, “questioning Wang Cen, understanding Wang Cen, becoming Wang Cen.”

I didn't expect that the big short seller Michael Burry would also take this path and start knowledge monetization.

A joke: How the big short Michael Burry made a 1.5x return in 3 months by “shorting” Nvidia:

1. Buy $8 million worth of put options on Nvidia;

2. Announce the 13F in advance and widely promote their short positions on the internet, sparking heated discussions in the market;

3. Start writing articles for paid subscriptions during periods of high market volatility, priced at $379/year;

4. Instant 80,000 subscribers, accepting a monthly subscription income of 3.2 million USD / annual income of 30 million USD.

In order to justify the $30 million subscription fee, Burry recently published two articles, and his core point is summed up in one sentence: This time is no different, it’s just a bubble.

He wants to prove that even in 2000, the ones driving the bubble were still the “rock star” companies that made huge profits back then.

In response to Burry's short-selling (he has repeatedly emphasized that “NVIDIA is like Cisco”), NVIDIA also retaliated.

Last weekend, NVIDIA's IR team sent a seven-page internal memo to Wall Street analysts, directly addressing a dozen concerns (including GPU depreciation and revolving debt financing).

NVIDIA responded by stating: Regarding the claim that “the current situation is similar to historical accounting fraud cases involving vendor financing and special purpose entities (SPVs) (such as Enron, WorldCom, and Lucent),” NVIDIA's situation is fundamentally different from historical accounting fraud cases because NVIDIA's core business is economically sound, our reports are complete and transparent, and we place great importance on our integrity and reputation. Unlike Enron, NVIDIA does not use special purpose entities to hide debt and inflate revenue.

Regarding the GPU depreciation issue, some companies use a six-year depreciation method for GPUs, but Burry believes that the actual lifespan of the chips is shorter than six years. Nvidia's customers inflate profits by spreading the depreciation costs over a longer time period.

NVIDIA stated in a memorandum: “Based on actual lifespan and usage patterns, NVIDIA's customers set the GPU depreciation cycle to 4-6 years. Older GPUs like the A100 (released in 2020) still maintain a high utilization rate and can generate significant profits, with their economic value far exceeding the 2-3 years some commentators claim.”

What do you think of Burry's short-selling viewpoint? Next, I will translate his first article published on November 24: “The Cardinal Sign of a Bubble: Supply-Side Gluttony (The Cardinal Sign of a Bubble: Supply-Side Gluttony), The Heretical Guide to AI Stars (Part 1):”

01. The Primary Characteristic of Bubbles: Supply-Side Greed

Folly (Folly)

Foolish actions can bring wealth. Creative destruction and crazy foolishness are precisely why the United States has become the global center of innovation. Here, companies are allowed to walk towards extinction through innovation.

And more companies will emerge like mushrooms after rain, repeating the same mistakes. Sometimes, a new company is just a transformation of an old one; sometimes, the current industry leader actually had a nearly identical “doppelgänger” in its early years, but that doppelgänger's product came out too early.

Our bankruptcy law, contract law, property law, and case law have withstood the test, protecting this innovative impulse while leaving other countries far behind.

This is glorious. It has nothing to do with reflecting thoughtfully in a restrained and mature manner. If the kids in Silicon Valley did that, foolishness would never happen, and that would be a tragedy.

So, foolish actions inevitably go too far. The result is—companies go bankrupt in batches. Investors permanently lose capital, and employees lose their jobs.

Today, we have not yet reached that point. However, to understand today's “glorious folly” and the corresponding strategies, it is necessary to break it down into different parts for analysis and investigation. This cannot be covered in just one article or post.

We will start with a brief review of the last farce of this scale—the so-called “dot-com bubble” of the 1990s. Many people misremember it or have no idea what it was about.

Our Federal Reserve Chairman is one of them, and he has become a victim of the most prevalent false narratives today.

“These companies (today) … actually have business models and profits, so this really isn't a thing.”

—— Jerome Powell, October 29, 2025

I cannot agree.

The slogan during that frenzy was “Internet traffic doubles every 100 days.”

That has always been a data transmission infrastructure bubble, rather than a “.com” bubble about internet content.

Of course, that sock puppet (the mascot of Pets.com) is indeed very interesting.

Pets.com is one of the most astonishing failures in Silicon Valley's history. But before it collapsed spectacularly, they dazzled audiences with a series of charming sock puppet commercials. These commercials rank among our selection of the best Super Bowl ads of all time.

So, perhaps there are indeed small bubbles contained within the big bubble. There were many similar cases in the 1920s.

However, rather than being driven by unprofitable internet companies, it is more accurate to say that the powerful Nasdaq, which surged ahead in 1999 and entered the new century, was driven by high-profit blue-chip stocks. This includes the “Four Horsemen” of that era: Microsoft, Intel, Dell, and Cisco.

The Nasdaq is a market capitalization-weighted index, so the largest market capitalizations contribute the majority of the performance, just like today. A highly profitable company - Qualcomm saw an increase of 2619% in 1999, with a year-end market capitalization of $56 billion.

It's not an isolated case. Applied Materials rose by 198%. Amgen rose by 147%. Cisco rose by 125%. Microsoft rose by 68%. Oracle rose by 309%.

Sun Microsystems reached a market value of $135 billion by the end of 1999, a rise of 244%, but this was based on only about $1 billion in net profit. Lam Research's stock rose over 500% in 1999, and it was certainly not a flash-in-the-pan company. Comcast rose by 66%. Qwest Communication also soared.

All of these are trustworthy, real big companies.

The profitless companies that today's bulls might think of—such as VA Linux, Webvan, Ask Jeeves—didn’t go public until the fourth quarter of 1999. Other favored loss-makers—Pets.com and webMethods—missed the 90s entirely, going public only in the first quarter of 2000. None of them were major contributors to the historic rise of Nasdaq.

Therefore, the popular history regarding the internet bubble as the “no-profit company bubble” is unfortunately mostly fake news when it comes to the peak and collapse of Nasdaq. The vast majority of funds led by large-cap profitable blue-chip stocks first rose all the way up and then all the way down.

Even Gemini and other large language models got this wrong. In fact, I have tried to educate them, and they eventually admit they were wrong, but then they make mistakes again in subsequent questions. This touches on an issue with LLMs and generative AI (GenAI), which I will write about in another article.

It is not the point to say that some companies are not profitable. The most important thing is that, just like today, the capital markets once enthusiastically supported those zealous advocates of data transmission construction.

• AT&T spends $20 billion a year on data hosting, long-distance transportation, and other assets.
• MCI spent 15 billion dollars.
• Sprint burned hundreds of billions of dollars on its PCS wireless business, which actually required a smartphone that had not yet been invented.
• BellSouth and SBC (now merged into Verizon) also spent hundreds of billions of dollars upgrading their service areas to provide high-speed internet.
• Global Crossing spent $20 billion to lay underwater cables.
• Level 3 cost 20 billion dollars.
• Competitive Local Exchange Carriers (CLECs, such as Adelphia, Covad, XO Communications, etc.) spent $30 billion.
• Williams Companies (yes, that natural gas company) spent billions of dollars to align its data transmission department with its natural gas pipelines.
• 360 Networks and submarine cable combinations like Flag and the TAT-14 consortium, each costing hundreds of billions of dollars.

More fiber optics require more routers, and more routers mean more fiber optics, creating a positive infinite feedback loop. All of this is for that seemingly impossible task: transmitting data to every household around the globe at an ever-increasing speed, a task that seems endless.

“The tasks seem endless.” Doesn't that sound familiar?

I elaborate on these points to ensure that readers understand — this is not merely a “dot-com bubble” phenomenon, nor is it an issue of insufficient capital market support or cash flow.

No, this is a catastrophic oversupply and severe lack of demand issue. Especially in the United States, the adoption rate of broadband connections is very slow.

This is not something like “this time is different,” even though many people try to make it seem different.

First of all, we should pay attention to other significant frenzied booms since 2000. Notably, the real estate bubble of the 2000s and the shale oil revolution of the 2010s. I have a lot to say about the real estate boom, but that's a topic for another time, as it's not very relevant here.

Below, I have created a chart that divides the total net capital investment of the S&P 500 index (capital expenditure minus depreciation) by nominal Gross Domestic Product (GDP), and overlays the related stock market peaks.

I suggest you enlarge the chart and study it carefully before continuing to read.

Net Investment ( S&P 500 Capital Expenditures minus Depreciation ) / Nominal GDP, 1991-2025

The chart marks: the Nasdaq peak in March 2000; the S&P 500 peak in November 2007; the S&P Energy Index peak in July 2014; and the Nasdaq 100 record high in the fourth quarter of 2025.

This chart contains a lot of information.

Net capital investment is total capital expenditure minus depreciation, and I divide net investment by nominal GDP for equivalent comparisons across different historical periods. The data is plotted in histogram form; the blue bars represent the level of net capital investment for that quarter. The data ranges from 1991 to 2025.

The colored vertical lines from top to bottom mark the speculative peaks of stocks that are typically associated with investment frenzies. The orange line on the right represents the recent new high of the Nasdaq 100 index.

The key point is that the pattern you see is: the stock market peak usually occurs midway through an investment frenzy, and in some cases, such as in 2000, it even happens before capital expenditures peak.

This is a visual representation of the peak of frenzy — investors and speculators are wildly cheering for the CEOs' massive multi-year spending plans, awarding market value rewards of $2, $3, or even more for every dollar spent. Shocking stock-based compensation ensures that this incentive resonates deeply, even long after the market has peaked. Hope always exists, especially during times of bubbles.

Today, equity incentives are more shocking than 25 years ago. Let's discuss it in detail another day.

In the past year, Meta, Google, Oracle, and other hyperscalers have found that massive spending plans are the surefire way to boost stock prices. Even Alibaba has joined the game, announcing a large-scale AI data center spending plan that finally lifted its stock price out of nearly four years of stagnation.

Edited by Edward Chancellor, the book Capital Account: A Fund Manager Reports on a Turbulent Decade, 1993-2002 details the internal and other communications of Marathon Asset Management during the bubble period of the 1990s.

Marathon refers to the capital over-investment issue I described above as a product of the “Capital Cycle Theory.”

I only discovered this book recently, but I found that CCT provides a solid framework for analyzing the capital investment frenzy driven by mania. Additionally, this book is also an excellent firsthand account of the true history of the booms and busts of the 1990s.

This book has a retail price of over $1000, but you might find a free PDF version online. This is a page from my book.

This section cites the contents of the book regarding the oversupply and price collapse in the telecommunications industry, mentioning that until 2002, the utilization rate of optical fibers was less than 5%.

What I emphasize is an astonishing fact: by 2002, during the fervent construction of data infrastructure in the bubble, less than 5% was actually lit (put into use). This represents how brutally management and investors misjudged the future ahead. CEOs were too overly confident in the stock market.

I have never been in that position, but it seems to be a particularly unavoidable part of human nature throughout history.

I can personally tell you how the market peak on March 10, 2000, occurred. In other words, there was no obvious reason for its occurrence.

The main capital expenditures for fiber optics are still in planning and execution. The demand for network equipment remains high. Cisco's revenue in 2000 is expected to grow by 55%. Fiber optics require routers. Routers empower fiber optics. However, the stocks are declining.

The Nasdaq index has not returned to that high point for more than 16 years since then.

However, that top is really inconspicuous.

As we moved into the year 2000, shortages of components and capacity constraints became the norm. Demand seemed strong, but this was just technology companies selling to technology companies, all under the guise of data transmission expansion.

“We have not seen any signs that radical internet business transformation… is slowing down—in fact, we believe it is accelerating globally.”

—— Cisco CEO: Q4 2000 Earnings Report Released

“…Due to the capacity constraints across the industry and the unprecedented demand for our communication products, we experienced a shortage of components.”

—— Agilent CEO: November 2000

“Based on the momentum and strong order backlog we experienced in the first nine months, we continue to anticipate that the growth rate in 2000 will be around 40% compared to 1999.”

—— Nortel Networks: October 2000 Press Release

“Cisco is fortunate to be at the center of an economic revolution that is reshaping not only the economy but also various aspects of society.”

—— Cisco CEO: Press Release on September 24, 2000

“We have not seen any signs of slowing down. We have provided accurate guidance to Wall Street, and we can execute as planned.”

—— Cisco Chief Strategy Officer: November 3, 2000

Later, the collapse of Enron in the summer of 2001 shocked many people. Cisco's revenue grew by 17% in 2001, but it began to incur losses.

At the end of 2001, the Nasdaq had fallen an astonishing 62% from its peak, while Cisco dropped 78%. Investors who waited for signals of a slowdown before selling suffered significant losses.

AI stock market bulls almost unanimously declare that large-scale spending will continue today and in the foreseeable future, leading to a sustained rise in the stock market over the coming months and even years. I have heard that we are currently in 1997.

But the historical record tells a different story. The real estate bubble and the shale oil revolution have similar dynamics and are no exception.

The five major listed tech giants “knights” in today’s AI frenzy—Microsoft, Google, Meta, Amazon, and Oracle—along with several startups in their adolescence, promise to invest nearly $3 trillion in AI infrastructure over the next three years. Investors are absolutely delighted about this.

OpenAI, the private owner of the widely used generative AI platform ChatGPT, has committed to an astonishing $1.4 trillion expenditure over the next 8 years. Its revenue is less than 2% of this figure, while its losses exceed 2% of this figure. Its CEO Sam Altman was once asked about this matter:

“The reason why listings are attractive is because, in very few cases, when those people write absurd posts like 'OpenAI is about to go bankrupt', I really want to tell them that they could short the stock, and I would love to see them suffer losses because of it.”

A CEO of a struggling startup has been practicing his tirade against short sellers. This says it all, really. It has the support of the capital markets, with a valuation reaching $500 billion, and it hasn't even gone public yet.

$500 billion surpassed the total market capitalization of all publicly listed unprofitable internet and telecom companies in the 1990s. In fact, $500 billion was exactly the market capitalization that Cisco reached in March 2000.

In any case, today's five major listed tech giants “Knights” are interesting in many ways.

It turns out that these big companies that spend money like water have been extending the useful depreciation period of these assets while increasing spending plans for the same chips/servers. Old-fashioned vendor financing from the 90s is back, just in a new guise.

Once again, there is a “Cisco” at the center of it all, providing “shovels” for everyone, accompanied by a grand vision. Its name is Nvidia.

“There is a lot of discussion about the AI bubble. From our perspective, we see something completely different. Just a reminder, NVIDIA is unlike any other accelerator. We excel at every stage of AI, from pre-training, post-training to inference. With our 20 years of investment in the CUDA-X acceleration libraries, we also perform exceptionally well in scientific and engineering simulations, computer graphics, structured data processing to classical machine learning. Choosing NVIDIA is because our singular architecture enables all three of these transformations. Therefore, for any form and modality of AI across all industries, across every stage of AI, across the diverse computing needs in the cloud, and from the cloud to enterprise to robotics, it is a single architecture.”

——NVIDIA CEO Jen-Hsun Huang, November 19, 2025

In this article, I clarified some relevant history, laying the foundation for the upcoming content. I hope this is helpful to you.

I introduced the Capital Cycle Theory and demonstrated through charts that the stock market peak often occurs right in the middle of a frenzy of investment enthusiasm, which is a clear pattern.

The investigation will continue in the second part, where we will closely examine the accounting issues involving the biggest players here—something that is common to any “glorious folly.”

I will conclude with an obscure quote from a famous person. This has always been one of my favorites.

“If you go around popping balloons, you won’t be the most popular person in the room.”

Charlie Munger, R.I.P.

02, A Deep Post in the Comment Section

Of course, the audience who spent 379 dollars are also hidden talents, not ordinary people. I feel that a comment I came across is quite inspiring, so I will translate it along the way:

The Economics of AI Infrastructure: Depreciation (including Excel Depreciation Model)

Every major economic boom is inseparable from physical infrastructure. Recent observations by Barry have prompted me to delve deeper into the similarities and differences in the depreciation of artificial intelligence infrastructure compared to past eras of large-scale capital construction. This article will compare the eras of railroads, the internet bubble, and artificial intelligence from the perspective of depreciation.

Railway Era

The railway boom period of the 19th and early 20th centuries was characterized by huge upfront investments in civil engineering and an extremely long economic lifespan. The tracks, bridges, and tunnels built with equity and long-term debt financing can last for decades (similar to today's renewable energy capacity construction). The economic lifespan of these assets ranges from 25 to 100 years, in other words, their effective depreciation rate is about 1% to 4% per year. Railroads exhibit the characteristics of a monopoly utility: high upfront capital expenditure, low annual depreciation, and predictable demand, resulting in long-term cash flows. By investing heavily now, one can benefit for decades.

Internet Bubble Era

Compared to the railway era, the assets in the telecommunications era are more complex, as the economic lifespan of certain assets is longer than others (for example, the lifespan of fiber optics and pipelines is 30-40 years, while network equipment has a lifespan of 3-5 years). Across the telecommunications network, this amounts to an annual depreciation rate of about 5-10%. Even after the economic downturn in 2000, fiber optics and pipelines remained useful for many years afterward, with some still in operation today. In this context, the issue of capital expenditure obsolescence does exist, but it remains within a controllable range.

Artificial Intelligence Era

The composition of capital expenditure on artificial intelligence is fundamentally different from previous infrastructure cycles. Nowadays, a significant portion of investment is directed towards assets with a shorter economic lifespan, particularly GPU clusters and bandwidth-intensive networks. The accompanying Excel model allows you to adjust these assumptions, but the basic situation has clearly demonstrated this dynamic change.

Assuming the economic lifespan of cutting-edge GPUs is three years, about 78% of the economic capacity of an artificial intelligence data center will be exhausted within the first three years (note that it is the economic lifespan, not the usable lifespan). In fact, if you invest 100 dollars in cutting-edge artificial intelligence computing, only 22 dollars of economically competitive capacity will remain by the third year. The remaining value loss is not due to physical performance degradation, but rather due to technological obsolescence.

This has created a reinvestment obligation that is different from any previous infrastructure boom period (as follows), where the value generated by each generation of hardware must be large enough to offset the rapid and recurring depreciation cycles.

Depreciation Policy

In the past five years, the depreciation policy for GPU servers has been inconsistent. Microsoft set the depreciation period for servers and network equipment at 4 years before the fiscal year 2023, but now assumes an economic lifespan of 6 years. Alphabet has adopted the same approach. Meta extended the lifespan of its servers to 5.5 years in January 2025, which is expected to reduce depreciation expenses by $2.9 billion.

AWS has been phasing out some assets in advance and shortening their useful life from 6 years to 5 years. If the economic life of an asset is 2-3 years while its accounting life is 5-6 years, it can lead to: a) overestimated early profits due to lower depreciation; b) accelerated depreciation in later years, increasing capital expenditure pressure, thereby dragging down cash flow. The effect of extending the depreciation policy is to reduce annual depreciation expenses, increase short-term profits, and smooth out the impact of large-scale AI spending.

Can these assets really perform well in the long term?

Yes - but it does not apply to cutting-edge artificial intelligence development. Older GPUs are still useful for inferring small models and dealing with demand peaks. However, when their unit price loses competitiveness relative to their power consumption and infrastructure costs, their economic value declines significantly. In fact, AI infrastructure is evolving into two economic levels: training capital expenditures (which require the latest chips and architectures) and inference capital expenditures (which can use older chips, but have longer lifespans and are more sensitive to model optimization). This dual structure means that the rapid updates of hardware and the constantly changing demands of models will both bring depreciation risks.

What does this mean for investors?

A) Because hyperscale data center operators may find themselves in intense competition, they need to continuously increase capital expenditures to maintain competitiveness. This can lead to a structurally increased burden of ongoing capital expenditures, which needs to be explicitly reflected in the modeling and valuation assumptions of hyperscale data center operators.

B) The accounting profit has plummeted in the short term, requiring investors to assess whether the capital expenditures made can generate sufficient economic returns. This requires an in-depth understanding of how artificial intelligence will be integrated into existing products or new products.

C) The industry economy may have undergone fundamental changes (i.e., capital expenditures remain at a consistently high level).

Conclusion

When investors forecast free cash flow, they should pay attention to the implied reinvestment rate and the reasonableness of the assumptions regarding returns on artificial intelligence capital expenditures.

The core question is no longer just “How great are the opportunities of artificial intelligence?” but rather “Can this business sustain the necessary reinvestment to maintain its leading position?”

Companies with strong balance sheets, product integration options, and pricing power that can transform a three-year asset cycle into lasting economic value will emerge as the ultimate winners. Meanwhile, the losers will realize too late that in the field of artificial intelligence, depreciation is not just a footnote in accounting, but an essential characteristic of the business model.