저자: 미상 · 외국인
요약 보기
엔비디아가 마이클 버리와 Substack 필자의 AI 버블 비판에 맞서 내부 메모를 낸 논쟁을 다룬다. 버리는 GPU 감가상각 기간 과장과 출시 주기 단축으로 데이터센터 투자 손실 가능성을 주장한다. 예: 2020~2025 GPU 감가상각 연수 3~5년→5.5~6년 확대, 바이두 23억달러 손상차손, H100 임대가 2023.9 3달러→2025.11 2.09달러. 엔비디아는 라운드트립 투자 규모(1190억달러 약정)와 보증 공개 문제를 지적받는다. 영상은 엔비디아가 엔론은 아니지만 주가(PER 45)가 AI 수요에 비해 과대평가됐다고 결론짓는다.
다듬은 글
With a market capitalization in excess of $4 trillion, Nvidia is the most valuable company in the world. When you're already at the top, there's not much room to climb higher, so you're instead gripped by the fear of falling down. The fear across the AI industry is palpable, with even big tech CEOs talking about a potential bubble.
One of the biggest voices warning about an AI bubble is a former hedge fund manager Michael Burry. Burry is best known for shorting subprime mortgages in 2009. Burry inspired the Hollywood film The Big Short, where he was played by Christian Bale. Recently, Burry has taken to Twitter to warn about a bubble in the AI sector.
At the center of this alleged bubble sits Nvidia. Nvidia was so disturbed by Burry's criticisms that they sent a memo to Wall Street analysts to try to refute some of his allegations. In the memo, Nvidia tries to explain that they are not similar to Enron. To be clear, Nvidia is nothing like Enron, but the fact that they feel they need to say this is not a good look.
In this video, we'll look at the escalating war of words between Nvidia and Michael Burry, and what this can tell us about the state of the AI bubble. We've been running Wall Street Millennial for more than 5 years now. The majority, in fact the vast majority of the videos we've ever published contain stock footage.
We get all of our stock footage from Storyblocks, which is also the sponsor of today's video. Storyblocks offers unlimited downloads of diverse and high-quality media for one predictable subscription cost. Everything you need in one place. 4K and HD video, templates, music, sound effects, images, and more.
You pay a monthly or annual subscription, and there are no hidden or extra fees. A lot of times when I'm editing a video, I download many dozens of clips of stock footage from Storyblocks. As I edit the video, I only end up using a fraction of what I downloaded, but it doesn't matter because a subscription gives you unlimited downloads.
Everything is royalty-free and ready to use. This unlimited source of content gives you the freedom to test, experiment, and create more effective video. I use Premiere Pro to edit my videos. I know nothing about animation, but I can still use the Premiere Pro templates from Storyblocks. They also have templates for After Effects, DaVinci Resolve, and Apple Motion.
If you're an aspiring YouTuber, or if you want to make videos of any kind, I cannot recommend Storyblocks enough. To get started with unlimited stock media downloads at one set price, head to storyblocks.com/wallstreetmillennial or click the link in the description. And now back to the video. In a series of Twitter posts and Substack blogs, Burry lays out his bear case for the AI industry.
Burry compares the current AI spending boom to the dot-com bubble of the 1990s. In the 1990s, broadband companies including AT&T, WorldCom, and Global Crossing spent tens of billions of dollars to deploy broadband internet cables. The thinking was that demand for internet traffic would grow exponentially, so this broadband infrastructure will generate massive revenue.
Internet adoption was much slower than these companies had projected, resulting in the Nasdaq bubble bursting in 2000. To understand the scale of overbuilding, by 2002 only 5% of broadband internet capacity in the US was being utilized. The broadband companies generated far less revenue than expected.
Many of the broadband companies borrowed money to build out the internet infrastructure in the first place. In 2002, one of the biggest broadband companies, Global Crossing, went bankrupt. Their broadband assets were marked as being worth $22 billion on their balance sheet. They were sold by the bankruptcy court for just $500 million, about 2% of the book value.
During this entire period, the number of US broadband subscribers increased and continued to increase continuously over the next 20 years. The internet eventually did become as ubiquitous as the bulls thought. It just took a lot longer than expected, so the infrastructure investments were not justified.
The current AI data center boom has some similarities to the broadband bubble of the 1990s. Cloud service providers are spending hundreds of billions of dollars to make new AI data centers. The AI data centers are indeed being utilized. OpenAI, Anthropic, and other AI startups are currently consuming all the compute capacity the data center operators can provide.
But OpenAI is generating minuscule revenue compared to its expenditures. OpenAI loses money on every ChatGPT user. They even lose money on ChatGPT Pro, which costs $200 per month. The AI infrastructure is indeed being utilized, but end consumers aren't paying for it. It's being subsidized by OpenAI's venture capital investors.
The situation is not sustainable. Eventually, the venture capital will run out. OpenAI is expected to generate $13 billion of revenue in 2025. They need to increase their annual revenue to hundreds of billions within the next few years to pay for all the data center capacity, which is currently under construction.
But Michael Burry has another criticism of the AI bubble. Even if OpenAI can somehow afford to pay for all of its spending commitments, he contends that the data center companies will still lose money. The cloud service providers have spent hundreds of billions of dollars to buy Nvidia GPUs. Even if demand for AI compute remains strong, Burry argues that these GPUs will become obsolete before the cloud service providers can make back the cost of their investments.
Since the release of ChatGPT in 2022, Nvidia has accelerated its product launch cycle. In 2020, they released their A100 Ampere GPU. In 2022, they released their H100 Hopper GPU. That's a two-year gap. In 2024, they released the H200 Hopper extended memory, which is an improved version of the H100. That same year, they also released the B200 Blackwell GPU.
Again, there was a two-year gap. In 2025, they released the B300 Blackwell Ultra. That's only one year after they released the B200. In 2026, they plan to release a new data center GPU called the Vera Rubin. In 2027, they plan to release the Vera Rubin Ultra. Prior to 2024, they released a new data center GPU every two years.
Now it's every one year. This makes sense. Given the surge in demand for GPUs, the amount of profit Nvidia can make has exploded. It's now worth their while to spend far more on research and development to develop better GPUs faster. As Nvidia's product cycle compresses, old GPUs will become obsolete more quickly.
In March of 2025, Nvidia CEO Jensen Huang publicly joked that once Blackwell is released, you won't even be able to give away the older Hopper GPUs. I said before that when Blackwell starts shipping in volume, you couldn't give Hoppers away. And this is what I mean. And this makes sense. If anybody if you're still looking to buy a Hopper, don't be afraid.
I'm it's okay. But I'm the chief wreck revenue destroyer. My sales guys are going, "Oh, no. Don't say that." There are circumstances where Hopper is fine. That's the best thing I could say about Hopper. There are circumstances where you're fine. Not many. If I have to take a swing, and so that's kind of my point.
Um when the technology is moving this fast, uh you you and because the workload is so intense and you're building these things that are factories, you we really we really like you to to um uh uh to invest in the right the right versions. According to Nvidia, a 100 megawatt data center full of Hopper H100 GPUs will require 1,400 racks and produce 300 million AI tokens in a given period of time.
If you instead switch to Blackwell B200 GPUs, they'll only take up 600 racks, so less than half of the space, but they'll produce 12 billion AI tokens, 40 times greater than the H100s. As cloud service providers deploy B200s, the supply of computing capacity will surge. This will decrease the price of compute.
Eventually, the price will become so low that it's no longer profitable to operate H100s. They're too energy inefficient. That's why Jensen Huang was joking that you won't be able to give H100s away. Even if an H100 is in perfect condition and can work as intended, it will become obsolete. Over the past few years, the so-called hyperscalers have increased their estimated useful life estimates for their GPUs.
In 2020, Meta, Google, and Microsoft depreciated them over an estimated useful life of three years, Amazon four years, and Oracle five years. By 2025, Meta increased their estimates to five and a half years. Google, Oracle, Microsoft, and Amazon increased to six years. Amazon subsequently decreased it to five years.
The biggest increases came from Google and Microsoft, who both doubled their estimated useful life. These two companies are also believed to have spent the most money on Nvidia GPUs during this period. So how do the hyperscalers justify this? They claim that advances in technology and software make it such that GPUs physically last longer.
Even if this is true, an old GPU can become economically obsolete even if it's in perfect physical condition. We're already starting to see some signs of this. In the first quarter of 2025, Amazon reduced the estimated useful life of its GPUs from six years down to five years. The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.
This resulted in an increase in depreciation and amortization expense of $217 million and a reduction in net income of $162 million during the quarter. If the hyperscalers are forced to decrease the useful lives of their GPUs, this will decrease their reported net income, which investors will not be happy about.
Michael Burry points to the example of Baidu as the first hyperscaler to recognize a large impairment of GPUs. Baidu operates China's largest search engine, and they're also a cloud service provider. In 2023, Baidu launched its own AI chatbot called Ernie. In addition to its own Ernie chatbot, Baidu also sells computing capacity to third-party AI startups in China.
You can think of Baidu as a Chinese version of Google. In the third quarter of 2025, Baidu recorded an impairment charge of $2.3 billion. This caused them to post a loss for the quarter. In the previous quarter, the entirety of their fixed assets was valued at $4.5 billion, the vast majority of which were data centers.
So, this impairment represented more than 50% of their fixed assets. Baidu's CFO explained, "Quote, we've conducted a comprehensive review of our infrastructure portfolio. Some of the existing assets no longer meet today's computing efficiency requirements. Unquote." In other words, they have a bunch of GPUs that are now obsolete and no longer useful.
Up until 2022, Baidu purchased Nvidia's A100 GPUs. In late 2022, the US government implemented export controls, which banned Nvidia from selling its highest-end chips to Chinese customers. So, they stopped selling the A100s to Baidu. In response, Nvidia made the H800 and later the H20 GPUs. These have lower performance specifications to comply with the export controls.
These chips are made exclusively for the Chinese market. Recently, Baidu has ordered AI chips from Huawei and is making its own AI chips. These homegrown Chinese chips are inferior to Nvidia's top-of-the-line GPUs, but they're probably better than the watered-down export compliant versions. We don't know which GPUs specifically were responsible for Baidu's recent impairments, but it was probably either the aging Nvidia A100 or possibly the watered-down H800 and H20 export variants.
Baidu's situation is not directly comparable to the US hyperscalers like Amazon, Google, and Microsoft. The US hyperscalers are not subject to export controls, so they've been buying Nvidia's latest and greatest GPUs. But, this just means that the US hyperscalers are a few years ahead. Eventually, they'll face the same problems as Baidu.
Let's look back at Nvidia's product cycle timeline. The A100s are probably now obsolete. The AI data center investment boom started in earnest in 2023. The US hyperscalers bought tens of billions of dollars worth of H100s. The H100s are a generation ahead of the A100s, so they're still usable. But, what happens once Vera Rubin comes out in 2026?
The H100s will probably become obsolete, and the US hyperscalers will be forced to take massive impairments, just like Baidu. We've talked a lot about abstract accounting terms like depreciation and estimated useful lives. But, what does all this mean in practice? There's a company called Silicon data, which tracks the rental prices for Nvidia's H100 GPUs.
The full data set is behind a paywall, but we were able to find a few data points in publicly available news articles. In September of 2023, the average rental price was about $3 per hour. By May of 2025, it decreased to $2.50. By November of 2025, it decreased to $2.09. That's a decline of 30% in a little over 2 years, or about 15% per year.
Let's just assume that the rental price continues to decline at a rate of 15% per year. We'll analyze this using CoreWeave's financials. They're a pure-play AI data center company, which makes the analysis easier. As of the end of the third quarter, CoreWeave had $14.6 billion of technology equipment.
This is mostly Nvidia GPUs and other supporting chips and server components. Plus, they have $1.1 billion of data center equipment and leasehold improvements. CoreWeave depreciates its GPUs over an estimated useful life of 6 years. When you replace the GPUs, you probably have to replace most of the data center equipment as well, because the newer models of GPUs have different infrastructure requirements.
So, CoreWeave spent almost $16 billion on AI data centers. They need to make back this investment within 6 years. They also have $7 billion of construction in progress. These are new data centers that are not yet operational. For now, we're just looking at the $16 billion of data centers that are already operational.
In the third quarter of 2025, they generated $1.4 billion of revenue. Their cost of revenue was $370 million, giving them gross profit of a little over $1 billion. CoreWeave's cost of revenue consists of the direct cost to operate their data centers, rent, utilities, staff, etc. It does not include GPU depreciation.
In the first year, they'll make $5.4 billion of revenue. That's just their Q3 revenue annualized. They incur cost of revenue of $1.5 billion, so they generate a little less than $4 billion of gross profit. Each year, their revenue declines by 15%, while their cost of revenue remains constant. So, their gross profit gradually shrinks.
By the end of the sixth year, they've cumulatively generated $14 billion of gross profit. Their investment was $15.7 billion. They will not even make back their upfront investment. To fund the construction of these data centers in the first place, CoreWeave had to borrow billions of dollars at interest rates ranging from 7 to 15%.
They're probably not even going to generate enough gross profit to cover the initial investment, let alone the billions of dollars of interest expense that will accrue over the next 6 years. Keep in mind that all of this data center buildout is predicated on orders for AI companies, mostly OpenAI and Anthropic.
To pay for all of this, OpenAI, Anthropic, and other AI startups will need to increase their annual revenues by hundreds of billions of dollars within the next few years. Otherwise, they'll default on their purchase orders. But, even if OpenAI can make good on its spending commitments, the AI data center companies like CoreWeave and the hyperscalers will probably still lose money, because the GPUs depreciate so quickly.
In the short term, the fast product launch cycle is good for Nvidia. But, in the long run, if the data center companies have to recognize massive impairments, they will eventually realize that this is an unprofitable business, at which point they'll scale back their expansion plans, and Nvidia's revenue will decline.
Nvidia was so spooked by Burry's criticisms that in late November, they sent a memo to all the Wall Street analysts who cover their stock. The full memo was leaked to the media. In the memo, Nvidia responded to two critics. Firstly, they responded to Michael Burry's statements on Twitter. Secondly, they responded to a Substack blog titled, "The Algorithm That Detected a $610 Billion Fraud: How Machine Intelligence Exposed the AI Industry's Circular Financing Scheme." It was written by a guy named Shanaka Perera.
Shanaka's post is viewable for free on his Substack page. I've linked it in the description below if you want to read the whole thing. But, these are his key points. Nvidia's accounts receivables are increasing. This indicates that his customers are having trouble paying for the GPUs. Nvidia's inventory is increasing.
This may indicate that they're having trouble selling all the GPUs they're producing. Shanaka points out Nvidia's circular financing deals. He specifically flags Nvidia's $2 billion investment into Elon Musk's xAI in October 2025. Nvidia's xAI investment is a bit complicated. Nvidia isn't investing into xAI directly.
xAI created a special purpose vehicle, or SPV. The SPV will buy Nvidia GPUs and rent them to xAI. The total value of the SPV is $20 billion, of which $7.5 billion is equity and $12.5 billion is debt. Nvidia contributed $2 billion to the equity portion of the SPV. The terms of the SPV require it to spend the money on Nvidia GPUs and keep their utilization above 70%.
The SPV will sell the computing capacity to xAI. In the most inflammatory part of Shanaka's blog, he compares Nvidia to Enron. Enron created SPVs, which it owned and controlled. Enron sold its telecom and energy infrastructure assets to its own SPVs. This allowed them to book fake revenue from their SPVs, even though they were not receiving any cash from external customers.
Shanaka compares the xAI SPV to Enron's SPVs. Quote, "Nvidia provides equity capital to an entity that exists primarily to purchase Nvidia's products. The transaction appears as an arms-length sale in Nvidia's accounting, but economically, Nvidia is funding its own revenue. Unquote." In Nvidia's memo, they try to refute Michael Burry and Shanaka Perera's criticisms one by one.
Nvidia maintains the demand for its data center GPUs remains strong, and they are in fact supply constrained. Their inventory balance is growing, but this is not because they're having trouble selling their products. Inventory includes significant raw materials and work in progress. As they grow their revenue, you should expect them to have more work in progress at any given time, as they're producing more GPUs.
And while their accounts receivable has increased, they have not had any significant customer defaults. In my opinion, Nvidia's responses to these allegations are pretty solid. There hasn't really been any sign that Nvidia is having trouble selling its GPUs today. The more concerning allegation is about the circular financing.
In the third quarter of 2025, Nvidia made $3.7 billion of strategic investments into AI startups. This represented 7% of their revenue in the quarter. In the first three quarters of 2025, they made $4.7 billion of strategic investments, representing just 3% of their revenue. The companies in Nvidia's strategic investment portfolio mainly raise from third-party providers, not from Nvidia.
These companies are growing their revenues rapidly, indicating a path to profitability and strong underlying customer demand for AI applications. Furthermore, they predominantly generate revenue from third-party customers, not from Nvidia. Nvidia has a venture capital arm called N Ventures, where they invest in dozens of small AI startups.
For the most part, these investments are relatively small. Additionally, Nvidia sometimes makes larger investments into various AI and tech companies. Here, I've compiled a list of the strategic investments Nvidia has made in the first 9 months of 2025. Collectively, Nvidia invested $4.7 billion into these companies.
The companies largely fit into two categories. Nvidia invested into so-called neo clouds, including Lambda and Scale and Fermas. Neo clouds are cloud service providers that exclusively provide computing services to AI startups. These companies are direct customers of Nvidia. They buy Nvidia GPUs to fill up their data centers.
The second category is AI startups including Cohere, Together.ai, and Thinking Machines. These companies are developing AI models. They typically do not purchase Nvidia GPUs directly. They instead rent computing power from neo clouds such as Lambda and Scale AI and Fermyon. Thus, they are indirect customers of Nvidia.
Nvidia also made an investment into a company called Commonwealth Fusion Systems, which is trying to develop nuclear fusion power. Nvidia claims that the companies in its strategic investment portfolio predominantly generate revenue from third-party customers, not from Nvidia. While this is technically true, Nvidia's web of investments across the AI landscape is so broad that its money is flowing around almost everywhere.
Nvidia owns equity stakes in various neo clouds. The neo clouds generate revenue from AI startups. Many of these AI startups have received investment from Nvidia. While the neo clouds revenue does not directly come from Nvidia, a lot of this money comes from Nvidia indirectly. Even if there is some circularity in Nvidia's investments, they argue that it's too small to matter.
Their strategic investments only represented 3% of their revenue in the first 9 months of the year. Even if all of that is circular, that means at least 97% of their revenue is real revenue from unrelated external customers. This is technically true, but we also have to look at the trend. In the first half of 2025, Nvidia only made $1 billion of strategic investments.
In the third quarter alone, they made $3.7 billion of investments. Over the past few months, Nvidia has committed to investing $100 billion into OpenAI, $10 billion into Anthropic, $5 billion into Intel, $2 billion into Synopsys, and $2 billion into XAI. These commitments add up to $119 billion. All of these investments have round-trip characteristics.
In the trailing 12-month period, Nvidia's revenue was $187 billion. The investment commitments Nvidia has made now represent more than half of their annual revenue. This doesn't even include their off-balance sheet arrangements such as their $6.3 billion commitment to purchase CoreWeave's unused capacity in the future.
In Chanakya's Substack blog, he specifically calls out Nvidia's $2 billion investment into a special purpose vehicle associated with Elon Musk's XAI. The sole purpose of this SPV is to buy Nvidia's products. Nvidia categorically denies any comparison between themselves and Enron. Nvidia claims it has only one guarantee, for which the maximum exposure is $860 million.
They reference note nine. Unfortunately, the version of the memo I was able to find does not include the notes, so I'm not sure what they're referring to. However, this is already a bit fishy. Nvidia provided a $6.3 billion guarantee to purchase CoreWeave's unused capacity. As far as I know, this guarantee is still standing.
So, how can they say that their only guarantee is $860 million when they have a $6.3 billion guarantee to CoreWeave? To be fair, the CoreWeave guarantee does not include an SPV. Nvidia made the guarantee directly to CoreWeave. This is an off-balance sheet liability. We only know about it because it was reported in the media.
What other guarantees does Nvidia have that we don't know about? To be clear, Nvidia is not Enron. Enron created a convoluted web with hundreds of SPVs. Enron controlled all these SPVs and effectively had 100% economic ownership of them, but they were deconsolidated from Enron's balance sheet. Enron did two things with these SPVs.
Firstly, they sold various of their assets to their own SPVs to create fake revenue. Secondly, they transferred liabilities to their SPVs to make their own balance sheet look better. In reality, they had 100% economic exposure to the SPVs, so they were still on the hook for the liabilities. The reason why Enron was a fraud is because they own 100% of their SPVs.
For the most part, these SPVs had minimal or no outside investors, so it was fraudulent for Enron to deconsolidate them. The XAI SPV is different. It has $20 billion, of which $7.5 billion is equity and $12.5 billion is debt. Nvidia reportedly contributed $2 billion to the equity portion. That's only about 25%.
The majority of the money came from external venture capital firms. Nvidia has no obligations to the SPV. It's actually the other way around. The SPV has obligations to purchase Nvidia GPUs. Chanakya is right to draw attention to the XAI investment. It indeed appears to be circular, but comparing it to Enron is a bit over the top.
The final issue we'll look at is depreciation. Nvidia's memo states the following, "Quote, claim Nvidia's depreciating PP&E more slowly than it appears, indicating that depreciation expense is understated. If properly reported, depreciation would be higher and net income would be lower. Unquote." This is what you call a straw man argument.
In the source document section of the memo, Nvidia claims it is responding to Michael Burry's tweets as well as Chanakya Perera's Substack article. Neither Burry nor Chanakya have alleged that Nvidia understates its own depreciation. It was mainly Michael Burry who talked about depreciation. Burry said the data center operators, the likes of Google, Amazon, Microsoft, and Oracle are understating depreciation.
While Nvidia makes GPUs, it does not operate its own GPUs in any significant quantity, so it doesn't even make sense to talk about Nvidia's depreciation. Separately, Nvidia claims that A100 GPUs, which were released in 2020, continue to run at high utilization and generate strong contribution margins, retaining meaningful economic value.
They tell you to reference appendix A. But if you look at appendix A, all it has is a list of depreciation schedules of the hyperscalers. Nvidia provides zero evidence that any of them are operating A100s at high utilization. In my opinion, Nvidia's assertion about the economic longevity is almost certainly false.
Remember, earlier in the video, we looked at Baidu's massive impairment. We believe that this impairment was likely linked to their aging A100 GPUs. If A100s are already obsolete in China, they're even more obsolete in the US. The US is a couple generations ahead of China due to the chip export controls.
So, what's the takeaway from all of this? Nvidia is not Enron. It's not a fraud, but the massive revenue growth they've experienced over the past couple years is probably not sustainable. Nvidia's stock is very expensive. They have a $4.4 trillion market cap. Their price-to-earnings ratio of 45 is much higher than that of the S&P 500.
To sustain this premium valuation, they need to continue growing their revenue and profits. The underlying demand for AI services is not enough to support this. That's why they feel obliged to engage in round-trip investments of ever-increasing size. Nvidia is one of the most innovative and successful companies of our time, but their valuation is being propped up by a house of cards.
Nvidia understands this. They're nervous. They're so nervous that they're now getting into Twitter debates with the likes of Michael Burry and Chanakya Perera. All right, guys. That wraps it up for this video. What do you think about Nvidia? Let us know in the comments section below. As always, thank you so much for watching, and we'll see you in the next one. Wall Street Millennial signing out.
Comments 0