This New Chip Factory Should Terrify TSMC

Kind: captions Language: en Something colossal is unfolding in the Texas Plains. Elon Musk is building a 1.4 nanometer chip factory. More than 100 million square feet, [music] enough to fit roughly 30 TSMC Arizona fabs. And somehow that's not even the wildest part. It could cost up to 100 billion dollars, [music] generate its own power, and potentially even have a particle accelerator inside. I've spent [music] over a decade designing the world's most advanced chips and now building my deep tech startup. And this might be the most fabulous project I've ever seen. But now comes the real question. Can they actually build it? Why might a chip factory need a particle accelerator? And most importantly, what happens if it actually works? Subscribe to the channel and let me explain. To truly understand Terafab, first you need to understand the ridiculous journey an AI chip makes today. Let's take this Nvidia GPU. Nvidia designs the analog circuits and the digital [music] part, but it doesn't manufacture the silicon itself. And that's where the journey begins. The logic dies are fabricated by TSMC in Arizona on silicon wafers coming from Japan. Meanwhile, across the Pacific in South Korea, SK Hynix is building the HBM memory. Then those components travel again to TSMC in Taiwan where advanced COWOS packaging brings them all together. But we are still not done because it still isn't something we can plug in a data center. Then comes [music] testing, qualification, and final assembly happening somewhere else again. So this single Nvidia chip depends on a manufacturing chain stretching across multiple factories and continents. [music] And that's thousands of kilometers of supply chain for one accelerator. And every handle of creates another dependency. No foundry capacity, you wait. Not enough HBM, you wait. COWOS capacity is full? You wait. And it doesn't matter if 99% of your supply chain is ready. One missing piece delays everything. [music] And that is the fundamental idea behind the TeraFab. They exist entire map and start collapsing it into one place. Logic, memory, advanced packaging and testing, even parts of the infrastructure, turning an entire supply chain into a factory, which sounds like a solution until you actually ask, "How do you build it?" So, let's assume you have the land and you figure out how to build a hundred million square feet of factory and [clears throat] potentially you have a hundred ninety billion dollars to spend. And at that point you might think that the hardest part is over. Just buy machines [music] and start making chips. Well, not quite yet. Because TeraFab isn't just enormous, it's targeting roughly one terawatt of AI compute production per year, potentially around fifty times [music] today's global AI chip deployment. Eventually, TeraFab is targeting one million wafers per month and not simple chips, but advanced chips built on fourteen angstrom class gate-all-around transistors. And that's extraordinary volume at the leading edge. Now, a single TSMC fab can [music] produce in the order of forty gigawatts of compute capacity per year. [music] TeraFab is targeting one terawatt. So, very roughly we are talking about the equivalent capacity of around twenty-five leading edge TSMC [music] fabs under one roof. And each one needs fleets of the most sophisticated machines humanity have ever built. [music] Lithography machines, etch and deposition equipment, ion implantation, metrology, inspection, thousands of tools. Most of these tools can be bought, but at this scale one [music] becomes particularly problematic, EUV lithography. Inside an EUV machine, powerful lasers blast tiny droplets of molten tin tens of thousands of times per second, eventually creating a 13.5 nanometer EUV light used to print nanoscopic transistor features on the wafers. And the high-volume leading-edge fab can require 20 or more EUV scanners. And now scale that to what Terafab's ambition, and suddenly you are talking about hundreds of EUV machines, potentially over 300 for a single project. And here is a problem. Today, there is exactly one company supplying production-scale [music] EUV scanners, ASML. And in 2025, they shipped around 48 EUV systems. 48 for the entire world, including TSMC, Samsung, Intel, SK Hynix, everyone. So, at today's production rate, Terafab alone could need years of the world's EUV output. And this is where even tens of billions of dollars cannot solve the problem, because you [music] cannot buy what does not exist. Which leaves Terafab with two possibilities. ASML and its suppliers somehow radically scales production, or Terafab eventually needs another way to generate EUV light. And right now, there is something of a gold rush in lithography. There are over a dozen of companies and research programs in the US, China, Japan, Russia, and beyond are experimenting with new ways to generate the light to break the dependence on a limited supply of EUV machines. And at the scale of Terafab, Musk has already hinted at a much more radical solution, FEL [music] for the win, where FEL stands for free electron laser. And if you're wondering what a particle accelerator has to do with manufacturing chips, it's a good question. And this is where Terafab gets really strange. Here is how an FEL works. You accelerate electrons to almost the speed of light, then send them through a sequence of magnets that make them wiggle from side to side. As they move, they generate light that builds into an extremely powerful beam. Tune it to the right wavelength and you can produce the same 13.5 nanometer EUV light used to manufacture advanced chips. And what's cool is FELs aren't science fiction. We are actually building them. There is just one problem. They are enormous. The European XFEL in Germany is a gigantic facility stretching 3.4 km or over 2 miles underground with accelerator tunnels, vacuum systems, and enormous arrays of magnets. So, we are talking about a chip factory with a particle accelerator. And that's where it gets really interesting. Today, every EUV scanner has its own light source, but a sufficiently powerful FEL creates another possibility. Instead of every lithography machine carrying its own artificial sun, the factory builds one giant sun and shares it. One FEL could potentially generate enough EUV power to feed many lithography tools. And suddenly, Terafab's enormous size isn't simply the consequence of scaling the production. It could actually make an entirely new chip factory [music] architecture possible. For a normal fab, the FEL economics makes little sense. At Terafab scale, they might. [music] So, have you already solved the EUV problem? Not quite. We will [music] get back to that in a moment. In 2025, over half of all companies are already using AI. Research shows that 40% of people worry AI will replace them at their job. But in reality, people using AI will replace those who don't. And it's already happening right now. Microsoft, Google, Amazon are currently hiring people that understand AI, who know how to build with AI. If you build a startup like me, or create anything, or a working professional, AI isn't a threat to you. It's a leverage. [music] Because if used right, it can help you to save hours of time and thousands of dollars in cost. And you need to learn how to use it. Now, this is why I highly recommend joining this two-day AI training by AltSchool, which takes you from beginner to advanced AI professional in just 16 hours. It's valued at $895, but I partnered with AltSchool to provide 1,000 free seats for you. And the feedback from you guys on this training has been amazing so far. This two-day program offers 16 hours of live AI training spread across two days, happening on the coming weekend on this Saturday and Sunday, 10:00 a.m. to 7:00 p.m. [music] In this training, you will learn more than 20 AI tools, prompt engineering, developing AI agents, and more. And exclusively for my audience, you can join it for free. Register right now through the link below, or scan the QR code here. And thank you, AltSchool, for sponsoring this episode. Now, back to that particular accelerator. Even if Terrafab builds and fails, it could replace the EUV light source, but it doesn't replace the entire EUV scanner. You still need the mirrors and masks to project the chip pattern on the silicon, and ASML's incredibly precise wafer stages, alignment, and metrology to actually print it. So, the FEL replaces one critical piece of the machine, but that creates a new problem. Do you remember this idea of one giant sun feeding multiple EUV tools. And that's incredibly powerful until the sun goes down. If that one fell fails, dozens of lithography tools could stop with it. Terafab might remove one external bottleneck only to create a massive internal one. >> [music] >> And even if the fell works perfectly, the optics, stages, and everything around it still need to scale. Well, let's assume Terafab gets a machine and accelerator even works. So, problem solved? Not even close. Because now you have to power all of it along with thousands of other tools. And that brings us to perhaps the Terafab's biggest physical constraint, power. [music] And here Mas already has a useful experiment. When XAI built Colossus in Memphis, the grid couldn't provide anywhere near enough power [music] for the AI cluster they wanted. So, instead of waiting for infrastructure to catch up, [music] XAI brought the power with them. Temporary natural gas turbines went up while permanent grid connections were being built. And [music] this Colossus data center became a remarkable demonstration of how quickly XAI can move when infrastructure becomes the bottleneck. Terafab appears to take [music and clears throat] that lesson and design it in from the day one. The campus is planned with its own natural gas power generation and large battery systems. [music] Reported configurations involve more than 40 gas turbines at roughly 50 megawatts each. That's around 2 gigawatts of generation. And generating the electricity is only part of it. You still need batteries, substations, [music] transformers, and an internal electrical network capable of moving gigawatts of power around the campus. At this scale, Terafab isn't just building a chip factory. [music] It's building the power and water infrastructure that keeps the factory alive. But, let's assume they sold the machines and the electricity. And there is still one thing $190 billion can't buy overnight, decades of manufacturing experience. And that's where Intel becomes extremely interesting. Intel has been manufacturing chips for 58 years. And right now, Intel is ramping 18A process node at its new Fab 52 in Arizona, already manufacturing Panther Lake laptop processors on this technology. And reported yields have reached as high as around 85%. And this yield matters because designing an advanced transistor is one thing, manufacturing millions of them economically is another. But, Terafab target goes even further to Intel 14A and 14A process built on technologies Intel is learning to manufacture with 18A process today. First, Ribbon FET. 18A introduced Intel's gate-all-around transistor, where the gate wraps around the channel for much tighter control. 14A refines that architecture further, packing transistors closer together and getting up to roughly 30% higher density. In simple terms, that allows for more transistors in the same area, running faster at the same power consumption. But, the most interesting and revolutionary part here is power. 18A introduced power via. Normally, power and signals compete for the space in the wiring above the transistors. Power via separates them. Signals come on the top, power comes from the backside. >> [music] >> 14A process evolves that into power direct, bringing backside power even more directly to the transistor. [music] And that means shorter path, less resistance, and better power delivery. And then, there is a third innovation, turbo cells. And this is genuinely the most exciting part for me, because I used to work in timing closure in STA. And this is where exactly this timing and this kind of cells actually matter. Not every part of the chip needs to work at the maximum speed. Usually, only a few critical paths slow the whole chip down. So, for that part, Intel will use larger and stronger turbo cells only on those paths, allowing the next stage to switch faster without making the entire chip bigger and more power hungry. And I have to say, this is a kind of engineering which makes me a huge fan of Intel. And now, when they're trying to put all these innovations at the once, which [music] is extremely hard, I'm really rooting for them. It's manufacturing 18A today, while developing the technologies that could take Terafab toward 14A in the future. And that's exactly what Terafab needs. Yield learning, advanced packaging, and decades of manufacturing know-how. Terafab can't build the infrastructure. Intel can help shorten the learning curve inside it. But all of that leads to a much bigger question. What if Terafab actually works? Because the biggest impact it could make is not one terawatt of compute it will produce, but the new model that it proves. For decades, the semiconductor industry has moved toward extreme specialization. Size makes up chips, ASML builds lithography machines, TSMC manufactures logic, others package and test it. That's what it looks like in a nutshell. We created perhaps the most sophisticated manufacturing supply chain in the world. Terafab is making almost the opposite bet, integrating more of that under one roof, along with its own power generation, water treatment, and recycling infrastructure. All increasingly optimized around one goal, producing AI compute at scale and potentially Terafab could go even further, even questioning something as fundamental as does the clean room actually needs to be clean? And that is a fun idea. I explored that idea along with economics of Terafab and fundamentals in the first part of the video. So, if you want to have a full picture, it's worth checking out. But look at the bigger idea here. We are building particle accelerators to build AI accelerator. And AI accelerator isn't really just a GPU anymore. Its performance depends on the entire system, logic, memory, interconnect, packaging, networking, cooling, and power. And at Terafab scale, the optimization problem changes. You are no longer constrained only by transistor density, you are constrained by fabs, memory, packaging, water, [music] and ultimately, energy. Put it all together and Terafab becomes an extreme concentration of capital and manufacturing capacity [music] in one place. And that's also the paradox, becoming less dependent on the outside world and that creates a new risk because just one failure can affect much more of the system. Even a natural disaster can destroy everything. Although the inland Texas location reduces some of that risk. While TSMC and Intel spread manufacturing across networks of fabs and campuses, Terafab [music] is making a different bet, concentrate more to control more. But also more riding on one place. That could make it faster and more self-sufficient, but also means a single failure becomes way more expensive. And if this model works, the payoff is enormous because instead of scaling AI output at the speed of your slowest supplier, you now scale it at the speed of your own factory. And that's why the most interesting thing about Terafab isn't the scale, it's what Terafab is trying to turn a chip factory into. For 60 [music] years, we've competed over who can build the best transistor. Terafab is betting that the next competition could be much bigger. Who can turn energy, silicon, and infrastructure into the most AI compute? And that is to me is the real Terafab bet. If you enjoyed this video, you will love this one or watch this video on the race to build the light source alternative to EUV. Love you guys and I will see you there. Ciao.