Why Terafab Will Transform Full Self-Driving

Kind: captions Language: en All right, so this is the post that started it. March 14th, 9:41 in the morning. Elon Musk posts six words on X. Terrafab project launches in 7 days. No context behind it, no press release. Almost 78 million views by the time I'm recording this. Now, if you follow Tesla, you've heard Terraab since January. But I want to explain what it actually is, what it's for, and why it specifically matters for full self-driving. Almost all of the coverage thus far has been finance and semiconductors. The FSD angle has been largely ignored. That's what this video is about. March 21st is less than one week away. Let's get right into it. To understand March 21st, you have to go back about 9 months. Tesla's 2025 annual shareholder meeting. Musk stood up and said, and I'll quote him directly here, I've got chips on the brain. I dream about chips, literally, is what he said. At that same meeting, he went a little bit further. He said, Tesla may have to build a terapab. His exact words were, it's like giga, but way bigger. I can't see any other way to get to the volume of chips we're looking for. Now, Nvidia sells roughly 30 million chips a year. Apple ships around 300 million. Tesla needs two AI chips per vehicle and about two chips per Optimus robot. That's about 3 and a half to 7 million chips in 2026 alone. But that's at the current scale. The moment Optimus hits Musk's projections, those numbers become something no supplier has ever seen before. Then January 28th of this year, the earnings call. Musk told investors the following. when he looks 3 to four years out, he sees chip production as the bottleneck. His exact quote was, "It would be crazy to not try the tariff app." He said there would be a bigger announcement later. And well, March 21st is that announcement. Akos Gupta put together probably the most useful public breakdown that I've been able to find. And he draws a comparison worth spending some time on here, the Gigafactory. In 2014, Musk announced a $5 billion battery factory located in the Nevada desert. Tesla at the time was selling 35,000 cars a year. Analysts were calling it delusional. That factory now produces more lithium ion cells than every other American manufacturer combined. The Gigafactory was never about batteries. It was about removing the constraint that would have killed Tesla at scale. Terrafab is the same bet. It's just one level even deeper in the stack. So what is Terraab structurally? It's a semiconductor fabrication facility. The kind that physically manufactures silicon chips. Right now every chip Tesla uses to run FSD to train the models behind it to power CyberCab and Optimus. Every one of those chips is built by someone else. Primarily TSMC in Taiwan and Samsung in South Korea. Terafab is Tesla's move to change that. Estimated cost is around $25 billion. The target output that they're projecting is about 100 to 200 billion chips per year with processing technology of about 2 nanmters. The most advanced node currently in commercial production. Now, the first product Terrafab is designed to produce is the AI5. Small batch production is geared up for 2026 with high volume projected in 2027. The March 14th announcement lines up almost exactly with when construction needs to start for that 2027 target to hold true. And it doesn't stop at AI5. AI5 is going to be 2026 to 2027. AI6 is going to be in 2028. Then you have AI7 and AI8 following on an annual cadence. Terrafab isn't being built for one chip generation. It's going to be basically the infrastructure for their entire road map. Now, one thing I want to be precise about here is that the Terrafab launch on March 21st, that doesn't mean that a fab opens its doors next Saturday. These facilities take multiple years to actually build. So, what March 21st almost certainly represents here is either breaking ground, maybe a location announcement, potentially a partnership reveal, or a construction timeline of some sort. So, the clock basically is going to start on March 21st. The FAB itself is years away from its first production. Now, here's the part that almost nobody is talking about. FSD doesn't improve just because engineers write better code. It improves because the models get retrained on more data. And retraining requires an enormous amount of computing power. Tesla's Cortex Supercomputer in Austin started with 50,000 Nvidia H100 GPUs. They're planning to expand that by three times because the compute demand to larger models just simply keeps growing. Now, every major step forward in FSD capability has a hardware component underneath it. The hardware 3 to hardware 4 transition wasn't just a chip swap. It's what made certain FSD features feasible at all. Therefore, when Tesla says it needs 100 to 200 billion chips a year and can't guarantee that through their external partners, that is directly a statement about FSD's growth ceiling. Now, if you can't get the chips and you can't train the models fast enough, you can't put updated hardware in enough cars fast enough. Cyber Cab is already off the production line. That happened in February. But the Cybercap that just rolled off that line is running AI4, not AI5 yet. AI5 won't be available in volume until 2027. So the first wave of Cyber Cabs run on the current chipset. The fleet that runs on AI5 with roughly 50 times the performance of AI4. That's going to be about 2027 and beyond hopefully, fingers crossed. And that story depends directly on whether the Terrafab is on schedule. The AI5 specs are worth putting on screen here. Roughly 50 times the overall performance improvement over AI4, 10 times the raw compute power, 9 times the memory capacity, and five times the memory bandwidth. When you put those numbers on the chip that powers FSD, you're not talking about an incremental improvement here. you're talking about a different category of what's possible for the software running on it. One more angle worth raising here is that Phil argues Terrafab is likely to be a joint Tesla and SpaceX project. SpaceX needs chips for space-based data centers, and the capital requirements of a facility this size makes a shared project structurally logical. Whether that's confirmed on March 21st, we'll have to wait and see, but it tells you the scope. This isn't just a chip fab for cars. its infrastructure for an entire ecosystem. The chip demand problem also gets clearer when you add Optimus. Tesla confirmed the pilot production line is already running at Fremont. The target is about 50,000 robots in 2026. Then they're going to 10x that in 2027. At the 2025 shareholder meeting, Musk said Fremont will eventually produce a million Optimus units per year. Gigafactory Texas would scale that to 10 million. He called it the fastest production ramp ever attempted. Now run the chip math on this. Two AI chips per robot, a million robots per year. That's about 2 million chips from Optimus alone. Then you add the vehicles. Then you add the cybercap fleet. Add Dojo's training infrastructure. And the number you arrive at is something no external supplier can currently guarantee. That is the supply problem Terraab is designed to solve for. Tesla actually pulled the Optimus Gen 3 setup timeline forward to this summer ahead of schedule. Long-term Fremont goal is going to be 1 million units per year. Every quarter of chip supply delay is a quarter of delay in Optimus production at scale. That's the direct connection between Terapab and Optimus. Now, the most credible push back on Terrafab comes from Jensen Huang, CEO of Nvidia. His position is that creating a modern semiconductor fab requires deep mastery of engineering and manufacturing science. The kind of precision that TSMC has spent 30 years developing. It's not something that you stand up in a few years. The yields alone, the percentage of chips on a wafer that actually work a discipline that takes decades to optimize. Tesla has never run a semiconductor fab. And that's a very legitimate point. Huang isn't speculating here. He's describing an industry he operates inside of every single day. Now, Samsung had to step in on AI5 production because TSMC's capacity was already strained. Musk is personally attending working level technical meetings with Samsung engineers. That level of involvement tells you how constrained the supply situation already is. And AI5 isn't even in the volume production as of right now. But Tesla isn't trying to become a merchant foundry here. It's not competing with TSMC for Apple's business. The goal is a little bit more narrow than that. It's to produce the chips Tesla's own systems need at the volumes Tesla's own roadmap requires on a timeline external partners just simply cannot commit to right now. The success threshold for Terrafab isn't matching TSMC's yield rates in one year. it's producing enough chips to meaningfully reduce external dependencies. Now, that's a different type of problem and has a much lower bar. Therefore, the question heading into March 21st isn't whether Tesla can become TSMC. It's whether they can build a facility that solves a specific supply problem on a specific timeline. Those are very different questions, and one of them is considerably more answerable. There's going to be three specific things I'm going to be watching for on March 21st. The first thing is going to be location. Where is this being built? Texas is the obvious call, but Chips Act funding implications could point to other states. The site will tell you how far along the planning actually is. The second thing I'm going to be looking at is Intel. Tesla is reportedly planning to use Intel's EMIB packaging for Dojo 3 and the two companies are in discussions about an unprecedented Samsung Intel collaboration led by Tesla. If any of that gets formalized on March 21st, it changes the risk profile significantly. It's going to mean that Tesla isn't building from scratch anymore. They're going to be plugging into Intel's existing footprint. The third thing I will be looking out for is the actual construction timeline. When are they going to break ground? When is the first wafer expected from this specific facility? The answers to these questions are going to tell me specifically whether the AI5 volume target for 2027 is coming from Terraab or still coming from TSMC and Samsung. Every structural advantage Tesla has built in batteries, motors, software came from the same exact decision. Own the critical component and don't depend on it from external suppliers. Chip supply is now that critical component for everything Tesla is building next. The Gigafactory bet looked irrational in 2014 and terapab is a harder problem, but the logic is very very similar. Now whether Tesla can execute on semiconductor manufacturing at a scale is genuinely uncertain. I don't think anybody actually knows but the question of whether they should try I think that one is already answered. It has to be done. So March 21st again is less than one week away. I'm asking for your support. If you haven't yet subscribed to my channel, hit that subscribe button down below so that you can keep up to date with the March 21st news and Tesla and FS FSD news in general. I'd really appreciate all of the support that you guys are showing me on my channel. While you're down there subscribing, if you can hit the like button for me as well, that would be really, really awesome. I'll see you guys on March 21st. Let me know what you think we might see during that announcement. Thank you.