Tesla and SpaceX Just Became Much More Intertwined

Kind: captions Language: en Tesla and SpaceX just announced Terra Fab. It's a chip factory the two companies are building together in Grimes County. And Elon Musk went on a posing spree about it like it's the biggest thing he's ever built. And honestly, it might be. The bigger story is that two names on the door, Tesla and SpaceX. And what that tells you about the merger question. [music] Elon says it will be the largest and most valuable building on Earth by far. Over 100 million square feet. That [music] is 10 times the size of Giga Texas. 15 times the Pentagon. two and a half central parks. The goal is over one terowatt of compute per year, full stack chips, memory, and packaging. His rough split 25% of the output for the Optimus, 75% for AI [music] spacecraft. Today we have Phil Bell joining us, and he's put some research into place asking the question, I can't shake. Is that even enough? [music] Elon said on the earnings call that memory is the big bottleneck. Supply growing 20% a year while demand grows 200%. Phil's curves say demand outruns one teraf. There's no limit to intelligence. And if you're wondering why SpaceX would want to merge with Tesla, this is the clearest answer yet. Their futures now run through the same building. Optimus and AI spacecraft made from the same chips under one roof. Phil Belle was a founding member of the technical team at electric vehicle maker Rivian and has also held various engineering management roles at Apple. Welcome, Phil. >> Good morning. Thank you, Phil. Appreciate the work you put into here. You put together some uh graphs, research. We'll go through that. Before we do that, though, Elon has been and SpaceX and Tesla surprisingly have all been posting about Terra Fab. And um this is SpaceX. The foundations for for an exciting future are being built in Texas. Next up, it's Terraab. And this is one of the stills. I'll play the video shortly. This is one of the stills of that video. It's very futuristic. You can see uh Tesla semi, cyber cabs, robo vans, futuristic city, and so forth. Um Tesla themselves posted this. Okay. Terrafab will be built in Grimes County, Texas. In April, we broke ground on our research fab on the north campus of Giga, Texas. The precursor to Terra Fab is a joint project between the two. Both Tesla and SpaceX will need far more chips than current and future global production can supply. This is why we're building the largest chip manufacturing facility ever with a goal of producing over 1 terowatt of compute per year. U this is the photo of a render of what it could look like. It's gorgeous. Elon uh has been posting. He said sci-fi city is what we're aiming for. And uh Doge designer said Terrafab looks more like a sci-fi city than a chip factory. That's what he replied that this is Terraab and uh beautiful render by the way just gorgeous. I I wonder why they decided to make it this way. He said terrafab Texas will be the largest and most valuable building on earth by far and it will be stunningly beautiful. Phil, I noticed that he said Terraab Texas just terap terap Texas and I know you have uh thoughts about that. Tell us what you're thinking. [laughter] >> Yeah. Oh, it's just I mean that is a stunning render of that building. I mean it's going to change I think I think it'll change Texas if you know what I mean. like it'll it'll move the epicenter of of tech commerce uh a little bit further east in Texas. So, it's incredible. It's incredible. Look at that building. I mean, you know, [laughter] do you realize 100 million square? By the way, we don't know exactly. I mean, I remember a post I did some time ago when I sized up uh Terrafab and I said ballpark 100 million and Elon agreed with that and he said um >> in that in that in that order of magnitude you know like yeah not not 10 not beyond 100 not like you know a thousand but 100 and uh 100 million square feet that that is eyepopping. Yeah, 10 size the size of Giga Texas. And Giga Texas was supposed to be pretty well the largest factory on the US except for Boeing. They might be like part of part. >> I think Boeing had more volume because of height because of planes. But I think on a square footage floor space basis, I think Gigafactory might be the biggest >> and this is 10 times the size. So this is crazy. 15 times the Pentagon. Uh Elon also posted my very approximate guess is that Terrafab AI compute output would be 25% for Tesla Optimus. Yeah. >> 75% for AI spacecraft. And we already know >> that Optimus is supposed to be this billion number Optimus. So that's crazy that this is how what he's thinking in terms of size. >> Yeah, that's the split, you know, that's where he's gone with the Tesla chip team. uh you know that the the Optimus side is the AI5 AI 6 generation chips plus+ and then the rest is for Star Mind in >> what's that called B something or B >> uh no uh D3 >> D3 >> it's a bit of a homage to the uh Dojo project back when >> which interestingly in some ways is full circle and you know I think it's a it's extraordinary what's going on here because um essentially he announced just before the earnings call, which is now two days ago for SpaceX. I feel like things go like, "Oh my god, that was two days ago." Um >> I know it's crazy. >> It was like weeks ago. >> You know, they he talked about how he was partnering with Nvidia to build the first uh let's just call it a potted version of a compute, you know, CL uh compute center for the uh AI satellites. And what's interesting about that is also made reference to the fact that that same effective architecture with a different cooling structure would be the same one that gets planted on terrairma which not only means that it not only means that he's he's talking about that in terms of of what he intends to do with Colossus but he also means like as he goes out into that potting future of uh putting uh inference compute at say supercharger sites as you know it I what I think is really interesting about that is he's basically saying I'm going to use the same basic architecture across these these two two very different platforms um physical manifestations and uh and then what we know is then he'll swap that platform over time into his own chips that are being produced at >> terafab >> I [clears throat] mean it's it's a very nice clean story and it makes too much sense yeah >> and it makes that it's going to happen sooner than later cuz we all thought that we won't have we will be talking about this for years on end. All right, let's let's go through your deck uh that you put together here. >> Well, these were just two posts I made this morning. I mean, I saw when I saw that building, I was like, "Oh my god." Um I mean, we all expect it, but what a beautiful design, by the way. uh and and uh I make a point there about how I think it's it's uh it's massive, it's big, it's necessary, and then I kind of lead in on saying it ain't big enough uh or at least in terms of total compute required. And then the other side of the post was I just sized it up relative to what you started with the opening is that we >> now would have a I mean look at this this a highway goes through it pretty much, right? I mean, it's that it's literally that big. Um, if you know, you think about it, a building that's 10 times the size of uh Gigafactory, you know, it's uh 15 times the size of the Pentagon. Have you ever been in Central Park? Could you imagine having two Central Parks worth of building space? I mean, >> uh, you know, >> I've never walked the entire path because it's too big. Yeah. >> Yeah. This is this is this is quite quite amazing. and and you know, as we'll talk about in this uh in this pod today, uh it it isn't going to be big enough. >> So, I think that uh when he says Texas terrafab, you know, I think he he is saying that that he's going to build another and another and another. And I know it sounds like that's completely nuts. It's not completely nuts. That is where we're headed. That is where the scaling laws of compute are dragging us. And um so >> who wouldn't want to work here? This is going to be amazing. All right, let's go through your deck. >> Yeah, this kind of u this is what you know Elon's talking about the same thing that uh I'm about to talk about which is basically I mean he's he's the one that brought me here. I mean what I mean is you know if you go back in I think it was March uh mid-Marchch when Terrafab was announced and um again seems like a lifetime ago but it wasn't that a couple months ago really and uh you remember that wonderful graph he put up where he said hey look the world over here is running around 30 gawatts of AI compute all in and I'm going to go build a terowatt you know 30 gawatts to one terowatt and you see the sizing difference on that and you think to yourself, okay, he's absolutely crazy. But then look at this uh here alone what he's talking about. This is this is him talking in the earnings call two days ago. Um and he said, you know, the biggest bottleneck for AI expansion right now remains memory. And while the industry is growing at at a very healthy 20% a year, he's saying the demand side is pushing it at 200%. I mean, that's why Apple had that spilloff a couple weeks ago when it admitted that with all their um with all the accolades that come their way on their ability to manage the supply chain, somehow they didn't quite get it right on the memory side. And and frankly, [clears throat] the reason that that's true is because they might have seen this coming uh like, you know, a freight train, but the reality is they're not vertically integrated, so they can't do anything about it. Um, so what Musk is trying to do here is he's trying to tackle this problem. uh Terrafab you know uniquely is uh not only produces the logic side of this equation but produces the memory and does packaging. So that's you know it's not like think of it it is like dub 2x vertical integration. First you have vertical integration in the sense that that the companies Tesla and SpaceX will be consuming their own chips. There you have vertical integration but within the factory itself within the fab itself you have it fully vertically integrated in the fact that it would produce the memory. It would produce the logic and it would do the packaging. You don't see that. You don't see that memory is produced elsewhere typically shipped over to the packaging house. Um whereas logic is pretty you know Apple's chips are done that way right you've got t TSMC um um doing the logic side uh some of the packaging memories purchased elsewhere ship shipped back to the fab for packaging um and Elon's like no no no we need to do it all ourselves and what's interesting about this too is that you know people think of memory as just sort of like a a true commodity thing wrong memory is extremely hard to build at the densities that that these types of chips demanded at. It's a very very uh high tech piece of of um you know silicon. It's not you know everyone thinks of oh the logic you know let's focus on the the GPU side. Uh but memory I think is equally hard to to build in its stacked form and dense form that it is today. So, you know, Elon's telling us that right here there's a 10x gap and and uh I have thoughts. [laughter] >> Phil, would it be fair for me as a kind of an amateur at this think of it think of this? Okay, I hear I hear Terra app. I hear how big it is, but then I kind of gloss very quickly that they'll make their own chips, that they'll make their own memory, that they'll make, you know, as you just described everything. Is it fair for me to say it's actually going to be an Nvidia and a Micron that size companies into one thing that Terrafab could be thought of as a Nvidia and Micron put together? >> Yeah. I mean, essentially, you know, if if you think about Nvidia, for all intents and purposes, they're a design house. You know, they they design the chips. They don't make their own chips. Taiwan Semi makes their chips. Um maybe >> so so it's a Taiwan TSMC and Nvidia and a Micron all into one thing. >> Yeah, it's it exact. [laughter] Exactly. I mean the only thing it's not is producing today the lithography machines you know the actual machines that um ASML produces that are will be in extreme high demand uh already are you know those are the massive uh you know I don't $400 million machines that basically do the uh ultra ultra violet lithography and image the chips you know I mean yeah I mean it's it's true it is fair to say what you're saying Because if you just look at the look at the sizing of the building and try to figure out what it's what it's doing and why it needs that sizing and and what what Elon is effectively saying is I I want to I want a facility to produce an annual on an annualized basis one teratt of u of compute power that is massively larger. I remember what that graph said. It said that the world's supply of effectively GPUs like AI compute uh is running around 30 gawatt, right? Um now Elon's saying that ain't good enough. I need to move that up to you know one terowatt. So yeah, it's totally fair to say that that's what he's doing. That's the sizing here. We are talking about combining the fabrication side with the I mean the design house. I mean Tesla is doing te Tesla the company they're doing the effective chip design here right they have the chip design team in house um and and and for the longest time they've been sending their chips out to the leading fabrication houses to to build the you know the the AI4s that are in our cars today um it's either Samsung or Taiwan semi it doesn't really matter but um you know Elon saying you know I'm going to cut them off and it's not that so much that he's trying to really go around those guys it's just that what's happened is he's gone to those guys and said, "Hey, you realize how much I'm I want." And they and and by the way, when he sized, you know, when he was doing that sizing, back of the napkin sizing, he was really doing it for the Optimus product more than anything. Uh it wasn't I mean, back in September when at the shareholders meeting, which you you were at, um when he first talked about this whole idea of doing Terra Fab, I think that was the first time it was publicly talked about. Um, you know, you remember he came off the heels of I think in August and that's uh a year ago talking about kind of owning the full chip production of the Samsung Spec Factory in Texas. >> Y >> and everybody thought, you know, that was massive and it's not. It's just simply not. Yeah. I don't know what its output is, but it's it's it's it's uh one or two more. It's at least >> it's a whole facto's worth. Yeah. It's a whole factory >> said I I get everything you guys build and then you can't build what I need more. So I need to build my own because >> Yeah. >> I mean in in a weird way um uh terafab is so incredibly necessary for Tesla and SpaceX's future that without it the companies will fail. So I mean it is an absolute necessity. It's almost like he need to get he needed he needed to have gotten on it yesterday. >> Uh I know it feels like oh wow what is he thinking? What's he why is he doing this? And I'm well I'm about to say why. Um but in a way I felt like after looking at the problem a little bit I'm like wow he's he's already two years behind where he needs to be. >> It's interesting because obviously he's now people are now not only him but throwing out the number that Tesla or SpaceX or the combined company could one day be a hundred trillion dollar company. Okay the the most the the biggest company to get today in market cap between Nvidia, Apple and others is four trillion, right? Maybe five trillion. Everybody can't imagine a hundred trillion company. Well, here you go. This is why Terra is the reason. >> Look, I think if you think if you if you if you check a box that says Terafab will produce its compute, right? >> Uh we'll talk about this in a second. Uh the the reality is that Star Mind is really where it goes, right? I mean at at the end of the day um >> everybody that's out there on in the frontier space that's on the frontier of building AI today whether it's Anthropic whether it's Google whether it's open AI and I mean even open AI will have to go through SpaceX we'll have to go through Star Mine to satisfy the customer's demand I mean that sounds stunning but it's absolutely true there's no way around this pickle unless uh someone else gets to space at the scale and I don't know who that would be. So, let's take a look at this craziness that I put together here. And I think >> curve. >> Yeah. Uh, by the way, our our our dear friend CERN Passure named it um >> the bice curve. >> Yeah. Um, >> okay. >> And and and you know, I don't want it to be hung around my neck as is u the silliest thing I ever did, but I look at this sometimes and I say this is completely crazy. Like this can't possibly be true. And um I've asked a bunch of AIS and they they've argued with me in in in two directions to be fair. But let me explain what's going on here. What I'm really doing is I'm trying to uh build some kind of picture that talks about the gap between the demand and the supply of chips >> for compute. And the way I did this is I said, you know, if you looked at the the blue line here, uh the blue line over time, by the way, this is a logarithmic scale on the on the y-axis. So realize, and we'll show what the linear ones look like, which are kind of crazy, but what the blue line is represents is just sort of the unfettered demand at three I think it's 3.4. I can't read my own thing there. 3.4x um Yeah. per year. >> That's a huge number. That's a huge huge number. And we'll compare it at the end why that's a huge number. And that is where we've been historically and I say historically with kind of a funny like for the last four years that's what AI has been put those are the numbers that AI has been putting up. Okay. So as we increase the demand and we're increasing the demand in two ways. We'll talk about that in a minute. um it's dragging us you know it it in a sense at least it's continuing at that growth rate at least it's continuing at that growth rate which is eyepopping growth rate okay now I built another line here and I dragged it down I dragged that demand down um to the to the orange line and you know I don't know you could push it further down if you want there's uh this is mostly looking at trying to build efficiency through hardware, you know, and and the reality is that unless we are talking about the quantum space or something like that, which is not is more of a a non-reality today, we really have hit the wall in Moors law in terms of compute. We we did that with the CPU. And what we do with the GPU, ironically, is we think, oh, the GPU somehow broke the curve. It no, it didn't. It just put parallelism in place. That's what it really did. It just basically popped in more parallelism. And I mean, if you really zoom out, that's what's happened here. >> And what that means is that while the chips themselves may look physically somewhat bigger, there's a lot more parallel core in that chip to do more compute concurrently, which and the good news about AI is it's a highly parallel bio parallel wait screwed that up. It is a parallel problem. it is a problem that can be realized through parallel compute. So you know that's why it is so uh so well served by the GPU if you will or TPU same kind of um thing. So what I did here is I said okay look there's that blue line can't be true over time it's just insane you've got to build some efficiency into this now maybe that efficiency and anybody can comment and say that efficiency is not enough. Maybe it would drag it much further down. Okay. So now let's talk about the supply side. This if you look at the the the the green line that goes straight across, not not the curving one. Yeah. Think of it as the dotted green line. That's essentially um what supply looks like without tariff. So that's what you know all in Taiwan semi Samsung and anybody else that's building the logic side of this equation. And by the way, the memory is an analog of this. It's like basically the logic drags in memory. So it's not really much of a different story. Um that's essentially what the supply is. So there's this huge gap in that and that's quantified by that that red area uh between supply and demand. Now you're welcome to push the efficiency uh you will you're welcome to push that efficiency line further down uh to to kind of come closer into supply, but I but I argue that it won't hit it won't hit it. And so if you look at it at least on this basis, look at how terra fab makes a dent. It's the green area. Um so what that means is that when terra fab comes online, you produce a terowatt of of compute annualized that you will start to come up to catch that uh uh demand side. But look how it flattens off. And that's an argument that says one terapab is not enough. you know over time >> you need to add >> more and more terapabs. Now the good news is if you build one terra fab not that it's not that it's easy but the research part of it you know the the heavy lifting isn't the physical construction of the building mind you that's difficult for many people that seems to be you know where Elon is quite magical but the heavy lifting is coming now and that heavy lifting is how is he going to build his version of this through the supply chain of what is and is he going to take a traditional path lethog graphy, traditional lithography path, or is he going to do some, you know, changes to that? I think he's going to do a bit of both. But, um, you know, he's pinched. He's got to kind of do do it the way it traditionally has been done, which is already high art, and then he's got to figure out new ways of doing it to build more efficiency himself. So, you know, you're welcome to argue with the the realities of this, but what I am saying unequivocally is that uh demand will in this AI universe we're in will outstrip supply by appreciably over time. >> And and one terapab is enough. And that's why he's hinting terafab Texas. >> I think so. >> And that's why he's always said factory is the product. One factory, it's not a product, but multiple factories, it's the product. Who can build a factory? >> That's right. That's right. Um, okay, you can flip to the the next lovely slide. So, in my uh my need to argue with my own slide, I threw this through Grock and said, Grock, argue this for and against, you know, and I said, don't you worry, don't you spend your time worrying about supply constraint because it got hung up many times saying, well, you can't do this because of supply. Well, duh. That's the whole point. But what what this is is this is this is with not thinking about anything related to supply. This is just pure purely the demand equation and you have three curves here. Um purple is the multiplied curve between basically the two others. And the the first curve uh that you see the blue curve is is simply um represents growth in AI. And that simply means that as we use more AI, we use more AI. You know, it's sort of like saying um personal computers were sold to a very few number of people in the early days and then all of a sudden it exploded into a mass market product. And the same thing with you know phones. I mean now everybody has a phone uh whether they have a Samsung phone or an iPhone. So so basically this is just talking about use uh demand. The the orange line represents intensity and it what that is saying is it's saying hey guess what the more we use the deeper we go in our use our queries that used to be you know couple thousand tokens on output are now generating millions of tokens in output. Our our questions are deeper. our our use cases are expanding and and by the way those use cases are uh today with textbased types of AI are minuscule compared to say generative AI in say video video games or um movies or commercials or however AI might be used in the in the image generation space. So what what the AI is doing here for me by the way the irony should not be lost on the audience that the AI is doing the thing that is talking about AI demand here [laughter] um you basically >> and what I understand is that uh I mean a lot of people are using chat GPT now but less than 1% of people are using agentic AI we think like I'm using it but that like when you look at this you know globally it's very little at some point everyone will be using aentic AI when Next year when iPhone includes it automatically in your iPhone and an Android has it in their Androids and it's in every device every it will be so easy for people to just spin off an AI agent to do something for them and that's going to skyrocket and then once you use it more you'll you'll want it to do more and I think it's called the Jevans paradox right which is when a technology becomes easier to use or more efficient >> more demand >> more people actually end up using it uh than they have been in the past. You can't be looking at the past of usage and use that as your line. It's going to actually be exponentially more, >> right? And that's >> is that what this is? >> That's exactly what this is. Yeah. It's talking about, you know, just the use growing by the number of people using it in their devices and then the intensity at which they're using it, which is just means more compute, more token. You multiply these two together and you get the purple line. Okay. So now, now go to the next slide. >> Wait, wait. So this let me just take a quick look. We are here in July 2026 and you just want to show that because people people need to understand that that exponential cur grow growth of demand has already been happening. >> That's already happening. >> It's not it's not weird to think oh it's just going to keep expanding. Yeah. But remember this is logarithmic although >> that's logarithmic right. Yeah it's important it's important to recognize that that that this is a logarithmic graph and we'll show you what that means in a minute. >> Okay cool. So the demand is just explosive as you said. Uh demand is explosive. Okay. >> Yeah. So now go here. This is interesting. And now what I did was I said, "Okay, take that take that purple line that that we showed in the last graph and plotted against my curves." >> Yeah. >> My blue blue, you know, 3.4x, which you know, I came on a historic basis of what we've been doing. >> Okay. And then >> you were underestimating the demand. >> Yeah. No, it's crazy, right? I'm underestimating the demand. I mean, really? I mean, >> and you're trying to push it down. And >> truth is, it's actually dismik. [laughter] >> Yeah. No, I mean, by the way, the the the purple line does represent without these types of efficiencies built in. So, you could go ahead and push the purple line back down to the orange line. I don't care. We're still in a problem. >> You know, we're still in a problem space. You could push the purple line down to the blue line if you want and say, "Okay, with efficiencies, you get you catch the 3.4x um year-over-year." I don't know. You know, it's it's um look, I guarantee um irrespective of um here's what I here's what I think is absolutely true. The raw unfettered demand >> Mhm. >> is is will come true. Like it it's very clear to me where we gain efficiencies in the loop and that includes hardware efficiencies which in some ways have tapped out. uh software efficiencies, algorithmic efficiencies are sort of where the where we're going, right? In other words, we look at things like, oh, Deepseek, it can do this much with, you know, 10x less compute or something like that. Um, you know, that that that is possible. There will definitely be optimizations. But, you know, the other part of the the other side of this is that the optimizations may bring us down, may bring us back to where we currently are. And it doesn't change much because um you know I was listening to a podcast the other day where um I think his name is Jeffrey Hinton uh kind of known as the father of AI. >> The grandfather of AI. Yeah. >> Yeah. He said that the human brain has about equivalent to you know 100 trillion uh neural connections and that today's leading frontier models are are climbing into the two trillion space and uh so it means that in a sense the scaling laws are holding in AI and it means that people are demanding larger and larger models and um I have no doubt that the industry as a whole will climb into that space. So that's 50x the size and it's um it's a bit of a quadratic when you add what I mean by that is the >> right >> yeah you you've got it's not just the model it's not just the model size but >> to fill the model with the right parameterized data you've got to kind of double the data sizing too. So you've got both of those things going on in parallel when you you know you build a bigger model you push more data through it otherwise it's meaningless like the like what his argument was is the brain is a very large connection space with very little data in it in a sense he was talking about the experience data in the brain is very minuscule where he's saying that AI have overloaded on the experience data and sort of sense what they're trying to do is grow the brain bigger you know that's really what they're doing so um so what I'm saying is imagine that you get really good algorithmic efficiency But the reality is that you're dragging the compute side further up. So maybe you bring it back down to where we are, which is still at 3.4. So, >> yep. And also, uh, my understanding is that not only are we moving to aentic AI, I think this year more agents agentic AI did searches on Google than humans did. We we passed it this year. And so then agentic AI is unlimited. like you can I guess your graphs were showing users humans uh adopting AI and then asking AI to do things but then there's also now once you get inentic AI gentic AI ask other agentic AIS and then it just you can have like a million agentic AI per one person kind of thing you know >> it's there's a lot of reasons why you can see pretty easily the explosive growth side of things you know like you said you have essentially what you have is today you have AIs that are rather uh responding to our queries is your you know, they're like, but there's so much AI that's now doing work on our behalf while we sleep, so to speak. You know, what we would call asynchronously operating behind our back, which means it's still computing. Um, and then the types of things that they're doing, you know, that's why I was talking about generative use cases in in in um video or real-time gaming and those types of things where, you know, I mean, the the what is the future of Netflix? It's probably you and I design our own shows and and um and we see a show once because we basically or we change the the actors in a plot, you know, we like, "Oh, we like that series, but let's throw our favorite actors in it." >> Oh, I don't know how many times my wife has said this wrong casting. I would have chosen this person. >> Yeah. So, that, you know, that'll come because it's obvious that it will. And um >> Yeah. >> And that'll just drag us more into that demand side. Anyway, if you can continue with my lovely little graphs. I have like three more or two more. By the way, now I took it and I I just plotted it on a linear basis. And um >> yeah, there you go. This is a better way to show it. My >> it is. So the so the the y-axis is now linear. And that's >> scary huge. >> And I'll show you why it's scary huge. And I think the next slide does it. >> I just I just want to pause it for a second with the years. 2026, >> 2030. >> Yeah. I mean, and then 2032, 20, whatever. We're just we're not we're talking less than a decade. This things happen fast. Just think about where 2016 was. >> I mean, an iPhone was developed in 28 >> 2007. >> I mean, it's it's like this is going to go in a, you know, when my daughters were, you know, just go back how quickly time flies. >> I mean, just look at 2030 how quickly these lines slope up even if you even if you take the efficiency line. Um because this is yeah >> yeah we're getting you know you get huge huge sloping um so [sighs and gasps] and by the way I nothing in my brain said that this is not true like I you know now what's really interesting is to compare it against the little green line you know what the little green line is >> that's that is I I've layered in um a time period I've compressed and layered in the growth rate of the CPU the CPU error that includes >> which was more and Crazy. >> Yeah. That includes everything that we talked about in Mo's law. That's the growth rate through the >> gotcha. >> PC error and the >> internet, >> smartphone error. >> A smartphone error. And and of course we all know how incredible that that was and that's how come we got um you know my smartphone is equivalent to right you can do these comparisons to what people were using for the NASA whatever a thousand times a million times more right. >> Yeah. >> That is the point. If you were sitting around in 2010 and you thought the industry was growing like mad and you know everything was this that and that about tech then compare what that looks like on that little green line you know it looks kind of flat >> and by the way it didn't feel flat that's the point right didn't feel flat felt like felt like the world was exploding >> oh my gosh yeah yeah all the change has happened in 10 years has been crazy I just telling people right now I mean I've lived through right analog phones pick it up and dial to cell phones. I've done through everything, right? Computers, uh, keyboards coming in for the first time, internet coming for the first time, um, AI now coming for the first time, and I haven't seen anything yet. [laughter] >> Robots coming for the first time. It's just that, you know, these the types of things that we're doing with these AI require it's it's sort of like the first time we've had the ability to crunch this many numbers at in in a small unit of time and through this incredible parallelism, you know, I mean what I'm saying is AI researchers who thought about neural nets a long time ago said they're all theoretically possible but impossible to build because there's no way to do the compute either on the training side or the inference side, you know, the uses, you couldn't do it. So when we got to the GPU era and we got the parallelism um we started to see a path and of course what happened is not only did we see a path but it meant that people could experiment in that path and they did and then we got outcomes and then other people followed those outcomes and of course that the whole thing started to explode and you know we're sitting uh really on the flat part of these curves right now. You know, >> you could actually be completely underestimating honestly because there's a thing called convergence, right? Uh Ray Coswell's been talking about this where if one tech improves, it improves all tech immediately and vice versa and then it just >> Well, I mean, [snorts] you know, a good way of thinking about that Herbert is that there isn't any single robot. I mean, we thought if we thought the world of robots in, you know, last 20 years, whether they're the Roomba style or stuff that was made by uh was it what is the company? Boston Scientific, >> Boston Dynamic. Yeah, >> Boston Dynamics. >> Boston Dynamics. Um, >> you know, those robots were basically AI free. You know, they're just little, you know, now we're in a place where you don't build a robot without AI. Now, some will do it better than others, but the reality is that the amount of computation done in an Optimus with an AI5 is just it it it's mind-blowing. um just just to manage speech conversion you know in other words input output the HMI components to the human so you know I think you're right it's it's there's no doubt that again the audience can argue with whether which line to pick whether the orange is still too aggressive in terms of the demand side in other words whether efficiencies will will drag us closer and closer to say the um to the green line but even if it twice the green line. Uh, you know, Terapab's right in there for being an absolute necessity. And and that's just the production side. You know, you got to put it into use. And and this is why I have this slide called Darmine to the rescue because even if Elon built every chip, even if he he realized the dream of one terowatt of annualized compute, we can't power that on Earth. We simply can't power that on Earth. I mean the United States produces about has the capacity produce about a terowatt generatively uh produces about half a terowatt uh and um and we're talking about trying to run compute at a terowatt of annualized production. How would you power that thing? How would you power that whole beast? And the answer is obvious. It's staring right back at us here. the son. >> Um, you absolutely like like when you're when Elon says, you know, I'm gonna add two I'm, you know, I'm gonna have two gigawatt capacity or something like that in Colossus or I'm adding I you know, forget the numbers. It it it's a very small number compared to one ter and and and if you take a trip to Colossus and you go driving by there and you look at the number of behind the meter national natural gas generators powering that beast and you think about trying to replicate that out to the terowatt you'll never get there. You'll never get there. um China won't even get there and China sort of doesn't care about how it produces power at nuclear, coal, solar, you know, makes the United States kind of look like we're just we're not even playing in the game. Um so what what what's required is not only the production side of these chips, but then the con the the running of these chips and you're going to run them you have to run them in space. And the thing about space is you know the beautiful thing about the space idea is it's it really is infinitely scalable right this is the thing that when people think about and they say oh you know Elon wants to put a million but it's like a factory right if it can produce one doughut it can produce 10 donuts or or let's say it has a you know ability to produce a 100 donuts a minute then it can produce you know you can replicate that process and produce a thousand a minute and so forth and and so what he's saying is look once I get the tech stack uh I should I should have broadened that word but like the platform the platform for launch of course is is is Starship it has to get to reusability principally on ship um once I get the design set for this AI satellite and and by the way um he has a lot under his belt there right I mean it's not as if SpaceX today isn't the larger produc largest largest producer of satellites in the world. They are Star Star Link. Um, once he gets that thing cranking, uh, placing a million satellites in space isn't it's not like you're running out of room. >> No, >> in fact, that's just because they're sitting around the sunb belt orbiting Earth. There's nothing that says you can't push them further into their own orbits that are, you know, um, >> actually, the closer to the sun probably the better. That's why he wants to he wants to use the moon to launch these things because it's easier and you can get closer to the sun. >> I guess that I guess the point is that I think we're entering an era where there really is unlimited use for intelligence. We don't complete see it today, but it's pretty obvious, especially if you start to think about Yeah. Um, we kind of talked about that. So with this unlimited demand uh that demand gets satisfied through you know I mean massive chip production has to occur. Uh Terapab is the beginning of that and um and once we get to that point now we have to deploy it and run it and that can't be done in data centers on the planet. It's not really possible. Um I mean the data center to run the compute and to you know If you think Terrafab is a big building to run that compute might look like an equally bigger building, you know, um, and so that's not going to happen. Part two isn't going to happen. Part two is basically satellites. And to me, it's there are challenges to building the I1 satellite, uh, so forth and so on. But I think, you know, they have a lot of this well well understood. Um, they've been doing this for years with Star Link today. And you know there what's the difference between a Starlink satellite obviously the the structure of compute is completely different on Starlink satellite it is all about comm you know its job is communications that's its principal job whereas the AI one satellites are um star mine satellites are all about compute and so really they're just in a sense just larger heat generators is really what it comes down to. I mean if you really distill it into its essence um the compute part itself isn't confusing you know you you put the Nvidia chips then you put the eventually teraf D series chips there um you're the the collection of energy isn't confusing they know exactly what to do and how to do it u the radive part where you have to dissipate the heat is is fairly well understood people say they have no idea how to do that. Well, they're doing it right now. Every sing every single Starlink satellite is doing exactly that right now. You know, it has to um it's a bigger problem. I mean, like in other words, it's a bigger platform for the same type of thing. Satellites dimensionally bigger in terms of their uh energy in, heat out, but uh at the end of the day, >> they should do it. >> Yeah. >> Okay. Okay. So, a couple of things I wanted you to take a look again one more time at Elon's post because first of all, we pointed out already he called the terraab Texas will be the largest and most valuable building on earth by far and it'll be stunningly beautiful. >> Texas meaning to say may maybe hinting that there will be other tariffs and that's what you kind of pointed out that there probably is a need for another one in the future. Just incredible. But here he said this um so his guess is that terafab AI compute output would be 25% for Tesla Optimus 75%. But Phil look what the words he used. He said AI spacecraft. He didn't say AI satellites. If it was AI satellites it would refer to Starlink Star. But what's AI spacecraft? >> Yeah I think he's using it as a I think he's >> combo word for those for both of them. Or is it more than that? You know, I said hint more. >> I I I don't know. I mean, um, seems to me that what what I think he might be alluding to is just sort of the multiplicative effect of different products that would ultimately like in other words, the product line of AI compute satellites, those that have these orbits, those that are larger scale later that have maybe some type of orbits. um you know the the the the interesting reality >> bases of course. Yeah. >> What's that? >> Moon bases. >> Yeah. Yeah. You could uh think of of when when you think about compute the biggest issue about compute is bringing it back to earth. What I mean by that is if you inject compute too deep into space then it has to be rather asynchronous. It has to be completely agentic because it can't talk to us in real time. Um but if you bring it in a in an earth-based orbit and in effectively a low earth orbit, I mean like today you look at um Starlink. The reason that Starlink works for us and and we could do this very podcast in live form over Starlink. I mean we're talking live. So we have to be able to have very low latency connection is because of its distance is only you know those Starling satellites are only 300 miles over our head and that leads to um very very low latency and then if you attach that to the compute side in other words most of Star Mind as a if you think of it as a collective whole it isn't really about a training center in space it's an inference center it's all about our use of AI um and so it needs to be close enough to us so that it can be effectively a real-time AI, right? >> Um, and I made a a I posted an article the day of the um SpaceX earnings where I basically said that uh it's a it's worth a whole other discussion, but what I basically said is that Optimus in its really its realized form is probably running in Starmind and um that your Optimus on the ground the physical optimist is merely a physical manifestation and most of its compute on board is [snorts] mostly about uh AI to um interact with us and operate the actual bot itself like the movements of the bot um but its broader mind its broader context and its permanent existence is probably running in Star Mind. >> Yeah. um which allows that that Optimus experience be kind of allows you know the whole lead into digital optimist and the idea that Optimus just becomes kind of a thing that we interact with over a very long period of time. Like in other words, we could buy our Optimus without actually getting the physical bot a year before the bot shows up. Why? Because we're just basically essentially buying into digital optimist and we're starting to build a relationship with that thing. Where do you think it's going to run? I claim it's going to run in Star Mind. You know, all the instances, if you have a billion instances, two billion instances of Optimus in its physical humanoid form running on planet Earth, I I guarantee there are two billion instances of the mind sitting in Star Mind. And that is inference compute that really hasn't even been accounted for in a sense, right? because it's it's a whole different idea that it uh compared to the way we think of it today is running some model in space to do you know kind of like uh chat GDP on steroids well this is sort of like how do you operate these robots and uh and and why can you do that why can why is that possible it's possible again because the latency is low enough to still have that be a real time a real-time presence with us um you know if you put a satellite If if if you if if you put that compute on Mars, you couldn't do that, right? No, it could compute the hell out of a lot of stuff and it could have its own connection to every Optimus bot that's running around on Mars, but no real-time connection to planet Earth. You know, you're you're at a minimum of I think 8 minutes round trip away or something. So, um I don't think it says that much, but >> yeah, I like I like the way you positioned it that I I hadn't heard before, which is that these satellites are low orbit, which is 300 miles up in in sky in orbit or orbit, which is like when you said that made me realize, okay, well, that's halfway between San Diego and uh San Francisco, right? It's like it's 500 mile difference, right? And so it's like yeah that's faster for me to use an you know rather than my you know when I'm talking to you right now all these bits are going through some data center somewhere on earth and sometimes I'll have to go around the earth right it goes through underneath through the oceans and all that um >> even through round trip right >> yeah this >> this is faster >> this 300 miles >> this podcast be could be done I could be sitting on a boat in the middle of the Atlantic uh >> and still just be 300 miles away from data center. >> Yeah. >> Or a switch. Okay. Wow. This is a powerful research that you did there. This is terrafab. This is the rendering and then uh any any thinking about time frame, Phil. Um you kind of put in your graph there that there was a date of >> I don't know. >> 2029. Yeah. >> Yeah. I don't know. I mean come on Elon, get building. You're not going fast enough. [laughter] Well, first thing is to design it. >> Yeah. >> And uh once he's got We know how quickly they can put up buildings. Look at the Optimus uh building that's already being put up. >> I this thing I don't know. He's going to probably he'll probably make some comments um soon. But look, it's not there's no accident that he picked Texas. There's probably no accident that he made a deal in Grimes County and that, you know, 90% of the problem if you were to if you were to do this in California, you know, you'd spend the first three years in permitting. >> Yeah. >> Um but that's not what happens in Texas. You know, they want him to build and he's going to build as fast as he can he can get through the design uh and materials phase and it yeah, they'll break ground. Uh I don't know when, but it'll take >> Just playing the video one more time. take months to clear the land and prep the land. I mean, this is a very large space. I don't know how many physical acres this is, but >> I'm I'm just wondering if there Oh, I'm sorry. I thought I was playing it. Hold on. Let's start again. I'm wondering if there is any kind of um hints. Some people are going, "Look at the cyber cab. It