원문 보기
So, it's been about a year since Michael Bury opened a short position on AI. Back then, everyone only had theory. The bulls pointed to demand, the bears pointed to accounting , and both sides could look at the chart and think they were right. This is already in the past. We now have actual data for four quarters.
AMD reported. Palantir reported. Every hyperscaler has updated their spending plans. An accounting thesis always contains predictions. The validity period of the forecasts coincides. So let's check the results. But if you're short on time or simply refuse to take financial advice from a former analyst who hides his face behind an animated cat, that's fair.
Here is the main conclusion, without further ado. AI commerce doesn't collapse when spending slows. It only collapses when the hyperscaler predicts stagnation in capital spending. This is the claim I am proving, and here is exactly how . Three tests in a row. The first test is AMD, because if cost control were real, the second-tier supplier would be cut first.
The second test is Palantir, since that was Bury's biggest position, and his arguments do not apply to it . The third test is the hyperscalers themselves, because they are the ones who allegedly manipulate reporting. Hopefully, by the end of this video, you'll learn not to overreact to scary headlines and know what to look for .
Okay , let's refresh our memory and start with what Michael Bury actually claimed. But let's be brief. So, to boil down Bury's argument to a single number, this is how hyperscalers understate depreciation by about $176 billion cumulatively over 2026–2028 . And the mechanism is simple. They amortize AI chips within 5–6 years, while he claims they become obsolete in two to three.
Hopper, then Blackwell, then Rubin. If you know a little about accounting, depreciation is simply dividing the cost of an expensive asset over its useful life. But if you stretch that period too far, you reflect lower costs today. Naturally, lower costs mean higher profits, profits that don't exist organically.
As a result, Bury noted that by 2028, Oracle's profits will be overstated by about 26.9%, and Meta's by 20.8%. To understand this, let's look at the mechanism using the example of one rack of chips. This last line is the whole thesis. The money is the same, but the profit figures are different. And the price evidence backs up his words.
Renting an H100 chip cost about $8 per hour during the 2023 rush. Today it is about $1.50–$2. This is a 75% decline in the return on an asset that is written off as if it were earning income evenly over 6 years. Um, but that's not all. I need you to admit something before we move on. The strangest thing is that his largest short position wasn't even related to the chipmaker.
He bet about five times more against the software company than against the hardware. But look, his entire public argument was about equipment depreciation, and the biggest position was against businesses that have almost no equipment. Mm, think about this discrepancy. This explains everything that happened next.
I'll come back to this later. Now let's be fair. Because he's not crazy, and he really has made significantly more money than any random user on social media. But look at years three through six on this overlay. Gray color shows that the chip is still working, orange—that it is actually written off.
This red wedge is his stake. So for now, I would say that they don't actually throw away old chips. They are simply translated into inference tasks . Learning happens once, learning happens constantly. So the residual value is real, it's just significantly lower than the neat gray line. And maybe six years—that's really too optimistic.
But three years is also too tough, so I think Bury is right about the direction, but too aggressive about the numbers. The truth is probably somewhere in the middle, if you compare his $176 billion to the plausible hyperscaler spending of $2.5 trillion over the same period. This is approximately 7%.
Yes, 7% may be a real problem, but it's definitely not the apocalypse. Now let's move on to the time test: his thesis requires that the cost reductions be felt around now. So let's take a look at the suppliers who are being cut first. We don't have Nvidia reports yet , but we can use AMD. Two things on this screen.
First: the dotted line bends upwards, not straight. Second, and no one talks about this, look at the gray segment. Excluding data centers, the rest of AMD grew by about 9% . Client PCs, games, embedded systems—a completely ordinary, unremarkable chip business. On the other hand, the share of data centers increased from 42% of the company to 58% in 12 months.
If you own AMD stock, the business diversification you were counting on has already quietly evaporated due to growth, not any announcements. When hyperscaler costs change, there will be no second engine. And yet, management predicts a doubling of data center performance by 2027. Server processors just posted their fifth consecutive record quarter, with cloud and enterprise segments growing by more than 70%.
Remember, the second source of supply is always the first item to be eliminated when budgets are cut, always. AMD is the second source, and AMD gives optimistic forecasts. This is stronger evidence against short-term cost containment than anything the market leader could report. Next, using the example of Palantir, the company he hated so much, see how this safe fills up .
63 cents of every dollar of revenue turned into real cash. And remember, depreciation shenanigans may inflate accounting profit, but they cannot create cash. In addition, there are almost no depreciable assets to take into account. This software literally runs in someone else's cloud. So Burry's accounting criticism doesn't apply to this company at all.
This is decided , it is not up for discussion. Meanwhile, his Palantir put options rose in value by about 35% at their peak, while certain chip puts turned out to be unprofitable, which directly contradicts his own arguments. So, he simply made money by squeezing the valuation of stocks whose price was higher than ideal.
He lost to the equipment manufacturer, which was actually his main point. Profitable and false at the same time . Which is honestly the most "cat-like" behavior I've ever heard of. Push a glass off the table, land on your feet, and then walk away, as if that's what you're supposed to do. And now what I kept quiet about.
The reason why the gap in position sizes matters. Everyone raised their forecasts. " Bulls" call it reputation. The “bears ” point to a yellow signal: a drop in growth from 83% to 39%—and call it a peak. Both are wrong. Look where these two lines intersect. There have been reports that hyperscalers are now raising external funding as AI spending outpaces operating cash flow.
So, why is this important? You see, for years the whole "bullish " case for big tech companies was that they were self-funded " cash machines," but that doesn't seem to be the case anymore. When Amazon and Alphabet take out loans to build data centers, expansion ceases to be a matter of internal budget and becomes a matter of the credit market.
And credit markets change their minds much faster than capital budgeting. This is where Burry's accounting remark ceases to be a curiosity and becomes a structural issue . Not urgent, but structural. Overstating earnings doesn't matter much when you're funded by strong free cash flow. This becomes more important when part of the financing comes from lenders, as inflated profits inflate the coverage ratios that these lenders underwrite.
Underestimating depreciation and switching to external financing are not two separate stories. This is one story, and 2026 is simply the year when these two halves met. They met, but they didn't explode. Nothing in these four blocks indicates stress. Costs are rising, forecasts are optimistic, demand is outstripping capacity.
Only the source of funding has changed , and that will only matter later, when growth slows. Now it's just a state, not a crisis. So what to actually do? Mostly nothing, and that's the real answer. First, don't downgrade to hardware. Slowing costs from plus 83% to plus 39%—that's the middle of the game, not the ninth inning.
Nothing in the 2026 figures suggests otherwise. Your signal is hyperscalers' forecasts for stable capital expenditures. Set up notifications. Until this happens, you are in a hurry, and haste is just a mistake with additional interest payments. Second, if you own AMD, know exactly what you own. This is an AI infrastructure company with a legacy PC business attached to it.
This is not a warning. Business is showing success. Third, monitor the level of debt financing. Don't panic about this. Neo cloud services, running on 2023 chips for $8 an hour, are now selling those calculations for $1.50. This is where Berry's math is mechanically correct. There is no evidence that this will break in 2026, but it is the shortest fuse in the system and most of you do not have direct possession of it.
If it ever breaks out, it will change the price of the assets you own. And one more thing , because that determines the discount rate for all of this. I still lean towards no rate hike in September, despite recent market speculation. So, relatively cheap money supports a comfortable level of leverage, which largely explains why 2026 looks calm.
This is my vision, not something that can be solved with these numbers. If I'm wrong, the time frame will shrink. If this was helpful, a like would really help this one-year-old little channel become more visible to those who need it. Subscribe if you want to be notified when the fixed capital expenditure forecast comes out, because I'll be keeping an eye on it. Anyway , that's all for today, see you in the next video.
Comments 0