TERAFAB: Elon Musk’s Trillion-Dollar Gamble to Build a Space-Age AI Empire

Kind: captions Language: en Imagine walking outside tonight, uh, looking up at the stars and knowing that somewhere in the dark, a massive solar powered supercomput is silently floating right above you. >> It sounds totally crazy, right? >> It really does. But, I mean, that isn't science fiction anymore. According to a highly detailed, up to-them minutee research report dated today, April 18th, 2026, that is the ultimate goal of a brand new multi- trillion dollar manufacturing project. We've got our hands on the details of Elon Musk's staggering new endeavor called Terraab. And well, we are so glad you are here with us to explore it today. >> Yeah, it is an incredibly ambitious plan. I mean, the announcement at the Seahol Power Plant in Austin certainly sounded like a script from a sci-fi movie. >> For sure. >> But I want to establish right out of the gate that the supply chain economics driving this massive shift are very real. They're very terrestrial and frankly, they're rooted in a pressing industry crisis. >> Okay, let's unpack this. Our mission for today's deep dive is to explore how Terrafab intends to completely upend traditional semiconductor manufacturing. I mean they want to pull a fragmented global industry under one gigantic roof, >> right? >> But we also need to understand why this earthbased mega factory is actually just the first stage of a much larger transition to orbital space. Before we even talk about the stars though, we have to start with the why. Like why build this completely unprecedented factory at all? Well, according to the report, it comes down to something called the 2% crisis. >> The 2% crisis. >> Exactly. To understand why Terra Fab exists, you really have to look at the combined computational demands of Musk's three main companies. First, you have Tesla, which is pushing the limits of full self-driving and, you know, preparing to massproduce their Optimus humanoid robots, >> which take a ton of compute, >> a massive amount. And then you have XAI, which is building these massive multimegawatt superclusters to train next generation artificial intelligence. And finally, you have SpaceX, whose long-term roadmap requires vast amounts of orbital data processing. >> Wow. Okay. >> So, the computational demand from those three entities isn't just growing linearly. It's on an extreme exponential curve. >> Wait, let me jump in here. So, it's not just Tesla needing chips for cars. It's XAI needing them for huge language models and SpaceX needing them for satellites, right? >> They're essentially cannibalizing their own supply chain across these massive industries. and the current global setup simply cannot handle it. I mean their report notes that if you look at the traditional foundaries the giants of the industry like TSMC, Samsung and Micron and you project their manufacturing capacity forward into the next decade they can only meet about 2% of this future AI demand just 2% >> which is a catastrophic bottleneck. I mean if your entire business model relies on scaling autonomous intelligence running out of compute is a fatal flaw. Definitely >> there is a very blunt quote from Musk in the report that perfectly summarizes the forcing function here. He said we either build the terraab or we don't have the chips and we need the chips so we build the terra fab. It's purely an equation of survival. >> So to solve this bottleneck they are pivoting away from the traditional foundry model toward extreme vertical integration. >> Yeah. To put this into perspective for you, imagine you are trying to run the most ambitious high volume restaurant in the world. Under the current semiconductor model, your kitchen, where the food is actually cooked, is in Taiwan, >> right? >> Your pantry, where you store specialized ingredients like memory chips, is in South Korea, and your dining room, where the customer actually eats the meal, is in the United States. It is incredibly fragmented, and you know, shipping takes forever. >> Oh, absolutely. Terapab is essentially deciding to grow your own wheat, mill your own flour, and bake the bread in the exact same building so you can serve the customer piping hot immediately. >> That is a perfect analogy. Honestly, the modern semiconductor supply chain is brilliant, but it is deeply deeply geographically fragmented. What's fascinating here is that Terrafab wants logic manufacturing, memory production, specifically high bandwidth memory like HBM4 advanced packaging and final testing all under one physical roof. Okay, but let's define some of the jargon for everyone listening. You mentioned advanced packaging. The report specifically talks about something called cos. What does that actually mean physically? >> Right. So cos stands for chip on wafer on substrate. >> Okay. Quite the mouthful. >> It is. Yeah. But the best way to visualize it is to think about urban planning. Traditionally, computer chips sit side by side on a flat motherboard, kind of like singlestory houses in a sprawling suburb. >> Got it? If data needs to get from the processor to the memory, it has to commute across town, which takes time and energy. Kwas is like building a multi-story high-rise. You stack the logic chips and the memory chips vertically. >> Oh, I see. >> So, the data doesn't commute across town anymore. It just takes the elevator. It drastically reduces the distance signals travel, which massively increases speed and lowers power consumption. >> That makes total sense. But taking all of that, the logic, the high-rise packaging, the memory, and building a state-of-the-art fab from scratch is famously impossible. >> Yeah, virtually impossible. >> Which brings us to the crucial April 7 partnership with Intel. Intel is stepping in to provide their cutting edge 18A process node, which is their 1.8 nanometer technology. >> Yes. And we really cannot overstate the physics of what 1.8 nanometers means. I mean a single silicon atom is about2 nanometers wide. So at 1.8 nmters you are literally sculpting electrical pathways that are less than 10 atoms across. >> That is just mindboggling. >> It really is. The margin for error is essentially non-existent. This Intel partnership potentially through a buildup or transfer or licensing model is the realistic anchor that makes Terrafab viable. Intel CEO Leboutan explicitly called this a step change in refactoring silicon fab technology. But Intel's partnership is just IP on paper, right? To actually see if this unified stack works, they have to physically prove it without blowing their entire budget. And that's what is happening right now in the dirt of Giga Texas. They are building a massive prototype called the Advanced Technology Fab. It's an initial investment of 20 to$25 billion targeting an output of a 100,000 wafers a month. But the number that absolutely stopped me in my tracks was the iteration speed. >> The 9-month cycle. Yes, they are targeting a 9-month product cycle, >> and that is completely unheard of. I mean, in standard semiconductor manufacturing, going from a chip design to a finished product takes years. >> Why though? If Tesla wants to shrink a product cycle to 9 months by literally just walking down the hall instead of shipping a wafer across the Pacific Ocean, what is the actual friction in the traditional model that slows everyone else down? Doesn't that make the rest of the industry look like they're moving in slow motion? >> Well, yes, it does. But achieving this is incredibly difficult. The friction is in the physical mask making and testing loop. Think of a mask as the highly complex stencil used to print the circuits onto the silicon. In the traditional model, you design the chip in the US. You send the data to a specialist facility in Asia to create the physical mask. The mask goes to a foundry in Taiwan to print a prototype wafer. That wafer is shipped to another country to be cut and packaged, then shipped somewhere else to be tested. Good grief, >> right? And if you find a microscopic flaw during testing, which you almost always do, you have to send the data back to the US, redesign the mask, and start the entire multi-month, multi-country logistical nightmare all over again. Oh, and you have to wait in line at the foundry because they're also printing chips for Apple and Nvidia. >> Wow. So, by putting everything in one building, Terra Fab completely eliminates the shipping time in the foundry queue. You print the chip, test it down the hall, find the flaw, redesign the stencil, and print a new one the next day. >> Exactly. Consolidating fabrication, testing, and remasking in one location requires flawless logistical synchronization, but it removes years of dead time, and that speed is critical for their immediate hardware needs. The report details that their new AI5 chip has already taped out, >> meaning the design is locked and ready for manufacturing. >> And this AI5 chip is a monster. The specs show it hitting roughly 2500 TOPS. And just to give you a baseline for what TOPS or terra operations per second actually means, a brand new iPhone 15 Pro, which is incredibly fast, operates at about 35 Tops. The AI5 is doing 2500. >> It is a massive leap in processing power, and that 9month iteration speed directly feeds into Tesla's mass production goals. The AI5 is designed to power the brains of the Optimus humanoid robots. Tesla is aiming to produce 1 billion of these robots. >> 1 billion? That's a lot of robots. >> Yeah. And the real world is chaotic. A robot needs hardware level iterations to navigate unmapped environments. By shrinking the chip cycle to 9 months, Tesla can continuously upgrade the physical brains of these robots faster than any competitor. There's also mention of an AI 6 for the next generation Optimus and Dojo systems. It permanently transforms Tesla from a car company into the dominant autonomous technology leader. Okay, so a $25 billion prototype aiming for mass production by 2029 is mind-boggling on its own, but we have to remember this is just the prototype phase. When you look at the plans for the full-scale project, the financial and physical realities start to look genuinely terrifying. >> Oh, without a doubt. >> The full-scale Terafab is projected to be 100 million square feet. That is 10 times the size of the current Giga Texas facility. They are targeting an output of 1 million wafers a month, which translates to a 100 to 200 billion customized chips annually. And the capital expenditure, it scales into the trillions of dollars. >> Trillions, which is an entirely different economic classification. I mean, TSMC's largest, most advanced fabs cost in the tens of billions. >> Hold on, let me stop you right there because I'm playing devil's advocate for a second. I'm reading this and going from zero fab experience to a trillion dollar mega factory sounds like a recipe for a spectacular financial disaster. >> It's a huge risk, >> right? I mean, the report explicitly notes that teams are scrambling to get quotes from equipment suppliers like Applied Materials and Tokyo Electron, but you can't just throw money at this problem. They need extreme ultraviolet or EUV lithography machines to print these 1.8 nanometer chips. And those machines have a two plus year delivery backlog. >> You're absolutely right to be skeptical. And frankly, some media outlets are claiming this project reeks of desperation because Tesla, SpaceX, and XAI have zero prior experience in massive chip manufacturing. >> Exactly. >> EUV machines are arguably the most complex devices ever built by humans. You can't just invent these machines out of thin air to bypass a supply chain backlog. Add in a severe global shortage of specialized semiconductor engineers. And the hurdles are truly monumental. >> It's a massive reality check. >> It is. But we also have to balance this by looking at Musk's historical track record with engineering tasks the industry deemed impossible. >> That's true. >> The aerospace industry said landing a reusable orbital rocket on a drone ship in the ocean was impossible. The telecom industry said building a profitable low latency global satellite internet network like Starlink was impossible. He has a habit of announcing a timeline that seems completely detached from reality and then through sheer force of vertical integration forcing reality to catch up. But this raises an important question. >> What's that? >> How do you mathematically justify a trillion dollar earthbound factory? You don't. Unless the Earth factory is just a stepping stone to something much bigger. >> Here's where it gets really interesting. Because when you look closely at what Terraab actually plans to produce, 80% of their chips aren't for cars. They aren't for the Optimus robots walking around on Earth. 80% of the production volume is dedicated to a chip called the D3. And the D3 is a radiation hardened chip designed specifically for SpaceX. It is destined for orbit. >> This is the massive pivot that explains this sheer scale of the project. There are profound physical limits to Earthbound computing. On Earth, to build a mega data center, you need immense tracks of land. You need massive power stations. You need complex millions of gallons a day water cooling systems to keep the servers from melting. And you have to navigate endless zoning and environmental permits, >> right? >> But space changes the equation entirely. Space has no cooling limits, no land restrictions, and as the report puts it, it's always sunny in space. >> No, wait. You're probably wondering how space solves a cooling problem. Because I always thought cooling in space is actually incredibly difficult since a vacuum is a perfect insulator. It's like a thermos. You know, there is no air to power a cooling fan or carry the heat away. How are they solving that? >> It's a brilliant question, and you're spot on about the vacuum. Because there's no atmosphere to conduct or convect heat, you have to rely entirely on thermal radiation. They solve this by deploying massive specialized radiator panels that essentially bleed the infrared heat directly out into the deep freeze of the cosmos. Oh, yeah. And on the flip side of that equation, you have uninterrupted high-intensity solar power with no atmosphere to filter the light and no nighttime to interrupt the grid. >> Uninterrupted solar power and endless space. But space introduces a very hostile problem for computer chips, and that's radiation. >> Precisely. Outside of Earth's protective magnetic field, cosmic rays and high energy solar particles constantly bombard everything. >> Right. In a standard off-the-shelf silicon chip, if a single highly charged cosmic ray strikes the silicon lattice, it creates a microscopic electrical charge, >> which flips a bit. Right. >> Exactly. That tiny electrical spike is enough to flip a bit of data from a zero to a one. That's called a single event upset, and it can cause a massive system crash. Over time, that ionizing radiation physically degrades the silicon itself. >> And that explains why 80% of the facto's output is focused on the D3 chip. It has to be radiation hardened. You can't just put an Earth chip in a satellite. >> No, definitely not. >> They have to engineer specialized transistor designs, build redundant circuits right into the logic, and use unique physical shielding at the microscopic level so a cosmic ray doesn't wipe out the AI. >> Exactly. So, synthesizing all of this, Terrafab isn't just a factory. It is the terrestrial manufacturing engine required to launch an entire computational infrastructure into the sky. Internal chip production is really the only way SpaceX and XAI could feasibly construct a self-sustaining orbital infrastructure. The report states the ultimate logistical goal is to launch 10 million tons of mass into orbit annually, specifically to build satellite-based AI data centers. >> I love the official slogan of the project in the report. It says, "Terra will close the gap between today's chip production and the futures demand a future among the stars." I mean, just imagine staring up at the night sky knowing there's a massive solar powered supercomputer floating above you. >> It is a profound paradigm shifting image. >> So, what does this all mean? We started this deep dive talking about supply chain logistics and silicon wafer bottlenecks and somehow we have arrived at a road map for a galactic civilization because the ultimate production goal of Terrafab is to generate one terowatt 1TW of AI computing capacity per year. Just to put that in perspective, that is the equivalent of about 70% of TSMC's entire current global capacity just willed into existence from scratch. >> Yeah. And that one terrowatt threshold is not an arbitrary number pulled out of thin air. It directly connects to the Cardartesev scale. >> Right. The Cardartesev scale. For those who might need a quick refresher, this is a method developed by astrophysicist Nikolai Cardesev to measure a civilization's level of technological advancement based entirely on the amount of energy it is able to harness. >> Correct. A type A civilization can harness all the energy that reaches its home planet from its parent star. A type two civilization can harness the total energy of its star directly. Often visualized in science fiction as a Dyson sphere. >> Oh, like a giant shell around the sun. >> Exactly. by moving our most intensive computing up into space, powering it with uninterrupted solar arrays, and targeting that massive 1 terowatt threshold. This is an explicit engineered attempt to permanently move humanity from a type I planetary civilization toward a type 2 stellar civilization. >> It is absolutely staggering to think about. If they successfully execute this plan, it permanently transforms the entire Musk empire. Tesla, SpaceX, XAI. They stop being separate companies making cars, rockets, and chatbots. They merge into a unified autonomous energy and space exploration monolith. Think about the feedback loop. The robots build the terrestrial factories. The factories build the radiation hardened ships. The rockets launch the chips into orbit. And the orbital data centers power the AI that designs the next generation of robots. It's a closed self-improving loop of technology. >> It is the ultimate expression of vertical integration. But you know if we connect this grand vision to the bigger picture, we have to ground ourselves back in the immediate reality of today. The Carterf scale is the destination. Yes. But the journey begins right now in the dirt of Giga Texas. Yeah. Back on Earth. Right. The next two to three years at that $25 billion Austin prototype are going to be absolutely critical. Proving they can actually achieve that rapid 9-month iteration cycle. proving that Intel's 18A physical process can scale under a unified roof and successfully navigating the severe shortages in extreme ultraviolet lithography machines and engineering talent. Well, those immediate highly terrestrial hurdles will completely determine if this starry future ever gets off the launch pad. >> The boundaries between Earth and space, between a traditional factory and an orbital supercomput are definitely getting blurry. Thank you so much for joining us on this deep dive into the Terapab report. We've covered a ton of ground today, transitioning from the 2% Foundry bottleneck that kicked this whole crisis off to the Intel 18A Austin prototype through the terrifying trillion dollar risk of full-scale manufacturing all the way up to solar powered radiation hardened AI data centers floating in orbit. >> It is an immense amount of technical and logistical information to process, but it is undoubtedly one of the most significant industrial pivots of our lifetime. It really is. But before we sign off, I want to leave you with one final thought to ponder. Something we didn't really touch on today, but that naturally flows from all of this. If 80% of our most advanced AI computing power eventually resides in orbit, completely powered by the sun, constantly iterating and operating entirely outside of any single country's physical borders. Well, how will terrestrial governments ever hope to regulate an artificial intelligence that literally lives in the stars? Keep questioning, keep exploring, and we'll see you on the next deep dive.