Musk’s TERAFAB Could Make Today’s AI Data Centers Look Like a JOKE!

Kind: captions Language: en March 21st, 2026, Elon Musk walks out onto the stage at a decommissioned power plant in Austin, Texas, which is already a very Musk thing to do and drops something that genuinely caught me off guard. And look, I cover this stuff for a living. I'm not easy to surprise anymore. He called it Terra Fab. Not a Marvel villain, not a new Tesla model, a chip fabrication mega project that if it actually works, might be the most consequential piece of infrastructure built in my lifetime. Maybe yours, too. And I don't say that lightly because I've sat through enough tech announcements to know that most of them are smoke. This one felt different. Let me explain why. Here's the core of it. Terraab is a $25 billion joint venture between Tesla, SpaceX, and XAI. The goal is to produce 1 terowatt of AI compute every single year. One terowatt. That's one trillion watts of computing power annually, coming out of what Musk described as the most epic chip building exercise in history by far. I mean, yeah. Okay, but here's the part that actually got my attention. The semiconductor industry today is a scattered globalized mess. [music] You've got one company designing chips, another making the machines that etch them, foundaries in Asia doing the actual fabrication, different facilities handling packaging and testing. It's extremely complicated, extremely fragile, and extremely dependent on geopolitical stability that honestly isn't guaranteed right now. Terraab is designed to collapse all of that into one building. design, fabrication, testing, packaging, masking, all under one roof, running in tight iterative loops, so the feedback between we made a chip and let's improve it is as short as possible. Musk said flat out that to his knowledge, nothing like this exists anywhere in the world. That's either genuinely true or it's grim way to great marketing. Based on what I found digging into this, it's actually mostly true. There are two chips coming out of Terrafab. You need to know about both. The H. First is the AI5. That one powers Tesla's self-driving system, the robo taxi network, and the Optimus humanoid robots. So, every time one of those robots folds laundry or drives you somewhere without a human at the wheel, there's an AI5 inside making it happen. That accounts for about 20% of Terraab's production. The other 80% goes to something called the D3. And this is where things get genuinely [music] wild. The D3 chip is purpose-built for orbital AI satellites, as in chips designed to run in space, handle the radiation, deal with the thermal constraints of the vacuum, and power an artificial intelligence network deployed above the Earth. SpaceX actually filed with the FCC earlier this year to launch 1 million satellites to create what amounts to an orbital data center. A million satellites running AI in space. And Musk's reasoning here is actually pretty solid if you think about it from basic physics. Solar irradiance in orbit is roughly five times stronger than what hits the surface through our atmosphere. And in vacuum, you radiate heat directly into space. No fans, no cooling infrastructure, no land footprint. His projection is that orbital AI compute could undercut groundbased alternatives on cost within 2 to 3 years. He closed the presentation with animations of satellites being launched from a mass driver on the moon. And he said, and I'm quoting here, "I want us to live long enough to see the mass driver on the moon because that's going to be incredibly epic." Which, yeah, same. This is the part that gets undersold in every headline I've read about Terraab because it's easy to frame this as Musk's ego going supernova again, but the supply chain situation is actually real and it's actually urgent. His argument is blunt. Build Terapab or don't have the chips. Not we'd prefer to have our own fab. more like the existing system [music] cannot physically supply what we need at the scale we're heading toward. He told investors back at Tesla's Q4 2025 earnings call that external capacity from TSMC, Samsung, and Micron would hit a hard ceiling within 3 to 4 years. [music] And the number he throws out is alarming. Today's total global chip output represents roughly 3% of what his companies will need in the future. 3%. Even if that's off by a factor of two or three, you still have a catastrophic gap between supply and demand that nobody else is really building toward. The companies that close that gap own the infrastructure of the next era. That's just how it works historically. You control the pipes. You control what flows through them. I have to be straight with you here because I'm not running a hype channel. The skeptics have legitimate points. TSMC spent $165 billion and years of work building six fabs in Arizona. And those won't even hit two nanometer production until 2029. A single advanced fab with serious production capacity costs around $28 billion and takes over three years just to construct in the US Tesla has never made a semiconductor in its life. And the financials are fuzzy. Tesla's CFO confirmed that the 20 to2 billion cost [music] estimate isn't even in the 2026 capital expenditure plan yet. A plan that already exceeds $20 billion on its own. There's no groundbreaking date, no construction timeline, no date for First Silicon. That's a problem. That's a real honest problem. And Musk's track record on timelines is what it is. Full self-driving took years longer than he said. The Cybertruck came in overweight and overbudget. The Hyperloop just kind of never happened. These aren't rumors, they're documented. But here's what I keep coming back to. The SpaceX exception. Rockets were considered just as absurdly complex and capital-intensive as chip fabrication. Governments with decades of institutional knowledge said private companies couldn't do it. SpaceX came from nowhere and eventually outpaced them all. That is a real precedent. You can't just wave it away because it's inconvenient for the skeptic narrative. What I find most genuinely compelling about Terraab isn't any single piece of it. It's that it actually connects everything. the XAI SpaceX merger earlier this year. Tesla's push toward 100 gawatts of domestic solar manufacturing, the Optimus robot program, Starship's development trajectory, Terraab slots into all of it as the central nervous system. [music] Chips made domestically, launched by SpaceX, powered by Tesla solar arrays, [music] running XAI models, ultimately extending compute infrastructure beyond Earth entirely. That's not just a company strategy. That's a blueprint for how a civilization becomes multilanetary and eventually maybe something more. Musk used the phrase galactic civilization. And yeah, it sounds like something off a science fiction paperback cover. But when you trace the actual logical chain from today's compute bottleneck to orbital solar to lunar infrastructure, the direction actually makes sense, even if the timelines are aggressive. I've been watching this space long enough to know that the wrong question is will this happen on schedule. It won't. It almost never does with projects this size. [music] The right question is whether the direction matters. And on that I think yes. Unambiguously yes. The world genuinely needs a step change in compute infrastructure. The current model of concentrating advanced chip production in Taiwan and hoping nothing destabilizes that region is fragile in ways that keep serious analysts up at night. Moving compute to space powered by solar built domestically. That's not just ambitious, it's arguably the only path that makes sense at civilization scale. If Terraab delivers even 60% of what's been promised, it reshapes the global technology landscape in ways that are hard to fully picture right now. And I'm rooting for it genuinely. Not because Musk is infallible. He's not. But because the problem being solved is real. The direction right. And frankly, nobody else is even trying to think at this scale. We're watching someone attempt something that's never been attempted before. Not just a fab, not just a satellite network. The whole stack from chip to robot to rocket to orbit to eventually the