저자: HamishHodder · 외국인
요약 보기
마이클 버리가 AI 하이퍼스케일러(구글·메타·오라클 등)가 서버 감가상각 내용연수를 늘려 이익을 부풀린다고 주장한다는 내용. 메타·알파벳·아마존 등의 내용연수 연장으로 수십억~수백억 달러의 비용이 이연되며, 바이두의 대규모 손상차손 사례를 들어 향후 실적 충격 가능성을 경고. 다만 엔비디아 CFO는 6년 전 A100이 여전히 풀가동 중이라고 반박하며, 회계 변경이 공시된 만큼 실질적 사기라는 증거는 없다고 평가.
다듬은 글
I mean, there were very specific identifiers. Extremely recognizable. The one one of the hallmarks of mania is the rapid rise in complexity and the rate of fraud. And did you know they're going up? Okay, that was a bit dramatic, but I couldn't resist. Michael Burry is back. He just broke a 2-year silence with a return to his cryptic tweeting ways.
Plus, he launched a Substack with some analysis behind a $400 paywall. And within his recent commentary is the claim of fraud being committed by the AI hyperscaler companies, Google, Meta, Oracle, spending hundreds of billions of dollars on Nvidia chips for their AI infrastructure build-outs, but using accounting tricks to overstate their profits.
I want to break down Burry's claims and the consequences on the stock market if he's right, as well as the counterarguments and response from Nvidia's own executives. You probably know Burry from his Oscar-worthy portrayal of random office employee in The Big Short, but few know that he was actually one of the first to correctly predict and profit from the 2008 housing crisis.
Since then, his predictions have been a bit more of a mixed bag, to put it lightly. But, his tweets are always filled with incredibly interesting insights, if you can decipher what he's saying. And that's why I'm here, because you're not going to understand Burry's tweets by reading Twitter comments that have been clearly written by ChatGPT to farm engagement.
By the way, I'm taking full credit for summoning Burry back to Twitter, because I was also on a 2-year Twitter hiatus. I came back for the first time on the 29th of October, and 45 hours later, Burry ended his 2-year Twitter break. But, the tweet I want to talk about is the one on the 11th of November, where he wrote, "Understating depreciation by extending the useful life of assets artificially boosts earnings.
One of the more common frauds of the modern era." By his calculation, the accounting change will cause hyperscaler companies to understate depreciation by $176 billion from 2026 to 2028. Or, in other words, the profits that are being reported by these companies are significantly higher than they actually are in reality.
In 2028, he estimates Meta's profits could be reported as 21% higher than the real figure. So, are companies doing this? Well, yes. All of the big players have made accounting changes that have boosted their reported profits considerably. But, before we go further into the numbers and the potential consequences, we need to do a little bit of accounting 101.
Imagine a company needs to spend a million dollars on Nvidia chips every 5 years. On the income statement, that cost is spread out evenly over the useful life of the chips. That's called depreciation. Of course, in the real world, the company is effectively just guessing. If the chips actually needed to be replaced after 3 years, then the depreciation expense was understating the actual cost and overstating the firm's profits.
How companies are able to depreciate assets is determined by US GAAP accounting standards. Basically, they're estimating the expected useful life of the asset, where useful life means the period over which it contributes to future cash flows. So, things like wear and tear and technological obsolescence play a big role.
Even if a chip is still working at the same performance as day one, if there are newer, faster chips now available, its economic useful life might be mostly gone. When it comes to companies building out their AI infrastructure, the true useful life of these assets is fairly unclear, and that gives management some leeway in deciding over how many years they'll spread out that cost.
The questions being, how far are AI companies stretching it? Are they justified in doing it? And if not, how badly will it come crashing down when it's exposed? The internet bubble was a different environment than today's market, but companies extending depreciation to make earnings look better was a feature back then, too.
Cisco Systems was the poster child. The explosive demand for their infrastructure products drove huge profits before collapsing spectacularly. This, by the way, is the Warren AI feature from today's sponsor, investing.com. Their investing pro subscription opens the door to the best real-time data in the market and a suite of tools to utilize it, such as tracking your stock investments with expert valuation estimates, the ideas tab for seeing notable investor trades the second they're reported, stock screeners, pro picks, and Warren AI, the ChatGPT of investing that will save you a ton of time.
And of course, I have an enormous Black Friday discount for you. It's currently 60% off, plus a little extra for my audience, so actually 75% off. This is an insanely good deal, and it's only going to be available for a very limited time. So, take your portfolio to the next level and try Investing Pro by clicking the link in the description below.
The first part of Burry's claim is definitely true. Mark Zuckerberg's company, Meta Platforms, historically depreciated servers over a 4-year period. In 2022, they said the servers would last a little bit longer and extended it to 4.5 years. Starting in 2023, they extended the server and network equipment life to 5 years.
And most recently, in 2025, they extended the useful life to 5.5 years. That's latest change alone will reduce the company's depreciation expense by about $2.9 billion in the financial year 2025. Or, to put it another way, this year so far, Meta reported a $58 billion profit, $12 billion higher than last year, and $2 billion of that increase is pure accounting trickery and has nothing to do with their operations.
Alphabet, the parent company of Google, extended the useful life to 6 years in 2023, essentially doubling the lifespan of their server assets. They made this change after doing a useful life study, which found that the advancements in semiconductor performance were slowing. So, they were able to use their existing hardware for longer.
This announcement, by the way, was made just 1 day after Meta announced their own depreciation changes. The next year, Amazon decided it's time to look at their depreciation. Starting in 2024, they also extended useful life from 3 years to 6, boosting their first quarter after-tax profits by $900 million.
Interestingly, Amazon actually backtracked a little bit this year. They shortened their lifespan for servers from 6 years to 5, citing increased pace of technology changes. Putting it all together, these companies and others like Microsoft and Oracle have all been systematically pushing out their depreciation schedules by 1 to 3 years.
They claim it's because the chips and servers have genuinely seen improved longevity, but the effect, whether you think it was their motivation or not, has been lower expenses and higher profits. So, is it true? Is the AI hardware actually lasting longer? Well, before I get into the numbers, let's explore what happens if they're not.
A skeptic like Burry would argue that the hyperscalers have a different motivation for extending depreciation. These companies began seeing a slowdown in their core businesses, and magically, they all discovered that their servers actually last a lot longer, all at the same time. Reporting higher profits has supported and helped justify their stock prices surging in recent years.
Take Google, for example. The company is currently valued at about $3.6 trillion. 30 times higher than their expected profit this year. That's a hefty price tag, and the reason investors are willing to pay it is because they expect profits to rise a lot in the next few years. This financial year, 2025, Google will lay about $90 billion for capital expenditures, the vast majority of which is for tech infrastructure.
With the accounting changes, they'll record just $22 billion as a depreciation expense and report a profit before interest and tax of $130 billion. On the flip side, had they kept the 3-year schedule from just a couple of years ago, their profit would have been closer to $123 billion. So, instead of having to report a pathetic 5 to 6% growth in profits, they can report 16%.
But, of course, accounting wizardry doesn't just magically make chips last longer. And so, when the chips become obsolete after 3 years instead of 6, Google will have to report about half of the costs as an impairment in a single period. Burry has actually shared an example of this already happening to another company, Baidu.
In 2021, they extended depreciation from 4 years to 5 years and boosted profits by 814 million yuan. In 2024, they extended depreciation again from 5 to 6 years and boosted profits by another 1.2 billion yuan. Well, fast-forward to their latest quarterly results from this week, and the company was forced to report a 16 billion yuan impairment expense.
The management team stated, "Some of the existing assets no longer meet today's computing efficiency requirements." And that might just be the understatement of the century. The green bars on this chart are Baidu's pre-tax profits for the past 18 months. The orange bar is the impairment. So, before the impairment, total profits were 40.7 billion yuan.
In reality, it was more like 25 billion yuan. And no, this isn't because it's a Chinese company. They did the same accounting tricks as the American tech giants. The only difference is Baidu is tiny compared to them. So, if the same thing happens in the US, it could be ugly. Investors realizing that future profits are going to be much smaller than expected suddenly makes today's stock prices look egregiously high, and the fall back to earth could be violent.
Also, I've seen some people push back and say it doesn't matter because all of this is visible in the cash flow statements. Burry's implication that they are cooking the books or hiding accounting is completely false because all of the accounting is apparent in the cash flow statement. It's all there.
They're following GAAP standards, and then investors make a market. And they all decide, what do I want to value this company on? Cash flow? EBITDA? Yes, Burry is using the term fraud pretty liberally here. The spending is fully reported on the cash flow statements. But the point is that companies are telling investors that the 90 billion dollars for servers, for example, is a lump sum payment that will provide benefits for many years.
If that narrative changes to, well, actually, we just need to spend 90 billion dollars every single year to survive, the market is going to react negatively to that. Of course, Burry's entire argument completely falls apart if it turns out that the servers and chips are legitimately lasting six years instead of three.
So, are they? Well, unfortunately, there's not really a simple answer. If Nvidia comes out with a GPU that delivers twice the performance of the previous version, companies will have to buy them in order to remain competitive, making their old chips worthless. If this happens over a two-to-three-year product cycle, like Burry suggests, then a six-year depreciation schedule does seem excessive.
But it's also just not that simple because companies are still using their older generation chips in different ways within the company after they've replaced them. Just as an example, when Meta buys 350,000 brand new H100 Nvidia chips, they'll probably use them for a couple of years for training new AI models.
But even when Nvidia comes out with a better GPU, Meta can repurpose their old chips for inference, running the trained models for users. And even after that, GPUs get repurposed in smaller ways within the Meta organization. So, whether or not companies are producing revenue on their older generation chips is a key consideration.
And that's exactly what Nvidia CFO spoke about on their earnings call this week. The A100 GPUs we shipped six years ago are still running at full utilization today, powered by vastly improved software stock. Effectively, they're implying that their data center customers are still deriving value from older generation GPUs.
Burry's counterargument is that just because an old GPU is being utilized doesn't mean it's creating any value. He says A100s take two to three times more power per flop, or compute unit, so they cost two to three times more in electricity alone than H100s, which brings up another important consideration, energy costs.
Newer GPUs often deliver significantly better performance per watt of energy consumed. So, keeping old GPUs running may eventually make no economic sense purely from an opportunity cost perspective. And given electricity is expected to be a huge bottleneck, the savings from improved efficiency of the newer chips will probably make upgrading sooner the better choice.
But as intriguing as Burry's thesis is, there is one very compelling counterargument. Doubling the lifespan of tens or hundreds of billions of dollars in assets is a huge financial decision. Whether it's Google or Meta or Amazon, this change would have involved the real analysis of how productive their older generation chips are.
And the results are effectively printed right there in the public filings. So, if their internal assessment didn't show that the useful life of servers was really higher, that's just actual fraud. I'm not saying it's impossible. Uh wouldn't be the first time a company's committed fraud, of course. Uh but where's the proof, then?
Where's the paper trail? Where's the whistleblowers? Where's the evidence that Google's sitting on billions of dollars of useless chips? Of course, depreciation is really only one piece of this whole AI bubble debate. I could probably do a whole 'nother video on the circular financing mess, the energy constraints, and the sneaky under-the-table debts.
But for now, thanks for watching and consider checking out one of these videos next.
Comments 0