SpaceX Just Revealed Terafab Phase 1 — Location Confirmed

Kind: captions Language: en [music] Something enormous is beginning to take shape in Texas. Something that could reshape artificial intelligence, [music] robotics, advanced manufacturing, and even the economics of space. Let me show you. Back in March, Terapab was little more than one of Elon Musk's most ambitious presentations. [music] 5 months later, SpaceX and Tesla had selected Grimes County as its future home. The first phase alone is estimated at $16.8 8 billion and the full plan calls for more than 100 million square ft of manufacturing space. >> So then the key missing ingredient is therefore a terowatt of compute. So this announcement is about solving the key missing ingredient. We either build the terapab or we don't have the chips and uh we need the chips. So we're going to build terap. >> This facility is not being designed for one product. It would manufacture AI5class chips for Optimus robots and self-driving cyber cabs, while a separate D3 architecture is intended for something far stranger, AI data centers operating in space. Together, Tesla and SpaceX say their future demand could exceed 1 terowatt of compute. That number sounds almost impossible, but the reason this story is worth following is that several pieces already exist. Tesla has broken ground on a research fab. The final AI5 design is complete and its first large-scale Optimus production lines are being installed right now. This is the story of what SpaceX and Tesla want to make the largest chip manufacturing project on Earth and the ecosystem Musk believes will consume [music] everything it produces. Hey guys, Alfie here. Subscribe if you want the engineering behind the headline and let me explain how this entire system is supposed to work. And you know, Tesla and XAine and SpaceX have all done amazing things uh that people did not think would be done before. Uh so there's the the Giga Giga Texas fab here. Uh there's, you know, the the the Optimus robot that's being built. There's a a global supercharging network. Um there's really quite a lot. Um and it wasn't that long ago when people thought electric cars would wouldn't amount to anything. Um, and there were there were basically no electric cars for sale when uh when Tesla started. Um, 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, uh, has also built the first gigawatt scale compute cluster, which uh, in record time. Uh, Jensen Wong from Nvidia said he'd never seen anything built so fast in his life before. So uh, a great compliment from from Nvidia. And then SpaceX, uh, well, I guess you can read it for yourself. I Well, you already know. I mean, the reusable rockets, uh, 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. Um, and now we've landed over 500 times. Uh, and then we did the Falcon Heavy, and now we're doing Starship. And Starship is is a critical piece of the puzzle because in order to scale uh compute and scale power, you have to go to space, which means that you need massive payload to space. Um and Starship will enable that. In order to get to the terowatt uh of compute per year, uh you need about 10 million tons to orbit per year and at 100 kow per ton. So, but we're confident this is feasible like no new physics or impossible things are required to to get there. So, then the key missing ingredient is therefore a terowatt of compute. So, this announcement is about solving the key missing ingredient. [music] In March, when Musk introduced Terrafab as an almost absurd chip building effort, the presentation showed Optimus robots, [music] self-driving cyber cabs, and space-based data centers, all dependent on custom silicon. It looked like a collection of futuristic projects. In reality, it was a supply chain diagram. But this month, SpaceX and Tesla confirmed Grimes County, Texas as the planned site. Phase 1 alone is estimated at $16.8 billion. SpaceX say the project will create at least 3,000 jobs while a smaller research fab already under construction at Giga Texas will [music] act as the precursor. More than 100 million square ft of planned manufacturing space. By total floor area, that is roughly 15 pentagons, but the wilder number is 1 terowatt. SpaceX and Tesla say their combined future demand could exceed 1 terowatt of compute. To be clear, Terraab is not [music] a power station generating 1 terowatt of electricity. The number describes the projected [music] scale of computing hardware these companies believe they will need. It is still a staggering [music] claim and for now it is a projection, not production. That distinction matters because the real breakthrough is not just volume. It is the way Terrafab is supposed to operate. Today advanced chip production is a relay race. One company designs the processor, another manufactures the logic. Memory comes from somewhere else. Packaging and testing may happen in different facilities, sometimes in different countries. Every handoff adds time, cost, and risk. Terrafab is designed to pull logic, memory, advanced packaging, and testing under one roof. The goal is a much tighter loop. Build a chip, place it in a robot or spacecraft, [music] collect real world data, find the bottleneck, revise the design, [music] and push the next version back into production. That is the same vertical integration strategy Tesla used with batteries, software, charging, and manufacturing. Only now it is being applied to the most difficult [music] supply chain in modern technology. And this is where the picture splits into two very different futures. Musk's very approximate allocation is that about 25% of Terrafab's AI compute output would go to Tesla's [music] Optimus program. The other 75% would go to AI spacecraft. Start with the smaller share because smaller becomes ridiculous very quickly. Tesla completed the final design of its AI5 inference chip in April. Musk claims it delivers roughly five times the [music] useful compute of the current dual AI4 system with significantly more onchip memory and bandwidth. Unlike a giant training GPU sitting in a data center, AI5 is intended for inference, taking a trained neural network and making decisions in the physical [music] world in real time. That is exactly what a cyber cab needs when a pedestrian steps into the road. And it is what Optimus [music] needs when it picks up an unfamiliar object without crushing it. Now do the rough math. 25% of 1 terowatt is 250 gawatt. If an Optimus class AI5 system can operate near 250 watts, that represents enough compute capacity for about 1 billion Optimus robots. Tesla says the former Model S and Model X line in Fremont is being converted into its first largescale Optimus factory designed for eventual capacity of 1 million robots a year. A second generation line at Giga [music] Texas is being designed for a long-term capacity of 10 million, but designed capacity is not current output. As of the latest update, the first lines were still being installed, and the earliest Optimus robots were intended mainly for training, data collection, and development, not customers. [music] Tesla has repeatedly pushed Optimus production targets, and Musk has acknowledged that virtually every part of the robot needs a new supply chain. Think of this first phase as an Optimus Academy. Early robots help collect the examples that train later robots. those later robots may eventually enter Tesla's factories and help manufacture the next generation. It is a recursive plan. [music] So, Terrafab does not prove that mass market humanoid robots are around the corner. It reveals the scale Tesla is building for if that problem is ever solved. And remember, Optimus receives only the estimated 25%. The other 75% is where Terrafab stops looking like a Tesla project and starts looking like the manufacturing engine of SpaceX [music] AI. After SpaceX absorbed XAI, the vision for Musk's AI infrastructure became much larger. Instead of relying only on massive data centers on Earth, XAI could eventually spread its compute across both terrestrial facilities and orbital data centers with SpaceX launching the hardware and Starlink connecting the network. On Earth, giant AI facilities compete for land, grid connections, water, and power. [music] In orbit, solar energy is far more continuous, and Starship could theoretically launch compute modules at enormous [music] scale, but space creates a different enemy, heat. On Earth, a data center can move heat into air or water. In a vacuum, there is no air to carry it away. [music] A spacecraft must radiate heat, which requires large, heavy surfaces. That is why Musk has [music] described a separate D3 chip optimized for space and designed to run hotter. The hotter the chip can safely operate, the less radiator area and therefore mass the spacecraft may need. >> To give you a sense of um what we're talking about, uh the current output of AI compute uh is roughly 20 GW per year. This chart explains why we need to build the terra fab [laughter] 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 Terraab project. So uh you know we we we certainly uh want our existing supply chain to be clear. We're uh very grateful to our existing supply chain uh to Samsung, TSMC, Micron and and and others. Uh and we would like them to expand as quickly as they can. Um and we will buy all of their chips. Uh I have said these exact words to them. Um and uh but but there's there's a maximum rate at which they're comfortable uh expanding, but that rate is u much less than we would like. And so we we either build the terrafab or we don't have the chips and uh we need the chips. So we're going to build terraab and we're starting off with an advanced technology fab here in Austin. So in the advanced technology fab uh we 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 uh equipment necessary to make the lithography masks. So in a single building we can create a lithography mask uh make the chip uh test the chip make another mask and and and have an an incredibly fast recursive loop for improving the chip design. Uh 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 and just keep looping it. So, and we're not just going to do conventional compute in this. I think there's some very interesting uh new physics uh that is potentially that that actually I'm confident will work. It's just a question of when. Um, so this is going to we're really going to push the limit of physics in 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 and then make and then change the design do another one and have that in a single building. >> That one design decision shows why Terraab cannot simply stamp out the same chip for every product. AI5 is optimized for efficient edge inference inside cars and robots. D3 is intended for high power compute in an environment where every kilogram affects launch economics and SpaceX controls the one machine that could make those economics work. Starship now the ecosystem closes its loop. [music] XAI develops models. Terrafab produces specialized compute. Tesla deploys edge chips across robots and autonomous vehicles. [music] SpaceX launches the orbital hardware. Starlink connects it. Real world data flows back into the models and the [music] next generation of chips is redesigned around what the system learned. This is also why Intel's involvement matters. Tesla [music] can design AI processors and SpaceX can design spacecraft, but operating a leading semiconductor fab is an entirely different discipline. It requires decades of process knowledge, extreme precision, specialized equipment, and [music] yield management. Intel brings legacy manufacturing experience while Musk has pointed to its 14A process for Terrafab. At the same time, Tesla is not abandoning its existing suppliers. AI5 is already tied to Samsung and TSMC. Terraab is a long-term attempt to fill a demand Musk believes the existing industry cannot satisfy, not [music] a magic switch that replaces the global chip supply chain tomorrow. That may be the most grounded reason for this gigantic bet. Advanced chips are now infrastructure. Cars, robots, [music] satellites, secure communications, and AI systems all depend on them. While much of the world's most advanced production remains concentrated overseas, bringing more logic, [music] memory, and packaging capacity into the United States gives Musk's companies more control, but it also aligns the project with America's larger push to rebuild semiconductor manufacturing. Tesla is becoming the physical AI company. cars and robots acting in the real world. XAI supplies the intelligence. SpaceX supplies the network, the launch [music] system, and potentially the orbital data centers. And Terrafab is intended to supply the one component every layer consumes. That is why this is bigger [music] than a chip factory. Musk is not building Optimus, Cyber Cab, Grock, Starlink, Starship, and space data centers as separate beds. He is trying to make each project a customer [music] supplier and training system for the others. The cyberpunk city in that recent video is not really the product. The product [music] is the loop and terraab may be the final piece that allows that loop to scale or the bottleneck that proves the vision was always too large to manufacture. >> What do you do after a terapab? U don't think small. Uh well yeah good point. Uh so we um you know how do you get to a pedawatt uh is is the obvious next question and um you get there by having um an electromagnetic mass driver on the moon with robots with optimi and obviously lots of humans and uh with that you can send uh a pedawatt you can create a pedawatt of compute and send that to deep space because on the moon moon has no atmosphere and has 16 earth gravity. So you can you you don't need rockets on the moon. You can literally accelerate it uh to escape velocity from the surface and uh that dramatically drops the cost once again of harnessing power and uh and enables you to go a thousand times bigger than a terowatt. So, for sure the future I want to see I I want to just live long enough to see the the mass driver on the moon cuz that's going to be incredibly epic.