저자: Kevin · 외국인
요약 보기
CDS 상승 등 신용 스트레스로 네오클라우드(CoreWeave, nScale)가 흔들리지만 하이퍼스케일러 투자는 지속된다고 봄. 이란 협상 타결 시 자금조달 비용이 낮아져 AI 인프라 전반에 우호적, 결렬 시 네오클라우드 중심으로 악화. ASML 계약부채 급증과 2027년 약 50GW 데이터센터 가동 전망을 근거로 칩 수요 강세 유지, 메모리 가격은 내년 하락 가능성. AI 실제 업무 능력 확대로 노동시장 충격은 수년 후로 전망하며, 버블 가능성은 인정하나 현 시점은 랠리 초입으로 봄.
다듬은 글
Another day, another standard swap schedule. Yes, here is Oracle literally to the moon. But this isn't the view of the Moon you want. It's not just Oracle, it's also Coreweave, of course. Coreweave is even more on the moon. Just when we thought it couldn't get any worse , it got even worse. But remember, this is the loop we were talking about.
And so we need to know, is this really the end? Is this just part of the process to get to where this economy is going, where going? And what could stop all this much sooner than you think, waiting for the chips to go live and waiting for Anthropic's IPO. Let's talk about all this and figure it out.
First , obviously, and it's worth starting with this, because it's one of the most important components. Before we talk about building a data center, we know the following. We know that when credit default swaps rise rapidly, the underlying company's stock price falls, its bonds sell off faster , which leads to more investors being drawn into the credit default swaps, which in turn prolongs the cycle of selling these companies' stocks and bonds , which increases the cost of capital for these companies to raise funds to finance AI development .
Now the richest companies, such as Amazon, Microsoft, have no problems. But these smaller neoclouds, like Nscale, which just lost a deal with Microsoft in the UK, these companies need lower returns. So how do we resolve some of these anxieties faster, not slower? Obviously, it would be good if we could make a deal with Iran.
In my opinion, the deadline for the agreement with Iran is between October 20 and 27. And there were rumors that perhaps a deal was coming. There's this report here in New Arab that says, and I quote, " An exclusive breakthrough in mediation efforts between the US and Iran is expected this week." Who knows?
Perhaps this is just an attempt by the Trump administration to spread a little headline to stabilize the markets. No one knows for sure. But the comment here is this: sources, I quote: "People familiar with the negotiations say that Qatar and Pakistan are continuing their diplomatic efforts to prevent a resumption of war." They say talks are continuing to narrow differences , encouraging mutual concessions aimed at preventing another military confrontation.
This morning on Al Jazeera we saw notes from Aragachi's interview, which claim the same thing. That the key is to prevent further military strikes, and that they prefer to work on political and diplomatic solutions. This is coming from the Iranian side, and it is consistent with what we are hearing here.
Now , could this all be for Fugazi? Of course. People ask me, " Kevin, why do you think the deal will happen?" Because if there is no deal, the financing costs of these AI companies will potentially increase. So far, we've reached the peak return that we were talking about at about 527, you know , give or take a little over 10 years.
So far, so good. Knock on wood. But if we go back to war, and we go back to conflict, profitability will deteriorate, then the AI trade will deteriorate in the short term. Everything will go awry. Then people ask, “Kevin, why are you so optimistic about, first of all, data centers, we’ll talk about this data center construction in a moment.
But why are you so optimistic about the Iran deal? And I think the easy way to sum it up, the Iran deal, is that Donald Trump has already said he's willing to sign the deal that Iran has proposed . Iran suggested, hey, we want the Strait of Hormuz to reopen, and then we'll negotiate on the nuclear issue.
Donald Trump said he would be willing to sign this agreement last year. So it was a framework agreement that could work. A concept plan that could work. Iran is desperate to get air traffic flowing. They want the cargo to arrive. They want trucks to be able to freely cross borders from Pakistan, Iraq, and Afghanistan.
They want to be able to trade with the ports of Karachi. They can't do anything because of the naval blockade, the land blockade, and the restrictions on air travel. So they are really being crushed, and that includes blocking their ability... to get the raw materials to produce more drones and weapons.
Donald Trump says, "Hey, well, listen, but since we've got you in a bind with this blockade, we need a little bit of a better deal." And that includes nuclear talks, meaning let's not re. Why don't we talk about nuclear weapons now? And that's exactly how we stand. So I think that once Donald Trump gets enough of this, conveniently before the election, we could see a deal.
But you don't have to take my opinion on this. We can skip this. I just wanted to explain to you my thesis about why we think that way. We can be wrong. And obviously, if we're wrong, the construction of AI data centers will suffer more, but there's still good news. Because even if companies like CoreWeave and some neoclouds like nScale have issues like this, you have...Microsoft is removing nScale UK data.
center agreement after delays, delays in electricity supply. “ The richest of them are still expected to invest in data centers, which is really good for the overall AI trade. This simply creates more uncertainty about the timing. If we make a deal, great. Then everything can develop rapidly. If we don't make a deal, we'll forget about the little guys.
The small guys—the neo-clouds— will have more trouble without a deal because financing costs will remain high for too long, meaning that only the big guys can really play. Microsoft, Amazon, Meta, Space X. So to simplify, maybe the investment strategy around this, neocloud is basically untouchable without an Iran deal.
So if I'm talking about neoclouds, these guys need a deal with Iran because it will reduce the cost of funding that has already flown to the moon, okay? Then we can move on to hyperscalers. They will feel good with or without a deal. So hyperscalers, uh , are suitable in any case. So, again, Microsoft, Meta , Amazon, the big guys, you're safe here because then you..." is not dependent on the Iran deal.
You are not dependent on financing costs falling in the short term. Well, the analysis of semi- finished products, and they can sometimes have a very optimistic view of the semi-finished products industry, I feel like that's the case, but I actually agree with that analysis. Uh, and that's consistent with what I've researched elsewhere .
This idea that the construction that we're going to see now in the data center space is essentially rebuilding every data center built in the world from the beginning of time until 2024. Restoring all of this in terms of electrical capacity in 1 year, just in 2027. This is about 50 gigawatts of power.
This is equivalent to approximately $2.5 trillion in spending. It is not necessary that this $2.5 trillion must be spent during 2027. Just because this energy comes online doesn't necessarily mean the chips are already available. It also means that some of that money was spent in the previous year, you know, plus or...
minus. So, you're roughly limited to 2027 here. But semi-analysis suggests we could see 50 gigawatts of capacity coming online in 2027. They already show that 39.4 gigawatts are planned , or, sorry, being built and commissioned in 2027, roughly by 2027. Of course, delays are possible. Plus another 11.9 planned, bringing us to over 50 gigawatts that will be commissioned in 2027.
And this is where they argue that the next big shortage won't necessarily be data centers, but it will potentially be a glut of data centers because we have a shortage of chips. So as soon as I saw this note about chip shortages, I started thinking to myself: well, what do contractual obligations look like at ASML?
Now, ASML, if you don't know, they make the machines that companies like TSM or Intel use on these two- and three-nanometer advanced chip packaging lines. This is where you're going to get your memory, your Nvidia chips, whatever. Your cutting-edge AMD chips, you're going to get them on some of these smaller nodes, the two-nanometer and three- nanometer node sets.
A larger chip like Cerebrus is running on the 5- nanometer line, you know, which hopefully will allow them to increase supply , but you know, this company really needs to get their supply up and running, which is expected in 2027. They can be a significant part of this process. But looking at ASML, you can look at a certain portion of their revenue, especially their annual report, and you can look at the contractual obligations that relate to the remaining performance obligations, essentially, of their machines.
In other words, people send ASML money as a deposit. ASML says, " Okay, that sounds good. We'll make the machine for you and deliver it to you when it's ready." "These contractual obligations started to grow rapidly. If I look at the financial statement, I'll see the total contract outstanding, uh, where are we each month?
Where is she? Where is she? Where is she? I want to show that I moved it because it's ours...Yes, I think it will work. Uh, contractual obligations, I don't know where I put it. I swear I had her, and then...Oh, there she is. She's right here. She's right here. Okay, I was...Okay, perfect . I knew I wouldn't go crazy.
Contractual obligations, look at that. So, September 2025 - 3.5 billion euros. December - 3.3 billion euros - slightly down. March - 3.5 billion. June - 3.5 billion. September - 9.3 billion. What? The absolute rapid growth of ASML's contractual obligations. This indicates that something has changed in 2026, when suddenly people wake up and say, “Oh, my God.
This boom is about to begin in earnest ." “And let me make a side comment, by the way, about artificial intelligence and what we’re seeing. We personally don't know, you know, where he was. It was probably somewhere around , I don't know, maybe the last 3 months here, but it could have been a little earlier.
We've started to see a shift in how functional AI actually is. And even in the last few weeks, it's become even more extreme, where things that AI couldn't do intelligently before , AI can do very well now. Some of these things that have really exploded recently are things like video editing or even videography.
So it can be both filming and editing. What's happening now is that people are installing, for example, 360- degree 4K cameras. And the AI essentially chooses the exact angle from which to shoot, based on the location of the tennis ball or based on the location of the soccer ball, a passing car, or something else.
AI is filming it, but AI is not only filming, AI is now editing it. Where you could actually put a 2-hour podcast, you can now put it in Adobe, you know, whether you're using Adobe Premiere or Final Cut Pro or something else, and artificial intelligence, and we tested it, and we thought: this is crazy.
AI could probably do about 95% of the editing for you. Changing the angle, who is speaking , shortening scenes, reducing dead air, whatever. And what's happening is that AI is actually becoming much more functional every day in places that people didn't think it would become functional as quickly as it has.
So we've seen this coding moment, which is now starting to turn into a web design moment, a video editing moment, and a healthcare moment is coming. All these additional changes that Jensen himself talks about. You know, Jensen says we had to go through an era of hallucinations, which is actually about 24 and 25, to get to the breakthrough moments.
And now we're starting to see these breakthrough moments. We see it too. And we think the boom is still at its beginning. You know, eventually this will all end...it will be over- inflated and a bubble, but we still think the boom is just beginning. And this line from ASML is probably one of the most important ones we could pay attention to.
And if you see a difference even after 3 months, or when they report on their backlog , because this is one of ASML's problems, they don't report on their backlog, but only once a year. So we only learn about their backlog once a year. And in the meantime, you know, from time to time we can make assumptions just based on how their contractual obligations are moving in terms of the amount of deposits that they're receiving.
And some institutions on Wall Street, I believe it was Barclays, but maybe I'm quoting them incorrectly . Or reports, I read this this weekend, are reporting that yes, it actually looks like ASML's backlog is going to be significantly higher than people on Wall Street are expecting right now. Okay, that means more people are buying these machines to produce more chips.
Why? Well , this is about what is being discussed here in the semi-analysis, which is a major shortage of silicon. You see, they say we see a risk of a 10- gigawatt silicon surplus in data centers by the end of 2028. It means, "Oh my God, we don't use it enough." This is bad . This is the end of the bubble.
But what they really mean is, "No, it's being used. " This is not the end of the bubble. This simply shifts the problem to a shortage of chips. In other words, we saw this in mid-2025 when Oracle was unable to source enough GB200 racks. And you essentially had these empty buildings with power available, but you didn't have enough chips to install them.
Then it turned into us not having enough available power. Now they think we'll once again face a situation where we simply don't have enough GPUs . And so we're going to face this crazy shortage of GPUs again . So I made a small update to the ASML assessment. ASML currently has a price/ earnings ratio of around 35, and expected growth is around 24.
So, the PEG ratio is around 1.4. The lower, the better. Nvidia almost turns these numbers upside down . Nvidia, uh, has a price-to- earnings ratio of about 23 based on the projected P/E at the end of this year. And their growth is expected to be around 35. That puts them at a PEG level of around 0.65.
That means if it's true that we're about to move on from , uh, you know, these fears around neo-holarders, who is the most sensitive to interest rates, right? For example, your CoreWeave is your most interest rate sensitive construction type. This is actually Oracle CDS. They are less sensitive. Their balance has improved a little.
This is CoreWeave , still bad balance. These guys are definitely expanding only through debt . And that makes them the most sensitive to high stakes, which is obviously detrimental to the Iran deal or no Iran deal. If a deal is made with Iran, they can all win. So, neoclouds and hyperscalers can win.
Without a deal with Iran, the neoclouds will continue to have problems. But the reason I mentioned this at the beginning is that just because neoclouds are suffering doesn't mean hyperscalers will slow down. Actually, look at this. SemiAnalysis says: “ Things have changed dramatically since the second half of 2025.” So basically between January 2026 and now, in the last 9 months, I quote, "Things have changed dramatically." Oracle ushered in an era of giant leases, signing 4 1/2 gigawatts of IT data center capacity across three sites.
Everyone else followed suit, and the pool of buyers has never been wider. Meta buys, Amazon buys, Microsoft buys , Google buys, Nvidia buys, AMD buys, Anthropic buys, OpenAI buys. They're all spending money on these massive gigawatt-scale deals . Neoclouds is actually a small show. You know, they sign these deals for 100 megawatts, which is still a lot.
But even if CoreWeave goes bankrupt, who cares? The boom is still ongoing thanks to relatively wealthy hyperscalers. And this is where SemiAnalysis claims that data center rentals have skyrocketed, which is a key indicator, and that expanded construction, which they identify with their satellite imagery, is also skyrocketing.
So , here's just an overview of some of the different sites they track. It's worth knowing that anyone can buy satellite imagery, it's just expensive. For example, if you want to zoom to this level, you'll probably pay around $1,500 per shot. So if you make $1,500 per site*50 sites, that's $75,000 per shot.
You know, if you take a picture every 2 weeks...Hell, that's $150,000 a month for satellite imagery. Yes. That's a lot of money, is n't it? So, I mean, anyone can get satellite imagery, or just fly over these places with a drone. It seems like that's all that can be done. But you do n't just need satellite imagery.
You can look at many clues. You can look at the costs that hyperscalers incur. You can look at these ASML contractual obligations. You can look at other people's satellite images that they paid for. Uh, and we can see that yes, even with this short-term fear about CoreWeave, I think the fear that this is it, this is the end, the bubble is bursting, and the fear that this is close , I think, is greatly exaggerated.
Currently, we are still experiencing a huge memory shortage. I think the situation will change next year. I think we will start to see DRAM prices decline next year . And obviously, you know, the CMTX moment in China is not helping to stabilize prices, but it's actually good for the rest of the industry.
This means that Google, Meta, Amazon, Nvidia, AMD, and all the other players can benefit from more accessible and cheaper RAM. Maybe not necessarily cheaper, but, you know, prices that are n't rising so rapidly anymore . They touch a little on the D specification that's going on here. So, the D specification is sometimes a little difficult to understand, but I'll show you here.
Here they mentioned that the Rubin Ultra now comes with 172GB of high-speed memory per GPU compared to 288 on the previous device. And a long time ago, that was probably the case, maybe not so long ago, but around the time memory stocks were really showing their best performance at the beginning of the year, there was a belief that we would see memory chips as large as 1GB in these Nvidia GPUs.
And they create a small problem. Most of the problems are when you link them together, and after running the model, let's say you need to , oh, I don't know. Let's say you need 700 GB of RAM for the model. Uh, let's write this down. So, launch like a great Meta Muse model, let's just say. Okay, let's say we need 700 GB of RAM.
After we loaded the model, uh , let's call it a rack with eight graphics cards. Okay, and if each of them has 192 GB of RAM, uh, 192*8, we have 1536, which is about half. So this will load our model. And then this allows us to actually do things. So this will be a place where we can, you know, have our caches for our conversations, our KV cache.
This will be a place where we can, you know, handle context, essentially. And then we can also break it up into little parallel programs where people say, “Hey, you know, you’re going to use some of this. You're going to use some of this. You're going to use some of this. You're going to use some of this." The thing is , the reason I'm talking about this is that you don't necessarily need that much RAM per chip.
And if you can bundle these puppies together, load a model, a big heavy model, whatever you load, the Frontier model, and then you can split the rest into, well, we can run queries from 10 people at a time on that model. And so a lot of people here are saying, " Hey, you don't necessarily need to use such a heavy set of RAM on a GPU." Why do n't we reduce, uh, how much RAM we have per GPU?
And , uh, this way, we can actually print more GPUs .” This is really becoming more recognized now, which I think is a big advantage for Nvidia in the long run. Here we are also talking about quad-core and twelve-core drives. This largely depends on bandwidth. In other words, how high do we place DRAM?
Do we really need to put these twelve-core drives in? No, because we could have a chip with faster bandwidth , which would reduce our need for even more of that RAM. So the bottom line of all this, I think, is that on the front end, once you load a model onto these advanced chips, you don't necessarily need as much RAM per chip.
It's better to have more chips that can run more processes for other people. And another thing that kind of reduces the weight is once you start moving to some of the open-weight models or the compressed models, you know, this one here, instead of 700GB, might end up being 32GB, you know, or 70GB, or whatever.
So, the models are significantly reduced in size. So that makes me more optimistic about broader development with chips like AMD or Nvidia chips, or basically anyone sitting next to RAM. So, how do you put all this together ? Good. All of this together, in my opinion , is due to the fact that a significant part of the fears surrounding CoreWeave and Oracle are directly related to the surge we have seen over the past 10 years.
It is now 5.13. He sneaked up a bit today. When you enter into a raw deal, you almost immediately eliminate many fears about the profitability issue. If you don't, neo-clouds will continue to suffer, but the big players will continue to spend. They spend, spend. RAM prices may stabilize, which is actually very good for chip manufacturers.
And if we bring the developers together, I would say. And if we combine these advanced chips like the Rubin Ultra, you know, the processors that Nvidia makes, or the advanced AMD chips, or even the ASICs, the accelerators that we see at Google in partnership with companies like Broadcom or Marvell. Meta makes them too .
Putting it all together, everyone is trying to print more chips. And I think that's exactly what we're seeing with ASML . There are no signs of slowing down here. The only place where there are signs of pain is neo Clouds. That's all. Wherever there is money, more money goes into chips. Again, Meta, Microsoft, Amazon, Nvidia plus Anthropic and OpenAI, obviously.
So, whether it's custom chips like ASICs or cutting-edge chips from AMD or Nvidia, they all benefit from stabilizing memory prices. Although recently this has been causing some concerns. Oh no, the stocks of memory companies will fall. Who cares that it won't drag the entire market down with it. And yes, neo Clouds are suffering right now.
But it also won't drag the entire market down. Do I think this will all end in a bubble at some point in the future? Yes. But, I think that between now and then, we have another wildly euphoric rally ahead of us. Uh, and really, this debate about expanding capacity to 50 gigawatts, uh, potentially in 1 year—it's a wild thing.
And that will probably happen. And this expansion of chips is just beginning. Mostly because the applications for which AI becomes useful are only now being implemented. You know, I was in this, uh , conversation with this guy from Wall Street. And the question was, "Hey, well, you know, why has n't the job market turned around yet?" And their response was, "Well, you know, maybe if it was meant to happen, it would have happened by now." In other words, if the labor market were to be turned upside down by artificial intelligence, it would have already been turned upside down.
And I actually completely disagree with that . I think between the ages of 23, 24, and 25, we had a period of hallucinatory AI where we thought, “Yeah, he’s good for some things. This is a cool encyclopedia. That's all." The year 2026 has changed. And my opinion on this has changed and evolved a lot, I have to say.
Now we're actually starting to see it can perform functional tasks, whether it's accounting , video editing, videography, website design, app design, and more. It is capable of doing much more than just coding. This is a bit scary for the labor market. Therefore, I think that the real shocks in the labor market are still ahead.
They will probably happen in a few years. So I'm saying that all this fear that people had this weekend was, " Oh, Black Monday, everything's going to hell." It's still too early. It's too early. So I need to make some discounts. Want to see what we do, .
Comments 0