저자: Kevin · 외국인
요약 보기
스페이스X 실적 발표 후 주가 급락을 계기로, 상장사 보고 수치(데이터센터 구축비 $2.7M/MW)가 GPU 미포함 '쉘' 기준이라 과장되었다고 주장. 구글·앤트로픽 간 순환 계약(높은 임대가, 90일 해지권)과 구글의 SpaceX·앤트로픽 지분 구조가 주가 부양 목적일 수 있다고 의심. 현금 소진(6~9개월 내), 설비투자 85.8% AI 인프라, 메모리 공급 병목, 팔콘 로켓 축소 등을 근거로 스페이스X 주식에 회의적 입장. 예측 목적 발언은 아님.
다듬은 글
Something really strange is happening with SpaceX . And we need to talk about this because SpaceX reported earnings yesterday. Today the shares collapsed. And tomorrow the lock-up period for a large block of shares expires, which means that potentially more people will start selling them. But the real oddity, which I want to start with, looks like a small graph.
And this is not a stock price chart . Instead, it concerns the sale of computing power, which is what SpaceX seems to be spending the most money on right now. In fact, about 85% of all the money they receive, they spend not on rockets or Starlink V3 satellites, to make Starlink 10 times faster. No. They go to data centers.
Yes. Micro hard and mini hard. But something strange is happening, and I think I have the right to ask a few questions. Let's start with the graph. Let's say you have an annual income on the left side of the graph . OK? From your data centers. This would be measured in megawatts per, you know, month, let's say.
So let's call it MW per month, okay? And it will be in millions of dollars. So let's put a small letter M there. Let's say you make $4 million for every megawatt of computing power you sell. And the scale is from 0 to 30. And let's say you earn a four. Well , that would put you somewhere near the dotted line, right?
Pretty low, pretty close to the bottom. That's understandable, isn't it? Now on the X-axis we will have the cost of building these facilities per megawatt. And it will be in millions of dollars, and it will also range from 0 to 30. And in this case, let's say the average cost for companies like EBIAS or Coreweave is about $15 million per megawatt.
Well, then you would be somewhere in the middle. So we would have an intersection point somewhere here, right? So the average data center should be somewhere low and in the middle of this graph. Now let's assume that you are a real phenomenon. Let's think about this for a minute. A complete phenomenon: you spend all your money on the best chips in the world and literally pay the highest possible markup.
And at the same time you receive average income. Well, in that case, the average income would be down here on the graph, and you would be paying, well, a premium price, you would be right here on the right, okay? So, this is an average value. This would mean paying a lot and earning very little . Now let's assume you're a different kind of weirdo.
Let's say you somehow managed to convince the world that you build data centers much cheaper than others. And yet you managed to earn six times more than everyone else selling the same thing. That would be crazy, right? In that case, you would end up somewhere here on the graph, right? That would be madness.
So, if you are overpaying and underearning, you are here. If you are an average player, you are here. And if you're, well, crazy, I guess, or grandly crazy, then you're here, at the top of the graph. Okay . So why did we just figure out where you might end up on the chart? What if I said there was something happening with SpaceX that put it in a unique position on this chart?
What we would call a statistical outlier . Let's look at the actual graph. Wow . So we have Coreweave here, with the usual construction cost per megawatt. Nebius is here, with a typical construction cost per megawatt in this mid-range . And the earnings are somewhere between , say, 2 and 5 million dollars per megawatt per month, right?
Well, actually, in this case it's an annual figure. It doesn't matter if it's for a month or a year, the point is clear. Of course? SpaceX is right up there, earning six times more than its competitors and building five times cheaper. This is a statistical anomaly . Let's look at the real data. SpaceX says it is building for $2.7 million per megawatt.
Well, okay. And their annual revenue is about $30 million. Annual, I said "per month" on my board, I meant annual. Nebius collects revenue of about 4.5 million. Coreweave collects about 1.65. Core Scientific collects 1.75. The average hyperscaler collects 2.3. So, somehow SpaceX makes more money than all of them.
And they buy all this much cheaper than others. This intrigued me a little . What is going on here? How is this possible? Is this the “ Elon Prize”? No. Is this hatred for Elon? No. Let's analyze what is actually happening. You see, during the reporting conference, Elon Musk stated that they were able to implement computing power with a payback of less than a year, which seemed like an incredible statement, so I decided to test it for strength.
And then I found out where that figure of $2.7 million per megawatt actually came from. This is not an auditor's report, not an SEC report, not an S-1 form, it is not in any official document . This is simply the amount for which SpaceX claims they were able to build the Colossus 2 data center, not including any GPU equipment.
It is literally just a " shell". No audits, just management's opinion on how much it costs them to build the building itself and the energy infrastructure. So is this a lie? No. But it lacks a lot of context. See, Deutsche Bank claims that to fill this center with GPUs, the cost per megawatt would be between $25 million and $30 million , which is essentially in line with industry averages.
Hmm , interesting. So, in that case, what is Colossus 2? Well, Colossus 2, like Colossus 1, are actually considered brownfield data centers . A brownfield data center is different from a greenfield, and it has nothing to do with green energy. These are essentially data centers built inside existing structures, such as an old manufacturing facility that was previously used for other purposes and has now become a data center.
This is exactly what SpaceX did for Colossus 1 and 2 when it was the XAI project. They took over existing factories with ready-made power grids that are easier to license, although Elon still had to use unauthorized gas turbines, and SpaceX had to promise to remove those illegal turbines by July 2027, which is about a year away, so time is running out.
It's a completely different story in terms of speed and cost than taking a blank piece of land, going through permitting procedures for 18 to 24 months, compared to the 91 days for the rushed construction of Colossus 1 and 2, with the faster of the two, Colossus 2, being built in 91 days, which is admittedly quite impressive.
But when you hear about $2.7 million per megawatt, you have to understand that this is a brownfield type deployment inside an existing facility, where and communications were already in place. This is not building from scratch " on a vacant lot." And that doesn't include GPUs, servers, switching equipment, or other things that actually make a data center work.
This is really important because the marketing used by many SpaceX investors is based on the figure of $2.7 million per megawatt. And that will significantly skew the ratings for many people. So, this is important. To be fair, this is just the first part. Because even if it costs, say, according to Deutsche Bank's estimate, an average of about $27.5 million per megawatt to deploy, let's say, mini or macro equipment.
Of course? Let's say you were able to implement this. If Anthropic is willing to pay $30 million per megawatt per year, then Anthropic effectively pays SpaceX back its entire investment in less than a year. That could still be justified, right? So is this good? Does this mean we are ready to work? The graph meant nothing?
Not quite. You see, the problem with this graph is that it contains two worrying things for SpaceX . The first is what we started with, namely marketing around construction costs . How did we end up on the far left side? Oh, I understand. We're not really on the far left . If we want to use cutting-edge hardware and really consider GPUs, we're actually right here on the far right .
But the marketing implies that they are here, on the far left side. Okay, that's misleading, but hey, a little more research—and we'll figure it out. But there is a second problem. The second problem is how can SpaceX have such high revenue when everyone else has such low revenue? To understand this, we must understand incentives.
First, Colossus number one was leased to Anthropic. The entire object. Colossus number two was partially leased to Anthropic. And that's partly because SpaceX says they'll focus less on training and more on inference. And then they leased the Anthropic portion at a premium price, which we'll talk about in a moment.
There were rumors that Grok had latency issues. And that SpaceX actually couldn't properly use these facilities to train Grok, but that's just a rumor, so let's leave it at that. This was reported by The Information, a publication often considered reliable, but these rumors were never confirmed or denied.
They just disappeared. So, the revenue level under the Anthropic contract is $30 million per megawatt per year. This still provides a payback period of one year. That's about 10 months to full payback, right? If you consider the entire computing infrastructure. Google then signed a contract with SpaceX to lease some of their additional computing capacity for $50 million per megawatt per year.
Which is really interesting, because 50/30 is 67% more expensive than Anthropic. Google agreed to a price that was 67% higher than Anthropic's, even though Anthropic was already paying 6 times the market price of hyperscalers like Core Scientific, Coreweave, or Nebius. And why is this so? Is it simply because they want it and SpaceX has it in stock?
“ Come and get it, but you ’ll have to pay extra, ” or maybe Elon is just a really good bargainer? Perhaps. Or is it something else? Google owns $94.1 billion worth of SpaceX shares. This is as of June 30. Of course, the valuation has changed significantly, but in Google's revenue came from the rise in SpaceX's stock price.
Of this $94.1 billion , $80 billion is locked in on a short-term schedule, and $14.1 billion is locked in on a long-term schedule. So, many assets are frozen. Most of Google's net profit that quarter of $114 billion came from the rise in SpaceX shares. In fact, this was the first quarter that Google went into the red on cash flow, and people are saying, "Well, this is an anomaly." That's exactly how it was described.
But, despite Google's revenue, this video is not about the company's financial statements. We know that Google benefits enormously if SpaceX shares soar in price. Google also has a "friend" who rents capacity from SpaceX. In a moment you will understand why they are friends. This is Anthropic. Colossus 1 is leased by Anthropic.
Google owns a $ 124 billion stake in Anthropic , which is almost 30 % more than their stake in SpaceX itself. So, let's put this together. Google makes money when SpaceX shares rise. The fact that Google is overpaying SpaceX on a 3-year, cancellable contract could help SpaceX shares rise. Google has $15.3 billion in backup for an unknown customer's data center in Hubbard, Texas.
So who is building a data center in Hubbard, Texas? Anthropic. Anthropic may need financial support. Maybe you'll need money too. If Anthropic makes money, they could also overpay SpaceX for Colossus, which theoretically pushes SpaceX shares up and helps Google . And if Google gets bailout and Anthropic does well , then Google wins again because it owns a bunch of Anthropic shares.
So you have a closed loop that explains this anomaly. So the marketing figure of $2.7 billion per gigawatt does not include all GPUs. The reality is that they are not building these facilities cheaper, and even if they could make the frame 10 % cheaper, they are using a repurposing strategy to speed up the process.
And when they try to scale it up and go to building from scratch, meaning taking an empty lot and turning it into a facility, it takes much longer. This creates cost problems. So, cost savings are an illusion. You are actually at the level of the average hyperscaler. You are no different in these expenses.
Jensen doesn't love you enough to give you a huge share of his GB300 chips at a discount. This will not happen. And these premium rates, actually, we don't know if that's true, but they seem to come from people who are significantly benefiting from the success of SpaceX's stock. The problem with such agreements on inflated terms is their extension or, worse, cancellation.
You see, Anthropic leased 100% of Colossus 1's capacity with a 90-day cancellation right, meaning they could give notice of termination and exit the game at any time. There is a kind of initial waiting period. All of these initial waiting periods expire by the end of the year. Although one of them may end even earlier.
I'll talk about that in a moment. Google, well , and Anthropic are also leasing, in addition to 100% of Colossus 1 with a 90-day cancellation right, they are leasing part of Colossus 2 with similar terms. Google also rents from...So that was about Anthropic. Google also leases from SpaceX, and they too have a 90- day cancellation right.
So, essentially, 90-day cancellation rights are everywhere. After introductory periods—90-day right to terminate the lease. The problem is that Google can terminate the contract early if SpaceX does not provide functional access for Google's deal by September 30th. It's already 56 days later. Now the problem is memory.
You see, Elon doesn't want to lose this contract because it will hit their stock, which will hurt Google. Therefore, Google may not terminate the agreement. But Elon is already building a narrative about force majeure: they say we were unable to provide access on time through no fault of our own. How does he do it ?
Here. Then I look at the logic and memory production rates. And one thing has always been considered a deterrent here. Currently, the limiting factor is memory. Memory output is growing by about 20% per year. Normally this would be fantastic. But if demand increases by 20%...no, demand does not increase by 20% per year, it increases by 200% per year.
So, in other words, yes, RAM prices are going up. We already know that. They have grown, you know, 4 times compared to last year. But it also casts some doubt on whether he will be able to fulfill that contract with Google. Perhaps. Maybe I'm putting too much meaning into this. Of course? We cannot get memory.
But it's also worth thinking about this. A significant portion of SpaceX's capital spending will go to, at least, micro and mini hardware, Elon just talked about this at X. He said, "We're going to install"—and in the earnings call—"we're going to install 220,000 Nvidia GB300 chips and 800G network cards." Okay?
These are your Infiniband connectors for these data centers. Each rack holds 72 GB200 chips. So, this means we need 3056 racks per facility. Each of these racks costs about $10 million. A rack for 10 million dollars. Damn, this is a " hot" rack. Literally, they have to be cooled incredibly hard. But in any case, that's $30 billion just for the racks.
Devil. That's a lot. Uh, and of course you'll have to spend money on InfiniBand. This will cost you one and a half billion. You need to, you know, build a facility. It will cost between 4 and 8 billion bucks. You need to purchase storage. It's going to be ...and I don't mean, you know, for toilet paper, but real, uh, you know, electronic storage.
So that's another 1.5 billion bucks. So your estimated cost for the facility would be around $40 billion . You're going to spend about $80 billion on micro and macro hardware, which is really just a pun on Microsoft. The problem is that the only way to make it pay for itself in one year is if Elon can set such crazy computing rates that no one else charges and that no one else charges.
So while it's possible that they could do it, given this initial information that I'm seeing, I do n't even like to look at Wall Street speculation because a lot of the suits, Goldman Sachs, Morgan Stanley, a lot of these people have to stay on good terms with Elon or they won't be invited to the earnings call anymore.
They will no longer be invited to events. That's how it is. This is how they play this game. And Goldman has to be at Elon's events. JP Morgan should be at Elon's events because they have a lot of clients at the bank who love Elon and want information about him. They want to feel like their bankers are there for them .
This is endless marketing. Every time they answer that call. Oh, here's your analyst. Well, look, do you understand? Our next question comes from Edison Yu from Deutsche Bank. Your line is open. Marketing. Every damn time. Marketing. That's all it is . So I'm a little suspicious of what's happening with SpaceX's AI revenue.
And this is where we can move to profits, because you see that profits in AI have skyrocketed . They grew from 737 million to 2.5 billion. This was largely due to a jump of $1.6 billion, which is a very strange flat number, since we started offering cloud services to customers. Okay, obviously it's a fixed number because it's probably one or two customers who made a fixed deposit.
It's likely that Anthropic and Google made some form of deposit. We don't know for sure because they don't give us more clarity. Okay, great. So we are increasing our income here. We are obviously still losing money on AI. You know, I don't think anyone expects SpaceX to be profitable in terms of cash flow or net income tomorrow, and they certainly aren't.
You know, they've burned through almost $5 billion in the last 6 months. So there is a net loss . Their negative cash flow here is about $25 billion over the last 25 months. And if we look at their balance sheet, I'll give them credit. They did a great job during the " suction". During that period, they raised essentially over $85 billion , and then they came in and raised a refinancing facility, or as I should say, a line of credit after that.
They raised a lot of money. So here we see $36 billion in debt, $13 billion in accounts receivable, about 50, right? And then, after I subtract that from the marketable securities and cash that they have, it would be fair to say that they have $50 billion of free cash. This will probably be enough for the next 6 months of capital expenditures.
So I assume they will be spent within the next 6-9 months. So, over the next 6-9 months, they will burn through all their spare cash. And this free money will go towards the construction of facilities, which, according to Elon, will pay for themselves in less than a year, taking into account income from courses.
But when we analyze these contracts, we have strong suspicions that they are being entered into not because they really need this Colossus computing power, but to artificially inflate the value of SpaceX shares, which makes overexpansion dangerous if you expect to lease this equipment or GB200 at these prices.
So, call me a skeptic, but I think this is something to be approached with skepticism and some trepidation. So, I don't know what to make of these predictions about reaching a hundred billion dollars in annual revenue, or that we'll eventually reach a trillion dollars in revenue. To me, this is hard to justify when you delve into some of the notes I made earlier, namely the contract termination terms and incentives to overpay compared to what we see in Nebius and Coreweave.
And keep in mind what kind of business they are in . This is a business with extremely high capital expenditures, and the vast majority of it is on AI infrastructure. See? Look at this. In the second quarter, we spent $18.4 billion, of which 85.8% went to AI infrastructure. The vast majority. This means that only 14.2 % of the money they burn actually goes to missiles.
And I want them to burn money on rockets. Instead, they throw them into the AI infrastructure. So I'm a little disappointed by that. Because I think they're spending a lot of money on what is potentially a deceptive asset. Why do I call this a deceptive asset? Well , and this is not to criticize Nebius or Coreweave, but let's look at the real numbers of these companies.
We see that one of them is doing a little better than the other, but both have some problems. So, let's take a look at this. First, Coreweave. So Coreweave has a capital expenditure of about $7.7 billion in just 3 months. Free cash flow is minus $4.7 billion. We can see it right here. Truth? So our CapEx is 7.7.
Subtract that, and we get negative free cash flow of $4.7 billion . The net loss here is 740 million. dollars. So, as you can see, the results aren't great. They have significantly increased their accounts receivable, although current estimates indicate that most of it is due to just three customers.
Perhaps up to 89% of CoreWeave's receivables are from Microsoft, OpenAI, and Google. And it's not uncommon for these companies to pay 15-25% upfront and then receive a smaller portion of the receivable revenue over a period of time. The problem with CoreWeave , however, is not so much the cash flow statement.
We know they spend money on reinvestment. In fact, it's about their balance. Look at the balance sheet. Uh, I have about $4.3 billion in capital here as of March 31st. And I need to pay the bills. I have long-term debt of 17.3 billion. Okay, that doesn't look like much anymore, because, you know, I have, again, $4.3 billion and $17.3 of long-term debt.
Uh, I do n't like that ratio at all, but let's imagine for a moment that they have no long-term debt at all. How much short-term debt do they have, i.e. accrued accounts payable , accrued liabilities, current debt, and other current liabilities? If I just add these categories, we get $15.1 billion in accounts payable.
That's two and a half times the cash they actually have. To be fair, I'm pretty sure they raised funds after that last report. But the only way they can survive is through the issuance of securities. They must issue debt and shares . This is the only way they can raise funds or, more likely, survive.
It's like ...it's not...in my opinion, it's not like they're saying, "Hey, we're going to raise money and buy houses at 20% discount." They say, "We need to raise money to survive." This is a little different. On top of that, they now have to raise funds at over 10% interest rates on debt. The Wall Street Journal reported that they expected to raise funds at an interest rate 1.25% lower.
They actually had to raise it by 1.25% just to raise enough capital and get what they needed. Uh, okay. Not very well. Then we can move on to Nevious metrics. So, CoreWeave's numbers, yes, they leave a lot to be desired . Things aren't so bad in Nevious. They are in a much, I would say, much better position in terms of their balance sheet.
They have about $9.3 billion in cash and current liabilities of about $1.3 billion , including $700 million in deferred payments. They have 8.4 billion in long-term debt. So, you know, there are bills to pay, but they have enough money. However, they also spend a lot. They spend almost all of their cash flow.
They spend 2.5 billion dollars. And, of course, they also issue bonds , warrants, do everything to continue to raise funds or accept investments from Nvidia, which they did too. This is a very capital-intensive business. The problem is , why are companies like Coreweave and Nebius forced to practically give away their computing power for free compared to Google and Anthropic's contracts with SpaceX?
And all of this always points directly to Google's ownership stake in Anthropic and SpaceX. It's all a vicious circle. Evil. Not good at all. I don't know, something is very strange here. Now, if we move on to other parts of the reporting conference to see if there is anything else of interest. Yes, I think it's exciting to hear about a possible rocket launch rate that will reach at least one flight per day.
It is worth noting that in 2025 their pace was approximately one flight every two days. But, you know, I usually add a multiplier to Elon Musk's time frame. So if he says we're going to launch one a day, it could be years away. To be fair , by the 2030s, it is predicted that there will be several launches per day, resulting in thousands of launches per year.
The problem here is that you have two types of missiles. There's Starship, which will have its 14th test flight next after they, you know, crane Starship 13 out of the ocean. This, you know, is a test device. Elon says they are optimistic about thermal protection technology, and that's great. But Falcon is no longer taking orders.
The Falcon rocket will fly until 2029, and they are starting to turn down orders because they only want to work with Starship, which takes larger payloads. And it is estimated that by 2028, the Falcon will have made only 70 flights. Well, I don't think Starship will be ready by then, and certainly not in mass production flight .
That's why they're already building...you know, they need four launch pads for this. Um, and you know, these plans are already in the works, but usually these things take a lot longer, and I think , unfortunately, are not going towards Starship or launching the business. As we have seen, this is going to AI.
And I'm skeptical that it's a great investment. I think it's obvious, and I think there are reasons for that. By the way, no. It's being towed now, so let's remove that. This is, uh, Starship 13. I think it's cool that they're going to launch V3 satellites for Starlink, which should make the service 10 times faster, and it's already very fast.
So that's pretty impressive. I'm not sure exactly what metrics they're using to calculate this, but this is basically my review of what I think about SpaceX. When SpaceX first went public and even before that, we predicted a 30% growth, followed by a gradual decline. And so far, that's exactly what we're seeing.
This AI calculation bet makes the stock even riskier for me, as I have no clear idea of the actual contract rates, and I sincerely believe we are being misled. Thank you very much for watching, friends. If you like my information, join us at meetkevin.com. Pay once for lifetime access to trade alerts, buy, sell, fundamental analysis—you name it . This is a course about wealth creation. Some say it's tax deductible, .
Comments 0