저자: JCristina · 외국인
요약 보기
일론 머스크가 SpaceX/Tesla 생태계의 AI 칩 수요를 자체 조달하기 위해 텔라팹(520만 평방피트, 약 250억 달러)을 건설 중이며 2030년 월 100만 웨이퍼 목표. 자체 AI5 칩은 H100 대비 가격 약 1/10, 크기 절반 주장. 칩 과잉 공급 시 엔비디아 등 기존 GPU·칩 제조사 매출 타격 예측. 중국 DeepSeek V4의 메모리 분할 구조가 GPU 메모리 요구를 낮춰 엔비디아에 부정적이라고 주장. 대만해협 등 봉쇄 리스크도 텔라팹 건설의 배경으로 언급.
다듬은 글
Why Elon Musk is building a $25 billion Nvidia killer? It's a pretty good question. Let's get into it. Everybody, welcome back to the channel. So, today we're going to be talking about this massive shift that's going on over there with SpaceX and how they're moving from just a simple rocket company to now this conglomerate, this monopoly almost when it comes to everything tech, from space to communications to phone to AI, on and on and on.
So, there has been a lot of talk about orbital data centers. And what they ended up calling them officially is an orbital data center system. And this is really important for what we're going to talk about today because it is not a center where it's just a single structure. It is a node-based, let's call it a decentralized distributed network, a mesh network of nodes that all work together for the greater good.
You follow that? So, if one of these satellites that have the AI chips on them go bad or drop out or are destroyed or even if a hundred, it doesn't make a difference. Just like your mesh network in your home. Well, Elon Musk is not a fan of making cars. Um, he said it's probably the hardest business he's ever done.
And he really doesn't want to do it very much. I think he's going to end up keeping with the idea of this RoboCab stuff and maybe one or two cars kind of in the wings to keep that going also. But, we already know that one of the plants is being completely retrofitted to make Optimus robots. That's a major change.
So, we're going to see probably about 10 million Optimus robots. That is going to happen by 2027. I think it's about 1 million in the one factory just in the first year. So, it is a lot and we're going to see a lot of these robots. So, what does Elon need for the data centers in space for his Robo cabs, for his Optimus robots, for all of this stuff, right?
What does he need? Well, he needs brains. He needs AI. He needs chips. Well, how do you get the chips? Well, right now he's getting the chips from Nvidia, Samsung. The problem is is are you sure that you're going to be able to get the chips in the future? And he predicts no. So, what is he doing? He's making what is called the Terafab.
So, what exactly is this Terafab? Well, Terafab is a chip manufacturing location. A fabrication site. It's going to be the north wing or off the Giga Texas over there and is it going to be 5 million or 5.2 million square feet? It's huge. Absolutely massive. And just think about this, guys. Elon Musk has always stated this.
He said, "We need to build the machines that build machines." This is his mentality, his methodology when it comes to fabrication. Build machines that build machines. This is how it's going to happen on Mars. This is the way it's going to happen on the moon. Because we cannot be there to do this and machines could do it more efficiently.
Now, the hardest thing on the planet to do currently is manufacture 2 nanometer silicon. Those wafers are very difficult to manufacture and that's exactly what he's going to be doing in this plant. So, this is a very big, big deal. It is a 25 billion-dollar big deal. It is a lot. It is big. And what he's saying is it's going to be targeting about 1 million wafer starts per month by 2030.
1 million per month out of this specific factory. Once again, 5.2 million square feet. So, if you take those numbers and you just look at what TSMC is doing or Samsung. If you took TSMC for example, that is 70% of their global output would happen in the US under this one roof. 70%. That is a lot of chips, guys.
That is a lot of chips. So, what is the rationale behind this? Why does he want to make his own fabrication plant? Okay? Instead of just purchasing from Samsung or purchasing from Nvidia. I mean, if you guys don't know it, as of right now, if you look at Colossus, Colossus 2, Cortex, which is the Tesla computer as well as Dojo, Tesla also.
If you look at these computers that Elon Musk currently has running, these are supercomputer AI farms, let's call it. If you look at these, they are running over 300,000 300,000 Nvidia GPUs. Now, those could be like H100s, H200s, so on and so forth. Okay? 300,000. You're talking about 10 billion dollars of Nvidia purchasing over the last few years.
You think it's a little bit? That's a lot of bit, right? That is a lot. So, what he's doing here is say, well, I'm spending all this money with Nvidia. Why don't I just spend it on our own manufacturing plant? They've spent 10 billion already just on Nvidia chips. Why don't we spend 25 billion make our own fabrication location here in the United States?
Now, that goes once again back to the why. His rationale is this. He does not feel in the next couple of years Now, that goes back to the why behind this. Elon Musk believes in the next couple of years there is a high likelihood that there is some type of South China Sea blockade. Now, exactly does that mean?
Well, the same thing that's happening over there in the Middle East where they have the ability to block a certain passage so that energy can't flow out of it. You follow what I'm saying? The same thing could hold true here in the South China Sea. If they were to place a blockade at that location, what that would mean is we would go from millions of wafers, millions of chips coming into the US to zero.
Like that. That's a problem. That is a massive problem, right? So, having the foresight, having the wisdom, not just the intelligence. People confuse intelligence with wisdom. Having the wisdom to do this in the long term will insulate him from this type of problem and allow him to facilitate all of the growth that he wants to happen for all of his companies.
Very, very important. AI is everything and chips is everything that he does. So, every product is going to have a chip in it. A perfect example is that $650 macro hard. I love that Microsoft poke in the eye. Macro hard. That a genic AI system that he's building. I did a whole video on it. Where you'll have a $650 computer.
It's actually one of his AI Well, right now it'd be an AI 4 chip, but eventually an AI 5 chip on your desk that controls your computer. Does everything fully autonomously. Anyways, watch that video. All right, good stuff. Really good stuff. So, that's just one of the things. He's going to also need these chips in every one of the SpaceX Starlink version 3 satellites, which will be nodes to a supercomputer.
A decentralized supercomputer. Very important. He's going to need these chips for his Robocabs. He's going to need millions, billions of these chips for his Optimus robots. You see? By making his own basically allows for this verticalization of the business, where he controls everything instead of being beholden to a third party.
Once again, wisdom. Right? Not just knowledge, not just brains or intellect. Wisdom. So, in my personal opinion, when I look at this and seeing what's going on with Nvidia and the H100s, I say, "Wow, you know, Nvidia really needs to be on the lookout to what's going on here." You know, it's the same thing that I said with AT&T, Verizon, T-Mobile, and all the rest of them.
I said, "They need to really start shaking in their boots as Elon Musk eventually gets his DTC off the ground, where you could communicate with your phone, an unmodified phone with a satellite in space. What's the need, right?" And I also said that when he does get this working properly, what will end up happening is is he will become a fourth, let's call it global cell carrier.
Elon Musk says he doesn't want that, but he can definitely do it if he wants to. And and the big thing here is by buying his own spectrum means that now he's not, once again, beholden to a third party, to Verizon, T-Mobile, or AT&T, or someone else to piggyback off their spectrum to be able to get his DTC to work.
Now, DTC is direct to cell, but now they have a new name for it. It's called Starlink Mobile. That's for another video. Anyways, the idea here is is that Nvidia's chip or GPU that he's currently using is like the H100s or the H200. So, I did a little bit of a deep dive into the H100 cuz I wanted to know like why is this better or how is it better?
So, what I found out was this AI 5 is just next level, okay? When you do a comparison between the AI 5 and an H100, they are pretty close to identical when it comes to computational prowess, let's call it. But, where things change or where it's different is that the H100 from Nvidia, that GPU, costs $30,000, okay?
Whereas the AI 5 will cost probably right around $3,000. So, you're looking at a 10x difference just in cost. Also, the size of the wafer is half the size. So, if you were to take an Nvidia chip, you'll be able to put two of the AI 5s within that same amount of space. Let's just look at it that way.
It didn't make it's a little more difficult than that, but let's just make it super simple. Why is that or how is that? Well, with Nvidia, it's a general purpose chip, meaning that not only is it used for AI, it's also used for its original purpose, which is graphics. You follow? So, what did Elon do?
They ripped out all of that legacy nonsense, right? They wrote it custom, very tight, and just took out all the bloat. There's no bloat. All right? So, the chip can be half the size. Once again, it is the same speed, it can function at the same power, let's say, but it is half the size. And what is really incredible here is the amount of watts that it uses.
So, instead of using 700 watts, let's say, all the time, 700 watts to 800 watts, let's say, would be the peak performance of that AI 5 chip, and it falls way back when it doesn't need to be used that hard. So, we're looking at a chip that's going to run cooler because not only is it going to run lower in wattage, but also it is half the size.
That is a That's massive. That is absolutely massive. So, once again, instead of building something that is more general purpose, what Nvidia has been doing all of this time, well, he's building a chip that is highly specific. So, in short, what Elon Musk is doing is basically taking a $10 billion baseball bat to Nvidia.
You know, he was quoted in saying this. He said, "Our explicit goal is to have an oversupply of AI 5 chips." Now, if you read into that, what this means is that the amount of fabrication that will be able to be made with this plant will exceed the amount that they need for the data center, for the SpaceX Starlink version 3 satellites, for the robocabs, for the macro hard, for the Optimus ro- so on and so forth, okay?
It will exceed that. What does that mean? Well, number one, what that means is that Elon Musk and his companies, right? For his three, four, five different farms, AI farms, will no longer be buying $10 billion worth of Nvidia GPUs. So, those 300,000 GPUs that were bought, that contract is gone. We're not buying anymore, okay?
He makes his own. But, the thing that even is worse for Nvidia is, like he said, the oversupply. I don't know if people understand this. What that means is he's going to have more of these AI 5 chips than he knows what to do with. So, what do you do with them? You sell them on the open market, right?
So, now all of a sudden, you have a another means of making money because now you're facilitating the silicon for all of your ventures, but then you have extra to sell on the open market. You also deplete the amount of revenue that Nvidia could possibly make. And all of the chip manufacturers, like the TSMCs and the Samsungs and all the rest of them.
So, that's why I said, these companies are they're going to start hurting. Just like I predicted that's going to happen with AT&T and Verizon and T-Mobile, the same thing that I predicted with Dish Network and DirecTV, these companies need to be worried because once this thing happens and this terrafab is online, and they're making those those millions and millions and millions of silicon 2 nanometer silicon wafers.
I mean, this is going to be This is going to be big for those companies. So, while Nvidia will continue to do well as of right now and the rest of these companies, the rest of these fabs, within that point of time of 2030, when everything is at full ramp, full production, I think within 4 years, we're going to see a toppling of Nvidia and some of these other companies, unless they somehow can figure it out, which T-Mobile, AT&T, Verizon, they're trying to figure out right now.
Dish Network and DirecTV, they never figured it out. And then of course you have HughesNet and Viasat that definitely didn't figure it out, that they just went out of business basically. So, this is the type of thing that when I make these predictions, I just I just want you guys to just watch it. Sometimes I'm right, sometimes I'm wrong, okay?
But in my personal opinion, I'm going to be right about this one. And there's another piece of the puzzle here that is really, really going to be, once again, that thumb in the eye of Nvidia and the rest of these GPU manufacturers. And what that is is the Chinese DeepSeek version 4. Now, DeepSeek is basically an AI, okay?
It is a model. And what they've done with DeepSeek version 4, and I'm going to do a whole video covering this within the next week, once it's released, maybe 2 weeks. What they've done is they've taken the model and instead of taking the entire LLM or large language model and forcing it or injecting it into the VRAM, the high-speed VRAM of a GPU, making so that the GPU needs to have an excessive amount of that VRAM.
For example, like an H200, H100, I think they're like 181 gigabytes, I don't know, don't quote me, or 141. It's a lot, okay? So, instead of doing that or having something that's that expensive, what they're doing with this new model is they're dividing it. So, you have your brain here and you have your encyclopedia over there.
It's called an Ingram architecture. You're dividing the processes. Meaning that anything that it needs to figure out, like, I don't know, what is the capital of Florida, whatever. That's in that encyclopedia. That's in your slower, regular memory that you buy cheaper, okay, in your computer. That regular memory, that's where the encyclopedia is housed.
Whereas the intellect or the brain, the cognition that happens, all of the thought, all of the thinking happens in the high-speed memory in your GPU. What that means is is the GPU doesn't need to have a massive amount of memory anymore. So, you can get a GPU that has 16 gigs, 24 gigs of RAM, let's say, and be able to run the same model that you would on an H100, for example.
Why? Because that Ingram architecture divided it up. So, you can have, let's say, 24 gigabytes of RAM here in your fast-speed GPU, and then you can have 128 gigabytes sitting on your computer, where your encyclopedia could be housed. So, you could have this massive LLM on a budget. So, this is also going to affect Nvidia in a negative way.
So, not only Elon Musk is putting pressure on them, but also in their own country, their truly open-source large language model their AI. If you want to download it, you can. I will and I will show you how it works in the next couple of weeks. Is once again just making things really rough on Nvidia and all of the rest of the companies that are fabricating chips because you don't need as many.
Now, if you're fabricating regular VRAM, that's fine, but you're not going to get as much money for them even though they have doubled price lately and you know why because I just told you why. Regular memory has doubled in price. Go to Amazon and go see how much a gig of RAM is, 2 gigs of RAM. You'll see.
It's doubled and the reason it's doubled is because the writing is on the wall. People are going to buy more of this regular RAM for their computer in comparison to buying a very expensive GPU if they want to run AI. And that kind of is the point here. What is going to happen with this AI 5 chip where it does become something that you can just purchase on the open market when he has so many of them to be able to just sell it.
What will end up happening, the same thing that will happen with this macro hard, is we will now have a powerful enough chip to be able to run an AI truly locally. So, instead of paying to chat GPT who is not open source, who is now closed source, chat GPT, Gemini, Grok, Perplexity, Claude, whoever you use, right?
These are all in the cloud. What that means is when you ask questions, those questions can be monitored. They can use your information to train on. They could use your information possibly against you, who knows. But, if you have an AI internal to your network, well, it's yours. You could do with it as you wish.
And if you need it to go out on the web to get current information, you can, but all of those questions, all of that data stays, once again, on your localized system. It's not in the cloud. You ask for maybe some data, it sucks it back in, and then it churns and processes that data on your localized AI.
Hence, that's why I'm going to be doing this video on deep seek. So, I think that there's two real important things to monitor going forward. Number one is the idea of a transition from extremely fragile, cloud-dependent type of computing that we've seen in the past to what Elon Musk is doing on orbit with a fully decentralized computer system, node-based, mesh AI network, AGI, basically, Skynet.
We need to watch that, all right? And what we also need to watch is the ability for us to be able to have our own AIs in our homes, so that we don't have to use and give money every month to these corporations as they use our information continuously against us. As we can see, currently, ChatGPT, once you get to a certain number of questions, all of a sudden, at the bottom, if you're not paying them, you'll start getting advertising.
And the advertising is specific to the questions that you're asking. That's problematic. Now, they say that the information that you are requesting is not being shared with these advertisers, but they sure as hell are generating advertising specific to the questions that you're asking. And I think that's problematic.
I can tell you this, if these legacy giants fail to see this coming, like the AT&T, T-Mobile, Verizon, your Dish Network, and your DirecTV, your HughesNet and Viasat. If they don't see Elon Musk coming in the rear looking in that rear mirror, they are going to be left behind in the digital wasteland.
That's my personal opinion. All right, what say you? What do you think about all of this stuff? What do you think about the AI? What do you think about this Terafab that he's putting together? This massive 5.2 million square foot fabrication plant. What do you think about all of this stuff? What do you think about this AGI or this decentralized computer in space that I've been talking to you guys about for a long time that is coming to fruition for sure.
What do you think about all this? Anyways, guys, I hope you enjoyed this. If you have, throw the video a thumbs up. Don't forget to share it with your community, with your friends, family, colleagues, Reddit, Facebook, X, wherever you traverse the web. Please share the video, share the channel. That will help out a lot.
If you want to give back to the channel, there's a thanks button. Click on that. Give a dollar two if you like. If not, it's perfectly fine. The video's still free. Consider becoming a member of the channel. That would be even better. And if you want to check out some of my merch, I put together a website just for you.
Go to jcristina.com/shop. jcristina.com/shop. If there's some merch there that you like, pick it up. Helps support me and my family. And since you got to the end of this video, well, I have 650 other videos on SpaceX, Starlink, and Elon Musk, and helpful how-tos and tips and tricks and what to do and what not to do and what to buy and what not to buy and of course the why behind everything.
This is channel's always been about the why, not just the how. Many blessings to you and your family. Stay safe. Stay healthy. Stay connected, hopefully through SpaceX Starlink. And we'll see you in the next one. Love you guys.
Comments 0