저자: 캐시우드 · 외국인
요약 보기
ARK는 스페이스X의 AI 컴퓨팅 사업을 극도로 수익성 높은 사업으로 평가한다. 기가와트당 건설비 250~300억 달러 대비 임대 단가 340억~500억 달러로, 투자금을 1년 내 회수한다는 주장이다. 앤트로픽(연환산 매출 약 90억→650억 달러)과 구글이 데이터센터 용량을 임차했고, AI 컴퓨팅 수요가 공급을 크게 웃돈다고 본다. 인터넷 버블 비유는 부정하며, AI 활용은 아직 초기(day 0.000)라고 강조한다. 영상 중간에 유료 후원·건강식품 광고가 포함되어 있다.
다듬은 글
Here's Kathy Wood from Ark Invest, who recently spoke about one of SpaceX's many giant multi-tiered billion-dollar money machines. So, data centers. One of the epiphanies for us was what Elon and SpaceX did with their ground-based data centers. Elon realized that he was losing in the race of advanced models and decided to reorganize as best he could, and I don't know how many people he fired or how many he brought back.
This is also in his style. Uh, but what did he do in the end? What kind of U-turn was that? He took and leased his data centers, his Colossus data center in Memphis, to Anthropic and Google. We analyzed how much it cost him to build this data center, and you correct me if I'm wrong, Brett, but it's somewhere in the range of $25 billion to $30 billion per gigawatt.
This is a very important starting point. Of course, these are just estimates. We don't know for sure, but if we're somewhere in the range of, say, $25-30 billion per gigawatt, we can start to understand how obscenely profitable this SpaceX AI business is. The current spot price for AI computing power per gigawatt appears to be around $50 billion per gigawatt.
It is important to note that this is happening now. The demand for computing for AI is simply skyrocketing. No company can get enough of them. This means that the price is likely to increase rather than remain stable or decrease. But even if that doesn't happen, imagine that you as a company could spend, say, $25 billion.
You invest $25 billion in physical infrastructure, say, computing for AI, and 12 months after launch it brings you $50 billion. Not only did she recoup her costs, she repaid them double in the first 12 months. This is complete madness. These are pyramid scheme-level profits from a real, real business with high demand.
Even if we take the upper bound of the Ark estimate, say $30 billion to build, and take the lowest possible estimate of the spot price for AI computing, which is conveniently around $30 per watt, that is $30 billion per gigawatt. Instead of getting double the value in one year, it simply makes back exactly its initial outlay.
Spending $30 billion gets you $30 billion in a year. But the second year brings in, say, another $30 billion . Again, this is unprecedented in terms of return on capital investment . I have never seen or heard of anything like it , anywhere, ever. Most companies would be incredibly happy if they could get a 100% return on investment in five, maybe 10 years.
And sometimes they wait 20, 30, 40 years. The fact that the payback period here is six to 12 months is just crazy. It is important that this is not a one-time asset that generates income and just “poof”—disappears at the end of the first year. In the second year, it will probably bring you a similar amount of income, but you will have already fully recouped the costs of creating it.
There are small ongoing operating costs, but they are negligible. So you spend $30 billion in the first year. You get $30 billion back in the first year. You have essentially covered all your expenses. In the second year, you just get another $30 billion at almost no cost. The third year—another 30.
Of course, these are very basic and simple numbers. The spot price for AI computing power may fluctuate, but where else in business history has it been possible for a company to invest capital, recoup all of its initial costs in the first year , and then generate the same revenue year after year for the foreseeable future?
This has never happened, at least I haven't heard of it . And then Anthropic comes along and says, "I'll lease this for 34 billion." So he spent a total of that much, and Anthropic comes in and says, " I need this data center capacity." "I'll pay you $34 billion per gigawatt." And then later Google comes along and says, "I'll pay you 50." Yes.
So, this is... Okay, perfect. So, these are just estimates. I mean, we have some publicly available information, but that's the point. We just heard the spot price. Anthropic will pay $34 billion per gigawatt. Google comes in and pays 50. So, according to the latest information, SpaceX AI is monetizing each gigawatt of computing power for 50 billion dollars.
So the estimates I gave earlier are a lower bound of $30 billion per gigawatt. You would pay for this thing in the first year. And at 50, you pay it off almost twice in the first year. The last point is very important. There is no magic button to immediately launch enough AI power to meet the extremely huge current demand from companies seeking access to AI computing.
At the same time, we are no longer able to meet the existing demand from companies for AI computing. So, spot prices have increased from a nominal 34 billion to 50 billion per gigawatt. The use of AI products, which is still in a very early stage. Most ordinary people, not those who watch such content, but just most ordinary people, have not even used AI yet.
Maybe, you know, something built into their Microsoft Office or Google Docs or something. But real agents, agents for programming and LLMs— man, it's so early. The use of AI is now at the lowest level it will ever be. This is a tiny fraction of what it will eventually grow to be. It will become better, faster, cheaper, more popular, and more user-friendly.
If we connect these dots, it means that as the demand for AI continues to grow at an incredible rate , the need for computing power for AI will grow proportionally. And we know that SpaceX has plans to launch many gigawatts of computing power over the next 12 months. And don't get me started on the orbital computing power of Starlink AI.
For the zillionth time , please don't ignore what's happening with SpaceX's AI computing business, or, if you prefer, the superintelligence computing business. You haven't seen anything yet . This just gives you an idea of the shortage in the market. And it wasn't just Google that then leased it to Anthropic on the backend.
So in a way he says But no, another thing: okay, how much do they make from this? Which is about $75–80 billion per gigawatt. So this is also a pretty important point. Many people don't seem to understand, but the reason companies can pay SpaceX AI $50 billion per gigawatt is because once they do, they get access to enough computing power to allow their users to scale up their models .
And in exchange for paying $50 per watt to SpaceX AI, they can earn $60, $70, or more in real income. Everyone wins here. Everything is logical. And a lot of people, we get so many questions: Is n't this a repeat of the internet bubble? Well , is that true? This is not true. How do you know? This is How do I know?
Because I survived that bubble. And yes. And what were the big investments back then? Fiber optic. Dark fiber optic. Fiber optic. And they buried it in the ground, and the customers didn't come, right? It was still dark, I mean, I think we just went through...We just finished using that power. It took 20 years, 25 years.
So it was just a crazy period. GPs today, this is an important indicator, they are in short supply. You know, these companies, given how much revenue they generate, Anthropic, and I'm sure everyone knows this figure, but we were shocked to see that last December the annual pace was $9 billion. Oh, hello everyone.
So, this is me. You probably don't know that I'm slowing down a bit, so you'll have to be patient , but I'm just trying to calculate how many months ago December of last year was. It's October now, so I think it was 10 months ago. 10 months ago. So, 10 months ago, Anthropic had annual revenue of 9 billion.
They didn't make $9 billion for the entire year , but in that month at the end of the year, their monthly income, if you multiply it by 12 months, would be about $9 billion . I think that's what I heard. Back to Katie. In July—65 billion dollars. This is extraordinary. Sevenfold growth . Think about it.
Think about it. Sevenfold growth . It took Salesforce.com 25 years to add that much revenue. Yes, actually, if you add up the revenue of Anthropic and OpenAI, their combined revenue, at least annually, has now exceeded the combined revenue of Microsoft Windows and Office. Yes. So, this has been building for four decades, and everyone knows Windows and Office, right?
And they don't just sneak by, they fly by. Yes. Ahem. They fly by. So, but why...okay, but isn't this "blown" money? How do we know that the GPs—not this year's, but next year's—won't become something like "dark fiber"? We have a bunch of GPs lying around , and we have to use them as pizza boxes or something like that.
No, there are rumors. We believe they are justified: Anthropic's gross margin is over 80%, and the company has been profitable on an operating income basis for two consecutive quarters. Not sure what's more interesting here. The numbers themselves, which are quite extraordinary, or the fact that Katie Wood went to great lengths to emphasize that the rumors were very well founded.
Here is the main conclusion. Today is already a reminder that it's still day 0.000 when it comes to how fast, cheap, intelligent, and capable AI will be. Even today, AI is so useful, valuable, and productive that the products companies offer now will seem primitive, slow, and expensive in a few months.
Today, AI is so incredible and exciting that companies like Anthropic are making crazy money, growing revenue and profits at a rate never seen before. Their limiting factor , the thing that prevents them from scaling even faster and making even more money— drumroll, please—is access to AI computing power.
This is the beginning of the beginning of the very beginning. You owe it to yourself, if you are not already using AI tools in your personal and professional life, to start doing so immediately. Not only will you benefit immensely from how much value they add, how much time they save, and how much mental resources they free up for you, but you will also realize how powerful these tools are right now, even though it is only day 0.000.
You need to use these AI tools, by the way, the best product of all time— Clockwork, buy it, use it, thank me later. But in doing so , you will also begin to understand how powerful they are now, how powerful they will become soon, and how the demand for them will be almost infinite for decades to come , which again means that the demand for AI computing will be almost infinite for decades to come.
Thanks for watching. If you appreciate this content and want to support the channel, check out the link below where you can get early access to daily videos, unlock hundreds of exclusive content, and more. My new 20- year Tesla and SpaceX valuation model is now also available on Patreon for investor-level supporters and above.
AG1 has been part of my daily health protocol for half a decade. It is packed with vitamins and minerals and helps me fill nutritional gaps by supporting energy, digestion, and immune function. It also contains prebiotics and probiotics to improve gut health. I take my health very seriously, and you should do the same .
Try AG1 today by visiting drinkAG1.com/SMR or clicking the link in the pinned comment to receive a free welcome pack of Vitamin D3 plus K2 and AG1 travel packs. This is drinkAG1.com/SMR.
Comments 0