저자: MichaelBurry · 외국인
요약 보기
AI 인프라 붐이 감가상각 회계 조작(내용연수 연장)으로 이익이 부풀려졌고, GPU는 에너지 효율·물리적 열화·임대료 폭락으로 실제 수명이 1~3년이라고 주장. 엔비디아 고객 간 순환금융(CoreWeave 담보대출·지분 투자)과 극도의 고객 집중도가 취약성을 키운다는 진단. 이를 근거로 팔란티어·엔비디아 풋옵션 등 하락 베팅을 하고 있으며, 석유서비스·제약 등 실물 가치주에 콜 옵션을 보유. 2025년 11월 해외 펀드 등록 말소 후 패밀리오피스 전환.
다듬은 글
Welcome back to the deep dive. So, if you've been watching the markets at all, you just you can't escape it. The AI boom. We're seeing numbers that honestly feel made up. Nvidia, for example, blasting past a $3 trillion market cap. It's pure euphoria. And whenever we see that kind of, you know, universal excitement, our job is to find the smartest person in the room who's saying, "Hang on a second.
Today, that voice is Michael Bur, famous, of course, for calling the 2008 crash." And he's not just talking about high valuations. He's alleging the whole AI infrastructure boom involves, and this is a quote, one of the more common frauds of the modern era. H So, our mission today is to really forensically unpack that claim.
Look at his sources, his math, and see if it holds up, right? And it's so important to get this straight from the beginning. Bur's argument isn't just, oh, Nvidia's PE ratio is too high. That's the easy surface level take. He's going much, much deeper. He's questioning the very accounting practices that make Nvidia's biggest customers, we're talking Google, Meta, Microsoft, look profitable enough to even afford this stuff.
The hyperscalers, the hyperscalers. The critique is uh totally structural. We're going to look at three things. First, how they account for these massive costs. Second, the physical reality of how long these expensive chips actually last. Okay? And third, the really scary part. These strange circular financing loops built on assets that might be depreciating way faster than anyone admits.
All right, let's unpack that first one. If he's talking about fraud or, you know, something close to it, where does he start? It has to be with the cost, right? The capex. These companies are spending hundreds of billions on GPUs. They absolutely are. And here's the thing. Nvidia's stock only stays this high if its customers can afford the chips.
And crucially, if those customers can prove to their shareholders that the spending is worth it by showing strong profits. Bur says those profits are basically being manufactured. Manufactured how? Through something he calls depreciation arbitrage. Okay, let's slow down on that because you said it's the cornerstone of his whole argument.
So, a company buys say a $20,000 GPU, they can't just write off the whole cost at once. They depreciate it over its useful life. Why is that little estimation so important? Because management gets to decide what that useful life is. It's an estimate. It's not a law of physics. Wow. So imagine you're a hyperscaler.
You buy 10 billion worth of GPUs. If you say their useful life is three years, your annual expense is about 3.3 billion. That hits your bottom line. Okay, simple math. But what if you just decide they'll now last for six years? You cut your expense in half. You just cut your expense in half instantly.
Your annual expense drops to about 1.67 billion. And that difference flows straight to pre-tax income. You've just created over a billion and a half in profit from a spreadsheet entry. Precisely. It is instant non-cash earnings growth. And what's Well, what's alarming is that basically everyone started doing this at the exact same time right when the AI arms race kicked off.
Yeah, the numbers on that, right? We do. The sources are specific. Meta, for instance, extended the useful life of its servers from four years to five and a half. that little change. It's expected to reduce their depreciation expense by almost $3 billion in 2025. Wow. And Alphabet, Google, they were even more aggressive.
They went from a threeyear life for some equipment in 2020 all the way up to six years by 2024. So you're doubling the lifespan of cutting edge tech in an era where the tech is moving faster than ever. That that feels like a complete contradiction. It is. And Bur points to Oracle as uh the most egregious offender.
The projections suggest that by 2028, Oracle could be overstating its real earnings by almost 27% just from this one change. So what's the total number? What does he claim is the scale of this this air pocket as you called it? When you add it all up, the allegation is staggering. He's saying that between 2026 and 2028, the hyperscalers will understate their real depreciation cost by a cumulative 176 billion.
176 billion. It's money borrowed from the future because eventually that equipment will become obsolete or it'll just break. And when it does, they'll have to take huge massive impairment charges all at once which could absolutely crater their earnings down the line. Okay, so this whole accounting trick, it really all depends on one question.
Can an Nvidia H100 chip actually be useful for 6 years? This is where we have to pivot from the accounting to the you know the physical reality of the server rack. So the bull case the industry line is that these chips don't die they just get demoted. Today's top-of-the-line training chip becomes tomorrow's workhorse for inference just running the models.
Why does Burie say that's wrong? He says it's wrong because of well three physical and economic facts you just can't ignore. The first one is the energy efficiency wall. This is maybe the most important part of the whole argument. the what the energy efficiency wall. See, it's not about if the chip breaks, it's about the cost to run it.
When the nextG chip comes out, say the Blackwell B200, it offers maybe two and a half to four times the performance for basically the same power draw. And if you're running a data center with tens of thousands of these, your biggest cost over time isn't buying the chip, the capex, it's the electricity to run it, the opex.
So after a few years, keeping that old H100 running is just burning money. You're spending way more on electricity for less performance than if you just bought the new one. Exactly. It becomes economically irrational. The chip's economic utility just collapses long before its six-year accounting life is up.
And what about them just, you know, physically breaking down? Can they even last that long under constant stress? That's the second point. Accelerated degradation rates. These AI workloads are brutal. They run hot. They run constantly. We've seen reports citing Alphabet engineers saying these GPUs have a practical life of only one to three years under heavy use.
One to three years. Yeah. And get this, Meta's own internal study suggested component failure rates could hit 9% a year. 9%. So if almost a tenth of your fleet is dying every year, how can you possibly claim a six-year useful life? It's just it's aggressive. And there's a third piece of evidence, right?
The one that's happening in the market right now. Yes. the collapse of rental rates. This is the canary in the coal mine. It proves the market knows these things are losing value fast. What's happening to the price? It's cratering. The hourly spot price to rent an H100 has fallen off a cliff. We're talking a 60 to 70% drop.
It went from about $8 an hour down to maybe $285 by late 2025. So, the open market is screaming that this asset does not have a six-year straight line value. Not even close. Yeah. And that erosion of its earning power leads us straight into the financing problem. Okay. So, if the actual collateral, the GPUs are losing value this fast, then any debt built on top of that collateral is shaky to say the least.
Bur makes a comparison here to the com bust, right? Things like roundtpping revenue at Enron. He does. He sees the same kind of circular logic. And the poster child for this today is a company called Necoreweave, the cloud provider. Exactly. And the feedback loop is well, it's pretty wild. Step one, Coreweave buys billions of dollars of NVIDIA GPUs.
Okay. Step two, Coree funds this by taking on massive debt like a $2.3 billion loan from big private credit firms like Magnetar and Blackstone. That seems normal so far. A company borrowing to buy equipment. What's the catch? The catch is the collateral. Those multi-billion dollar loans are secured by the very GPUs they just bought.
the same GPUs whose rental rates are collapsing. I see where this is going. But it gets better. Step three, Nvidia books that huge purchase from Coree as revenue, which pumps up its own stock. And then step four, Nvidia turns around and invest equity back into Core Weeave. Wait, so Nvidia is helping to fund its own customer, who then uses that funding to buy more chips from Nvidia, which Nvidia books as revenue.
That's the loop. It feels incredibly fragile. The whole system depends on rental demand for those GPUs staying skyhigh. If it falters for even a moment, the debt can't be serviced, the collateral is worthless, and you have a contagion problem, and it gets even more concentrated. We know Cororeweave posted an operating loss of around $40 million in Q3 2025.
Even with a massive $55 billion revenue backlog, the cash isn't there yet. And this whole system rests on just a handful of players, right? Nvidia's filing showed this pretty clearly. Oh, it's terrifyingly concentrated. Their filing showed two mystery customers made up 39% of total revenue. One of them, customer A, was 23% in a single quarter.
23% from one buyer. And if you drill down even more, Coreeave itself gets 71% of its revenue from one customer, Microsoft. So a slowdown at Microsoft could torpedo Coreweave, which torpedoes the private credit firms holding the GPU back debt, which torpedoes Nvidia. It's a house of cards. It's the classic definition of late stage bubble concentration.
So if this is what Bur believes, misleading accounting, obsolete hardware, circular financing, how is he actually betting on it? What is Scan Asset Management doing with its money? Well, his Q3 2025 portfolio shows what you'd call a classic barbell strategy. It's built for extreme outcomes, meaning one end of the barbell is loaded with massive short bets against the AI bubble.
The other end is long bets on really unloved old economy sectors. the anti-bubble, if you will. Okay, let's start with the bearish side. What's he shorting? About 80% of his reported portfolio was in PUT options. The biggest bet by far was against Palunteer. That was about twothirds of his exposure.
Just a pure bet against hype and valuation. And Nvidia, Nvidia was the second biggest, about 13.5%. That's the direct shot against this entire infrastructure thesis we've been talking about. The sources mentioned he's using high delta PUT options. What does that tell us? It tells us he's not betting on a slow decline.
He's betting on a crash. A high delta PUT is super sensitive to price moves. It's a way to get maximum leverage on a bet that a stock is going to fall and fall fast. So, total conviction. What's on the other side of the barbell? The long positions. He's buying call options on the real economy. Things he thinks are being ignored.
Fizer, the pharma giant, a classic value play. And Hallebertton, the energy services company. So oil and drugs over chips basically. Yeah. He's betting that the physical world, the energy needed to power these data centers of the basic materials is being massively undervalued compared to the silicon itself.
And there were two other big strategic moves he made that really signal his conviction level. The first one was the great China exit. Right in Q3 2025, he just dumped everything. All his long-held value bets on Chinese tech companies like Alibaba and JD.com gone. Why? He was freeing up cash. He needed that liquidity to fund these huge expensive short bets on US tech.
It shows his priority shifted completely to shorting this bubble. And the second move was even bigger. He decided to go dark. Exactly. In November 2025, he dregistered as a hedge fund and converted it into a family office. And the implication of that is no more public filings. No more public filings.
He no longer has to disclose his positions every quarter. It gives him total opacity, total agility. He's basically saying, "I have my position set, and now I want to navigate the coming crash without anyone watching my every move." It's a move you make when you think things are about to get very, very volatile.
That's the signal. Yes. So, to pull it all together, Bur isn't saying AI is fake. That's not the argument. He's saying the financialization of AI has created this incredibly fragile house of cards. The risk is hidden in the accounting. It's hidden in these weird financing deals. It feels a lot like the complex financial products of 2008.
It's the plumbing he's worried about, not the technology. Right. And we've laid out the evidence from the sources. That potential $176 billion earnings distortion from depreciation. The bet that's backed by collateral that's losing 70% of its value on the open market. The insane concentration risk. At the end of the day, the bullcase seems to need this unending decadesl long explosion in demand for inference to soak up all these chips and justify their six-year value on paper.
So if Burie is right, you have to ask yourself if the promise of AI revenue will really outweigh the crushing physical reality of electricity costs and technological obsolescence. And the final thought to leave you with is this. If that collapse in rental prices really is the canary in the coal mine, how much longer can the accounting tricks keep the silence muffled? Something to think about.
Comments 0