À la centrale électrique de Seaholm à Austin, Elon Musk dévoile Terafab – une mégausine de puces commune à Tesla et SpaceX visant à produire plus d'un térawatt de puissance de calcul IA par an.
Elon Musk
Important announcement to make, which is the most epic chip building exercise in history by far. This is really going to take things to the next level. So, yeah, a level probably people aren't even contemplating right now. This is not in the, I'd call this sort of an out of context problem. It's not in their context. So we're going to adjust the context by a few orders of magnitude here. See, it's, it's a joint effort. I'm pressing the button, but the button's not working. Oh, there we go. Okay. So yeah, we aspire to be a galactic organ, galactic civilization. So I think the future that everyone, well, most people I think would agree is the most exciting is one where we are out there among the stars, where we are not forever confined to one planet, that we become a multi planet planet species. Like the best science, science fiction that you've ever read, you know, Star Trek or Ian Banks or Asimov or Heinlein. And we want to make that real. Yeah, not just fiction, turn science fiction to science fact. That's the glorious, exciting future that I certainly look forward to. And it's worth considering. Sort of like how would you rate civilizations? You know, there's so there was a physicist, I think it was Russian in the 60s, Kardashev, he thought about at a high level, how would you consider any given civilization? And he said, well, if you're type one, you're using most of the energy of your planet. And we actually still have quite a ways to go to be properly a type 1. We're still using a tiny fraction of the, the Sun's energy that reaches our planet. See, there we go. But the Earth only receives about half a billionth of the Sun's energy. So the sun is truly enormous. The sun is 99.8% of all mass in the solar system. So sometimes people will ask me, like, what about other power sources of power on Earth? Like, what about fusion on Earth? Well, that is unfortunately very small because the sun is 99.8% of mass in the solar system and Jupiter is about 0.1%. And Earth is in the miscellaneous category. We are, I think, as Carl Sagan I think might have said, Earth is like a tiny dust mote in a vast darkness. Very, very small. The sun is enormous. So the way to actually scale civilization is to scale power in space. This is necessarily true because we actually capture such a tiny amount of the Sun's energy on Earth because we're just this tiny dust mode. Another way to think of it is roughly like electricity Production on Earth of all of civilization is only about a trillionth of the Sun's energy, which means if you increase civilizational power output by a million, you would still only be a millionth of the Sun's energy. I mean it's, it's all inspiring to consider that just how tiny we are in the grand scheme of things. And yeah, we, we often get sort of caught up in the sort of these sort of squabbles on Earth that are really very sort of minor things in the, when you consider the, the grandness of the universe. And so I think it is important actually to consider the grandness of the universe and what we can do that is much greater than what we've done before, as opposed to worry about sort of small squabbles on Earth type of thing. Not much point in that. Yeah, we want to be a civilization that expands to the galaxy with spaceships that anyone can go anywhere they want at any time. That would be epic. And have a city on the moon, cities on Mars, populate the solar system and send spaceships to other star systems. That sounds like the best possible future. So to do that we need to harness the power of the sun. And so a terafab, while it is enormous, a terawatt of computer per year is enormous by our sort of civilizational standards. It is still just one step along the way of being even a Kardashev. You still have a long way to go to even be a Kardashev 2 level civilization and you're not even registering as a Kardashev. 3. So it's a very big thing by current human standards, but still small in the grand scheme and, but very difficult for humans. So to, to accomplish this very difficult goal really requires a combination of efforts of SpaceX, XAI and Tesla working together to create this epic Terrafor project. So, And you know, Tesla, Next, AI and SpaceX have all done amazing things that people did not think would be done before. So there's the Giga Texas Fab here, there's the Optimus robot that's being built, there's a global supercharging network. There's really quite a lot. And it wasn't that long ago when people thought electric cars wouldn't amount to anything. And there were basically no electric cars for sale when Tesla started and people said it was impossible. And Now Tesla's making 2 million electric cars a year. And then Xai, although it's a new company now part of SpaceX, has also built the first gigawatt scale compute cluster which in record time. Jensen Huang from Nvidia said he'd never seen anything built so fast in his life before. So a great compliment from Nvidia. And then SpaceX. Well, I guess you can read it for yourself. You already know, I mean, the reusable rockets. People said that reusable rockets weren't possible and even if you did do them, they wouldn't be economically feasible. So we did them and then we made them economically feasible. And now we've landed over 500 times. And then we did the Falcon Heavy, now we're doing Starship. And Starship is a critical piece of the puzzle because in order to start scale compute and scale power, you have to go to space, which means that you need massive payload to space. And Starship will enable that. So this gives you sort of just a sense of scale. We've got Optimus there, Optimus for scale, and optimus is about 511. So it gives you a sense of the size of the Starship V3 rocket. The Sasha V4 will be much longer. Actually, the Sasha V4 will, will make Starship V3 look kind of short. So we'll, we'll expand with Starship V3 to 200 tons of payload to orbit from 100 tons. We'll start with V3 and then you can see that just the, that's just a rough approximation of the, the AI, the mini version of the AISAT. So that's roughly 100 kilowatts. It's showing the solar panels and the radiator to scale. So. So for some reason there's been a bizarre debate about radiators in space. It's safe to say SpaceX knows how to do heat rejection in space. With 10,000 satellites in orbit, might know a thing or two. So you can see the radiator is actually quite small relative to the solar panels. And we call that the Minisat, since that's just 100 kilowatts. We expect future satellites to probably go to the megawatt range. No. So in order to get to the terawatt of compute per year, you need about 10 million tons to orbit per year, and at 100 kilowatts per ton. So we're confident this is feasible, like no new physics or impossible things are required to get there. So I'm confident actually that SpaceX will get to 10 million tons to orbit per year. And then we're building up to a terawatt of solar. So that solves the, well, will solve the solar problem, the power generation. So then the key missing Ingredient is therefore a terawatt of compute. So this announcement is about solving the key missing ingredient. To give you a sense of what we're talking about, the current output of AI compute is roughly 20 gigawatts per year. This chart explains why we need to build the terafab, because all of the rest of the output from Earth is about 2% of what we need. So if you add up all the fabs on Earth combined, they're only about 2% of what we need for the terawatt project or terafab project. So. So, you know, we certainly want our existing supply chain to be clear. We're very grateful to our existing supply chain to Samsung, tsmc, Micron and others, and we would like them to expand as quickly as they can, and we will buy all of their chips. I have said these exact words to them, but there's a maximum rate at which they're comfortable expanding, but that rate is much less than we would like. And so we either build the terafab or we don't have the chips, and we need the chips. So we're going to build terafab. And we're starting off with an advanced technology FAB here in Austin, and. I believe Governor Abbott is in the audience. I'd like to thank Governor Abbott and the state of Texas for their support. So in the advanced technology fab, we will have all of the equipment necessary to make a chip of any kind, logical memory, and we will also have all of the equipment necessary to make the lithography masks. So in a single building, we can create a lithography mask, make the chip, test the chip, make another mask, and have an incredibly fast recursive loop for improving the chip design. To the best of my knowledge, this doesn't exist anywhere in the world where you've got everything necessary to build logic, memory and do packaging and test it, and then do the masks, improve the masks, and just keep looping it. So we're not just going to do conventional compute in this. I think there's some very interesting new physics that is potentially that actually I'm confident will work. It's just a question of when. So this is going to. We're really going to push the limit of physics in compute, and we're going to try a bunch of wild and crazy things which you can do if you've got that fast iteration loop that I can't emphasize enough the importance of being able to make a chip, test it, and then make and then change the design, do another one and have that in a Single building. I think that our recursive improvement with that situation is probably an order of magnitude better than anything else in the world. So, broadly speaking, we expect to make two kinds of chips. One will be optimized for edge inference. So that'll be used primarily in Optimus and in the cars, but, but especially in Optimus because I expect the robots, humanoid robots, to be made 10 to 100 times more than the volume of cars. So, you know, if vehicle production, vehicle production, earth is about 100 million vehicles a year, and I expect humanoid robot production to be somewhere between a billion and 10 billion units a year. So it's a lot. So yeah, Tesla's going to make a very significant percentage of those is our goal. And then we need a high power chip that is designed for space that takes into account the more difficult environment in space. Where you've got high power, you have high energy ions, photons, you got electron buildup. It's a hostile environment in space. So you want to design the chip, you want to optimize it for space, and you also want to generally run it a little hotter than you would normally run a chip on Earth to minimize the radiator mass. So there are just a bunch of constraints that you design something differently in space than you would on the ground. And for the space compute, my guess is that is the vast majority of the compute because you, you're power constrained on Earth. That's why I think it's probably 100 to 200 gigawatts a year of terrestrial chips and probably on the order of a terawatt of chips in space, just because of power constraints on the ground. That's probably how it ends up. Space has this advantage that it's always sunny, it's very nice. So I actually think that the cost of AI and deploying AI in space will drop below the cost of terrestrial AI much sooner than most people expect. I think it may be only two or three years before it is actually lower cost to send AI chips to space than it is on the ground. Because in space you don't need much in the way of batteries because of it's always sunny. And the solar power you get, you're going to get at least five or more times the solar power you get in space versus the ground because you don't have atmospheric attenuation or a day, night cycle or seasonality. And you're always normal to the sun. So you're really maximizing the solar power at that point. And space solar actually costs less than terrestrial solar because you don't need heavy glass or framing to protect it from extreme weather events. So as soon as the cost to orbit drops to a low number, it immediately makes extremely compelling sense to put AI in space. It becomes a no brainer, basically. Moreover, as you go to space you get increased economies of scale and things get easier over time. Whereas as you try to put more and more power on the ground, you run out of space and you start using up the easy spots and, and then you get next level numby. Nobody wants the thing in their backyard. So then, so actually increasing power on Earth has, becomes harder over time and more expensive over time, but in space it becomes actually cheaper and easier over time. These are very important points. So. Yeah, Sam sa. So what you just saw there was because of course you're asking what's on your mind is well, what do you do after a tariff app? Don't think small. Well, yeah, good point. So we, how do you get to a petawatt is the obvious next question. And you get there by having an electromagnetic mass driver on the moon with robots, with Octomi and obviously lots of humans. And with that you can send a petawatt, you can create a petawatt of compute and send that to deep space because on the moon, moon has no atmosphere and has one sixth Earth gravity. So you can, you don't need rockets on the moon, you can literally accelerate it to escape velocity from the surface and that dramatically drops the cost once again of harnessing power and enables you to go a thousand times bigger than a terawatt. So for sure the future I want to see, I want just to live long enough to see the mass driver on the moon because that's going to be incredibly epic. Yeah, That should hopefully get us to a millionth of the sun's energy at least humbling to think about that. But a millionth of the sun's energy would be a million times bigger than Earth's economy. So it's good from that perspective. And then yeah, you expand beyond that to the planets, to the other stars and create the most exciting possible future than that. I can imagine. This looks a bit like the opening in Idiocracy with Mike Judge, unlocking an age of amazing abundance. So yeah, obviously the elements of that are sustainable energy, space travel and AI and robotics that bring amazing abundance to everyone. And it's really the, it's really the only path to amazing abundance is AI and AI and robotics. Which is not to say it can't go wrong hopefully, you know, but I think it'll probably go Right? And it'll be a future that you, that you love. And it's the best future I can think of, at least. And then we go beyond the moon, beyond Mars, and we sail through the rings of Saturn. Now, wouldn't it be amazing if you could buy a trip to Saturn? Or frankly, if you just have a trip to Saturn, I think things will just be free in the future. It sounds nuts, but, you know, if you've got an AI robotics economy that is anywhere close to a million times the size of the current Earth economy, literally any need you possibly want can be met. If you can think of it, you can have it. So I think Ian Banks in his culture books has it pretty much right. Where there actually isn't money in the future and there's abundance for everyone. If you can think of it, you can have it. That's it. Which means anyone could have a trip to Saturn. It won't be just a few people. If you want it, you can have it. Yeah. So, Yeah, join us on this journey and help us design incredible chips and make incredible chips and build a terawatt of chips, a terawatt of solar and 10 million tons to orbit per year. Thank you.