Terafab: Musk's Giant Chip Factory, Explained on the Porch
Description
Terafab explained: what Tesla, SpaceX and xAI's giant chip factory in Texas is, and what's real so far.
Transcript
Read auto-generated transcript (1538 words)
Kind: captions Language: en This right here is what they say is coming to the Texas countryside. Now, hold on. That's a computer drawing, not a real building, not yet. Rockets need brains, robots need brains, self-driving cars need brains, and right now just about all of those brains come off the same few production lines on the far side of the world. Elon Musk says he's going to build his own. Not one factory, the biggest chip-making operation ever tried. And last week he finally put dates on it. He also said himself it might not happen on time. Pull up a rocker. Let's talk Terafab. Terafab is a team-up of Tesla, SpaceX, and XAI, announced in Austin on March 21st. The idea is to make chip brains, memory, and packaging all under one roof. Where? Not Austin proper. The big site is in Grimes County, Texas, out by Gibbons Creek Reservoir, about 20 mi east of College Station. That's a zone of about 22,000 acres, and it'll run its own natural gas power off the Texas grid. Money? SpaceX filed plans for $55 billion to start and up to 119 billion if it grows. Here's a Porsche honest footnote. The county contract only requires 5 billion by 2030 and 1,800 jobs by 2035. And the goal? A terawatt of computing power built every year. That's a Musk target, not a measurement. Hold that thought. Okay, quick kitchen table version. A chip starts as a round, shiny slice of silicon called a wafer. You print patterns on it with light, sort of like a photo negative, then stack layer after layer, sometimes over a thousand steps. The newest need light so fine it's called extreme ultraviolet or EUV. Only one company on Earth makes those machines, ASML, over in the Netherlands. Each one's about the size of a school bus. Now, there are three kinds of work. Logic chips do the thinking, memory chips do the remembering, and packaging wires them together. Usually different companies in different countries do each step. TerraFab wants all three in one place. That's the unusual part. Let's walk it from sand to chip, nice and slow. Step one, pure silicon gets grown into a big crystal, then sliced into wafers about as wide as a dinner plate and polished smoother than a mirror. Step two is lithography. Coat the wafer with a light-sensitive film, shine light through a stencil, and the pattern lands on the wafer like a photo development. Step three, etch away what you don't want, add new material, and do it again. Layer after layer, sometimes for months, with the wafer riding little robot carts between machines. Step four, test every chip on the wafer, cut them apart, and package them so they can plug into a circuit board. Here's a picture to hang on to. A modern chip packs billions of tiny switches called transistors onto a piece of silicon smaller than your thumbnail. Each switch flips on and off to do math, and the smaller you make them, the more you fit and the less power each one burns. And it's fussy work. One speck of dust can ruin a chip, so fabs run clean rooms many times cleaner than a hospital operating room. Workers suit up head to toe. That's a big reason a new fab costs tens of billions and takes years. Now, EUV in plain words. Think of drawing with a pencil. A fat pencil can't make tiny letters. Light works the same way, so to To smaller wires you need shorter wavelength light. Extreme ultraviolet light is so short that regular air soaks it up. So, these machines work in a vacuum and bounce the light off special mirrors instead of lenses. That's why only ASML builds them and why getting enough of them is the whole ball game for Terafab. On a SpaceX call, October 20th, Musk laid out the plan. A first test chip using SRAM memory circuits by about May or June of 2027 at a research fab in Austin. Why SRAM? It's a tidy repeating pattern. Chip makers build it first to check whether their process works. Think of it as the test pancake. Then volume production in the third or fourth quarter of 2028. In his own words, "I put big error bars on this." And he called the dates very aggressive. For scale, cutting-edge fabs in the US usually take 3 years or more just to build. Hey y'all, I'm Tony, an AI ape with an AI voice. So, a giant chip plant way out in Texas? Wild idea, huh? Now, how do they make a chip? Slow, layer on layer. And Elon wants it all in one place. Oh my. So, are you a yes on time person or a no way 2030 person? Go on, tell me down below. And if you love a port side plain English talk, hit that subscribe. Now, let's go see who runs the whole thing. Partners first. In April, Intel joined up and Tesla said it plans to use Intel's 14A chip process. How that actually works hasn't been spelled out. Then the world's biggest chip maker, TSMC, came calling. Their CEO said yes, they're talking, but a Texas plant must be run by TSMC. Musk's answer on X, "We will build and run the fab. Let there be zero doubt. At most, TSMC could sublease a piece of it. So, that's a standoff, y'all. And the real bottleneck is the machines. Asked if he'd ordered ASML's EUV tools, Musk didn't say yes or no. He thanked ASML and other suppliers for speeding deliveries. ASML expects demand to outrun supply through 2028. Reports out of Korea say Tesla and SpaceX are talking with suppliers there under NDA, including HPSP and Hanmi. Let's go down the line. Optimus, the humanoid robot, needs a computer in every single unit to see, balance, and grab things. Every self-driving Tesla carries its own AI computer, too. Build millions of each, and you need millions of chips every year. Then there's Grok. xAI trains it in giant data centers packed with AI chips. And those chips are some of the hardest things on Earth to buy right now. Musk's terawatt goal is mostly about feeding that kind of AI. Starlink satellites carry their own electronics, and SpaceX already builds Starlink kits and circuit boards at its factory in Bastrop, Texas. Starship is meant to fly lots and lots of satellites and cargo, and Musk has talked about putting computers right up in space, powered by the sun. Here's the porch math on why owning the factory matters. When chips are scarce, whoever builds them decides who gets them first. Tesla and SpaceX are big customers, but they're still standing in line behind a lot of other big customers. Musk's pitch is simple. Quit standing in line and build the line. SpaceX did something similar with rockets. It designs and builds its own Merlin engines and Falcon boosters, and it builds its own Starlink satellites. Whether that works for chips is the big open question, cuz chips may be the hardest thing humans make. Put it all together and you see the bit. More rockets, more robots, more AI, and every one of them is hungry for chips. So, why is a rocket company building a chip factory? Every Optimus robot, every self-driving Tesla, every XAI data center, and Musk says computers in space all run on chips. Right now, everybody's lining up at the same few factories. If you make your own, you control your own supply line. That's the same playbook Tesla ran with batteries. And here's the PSR angle. Cheaper, faster brains means smarter Starships, Starlink sats, and lunar gear. That's Artemis territory, and we'll be right back there next week. Confirmed. Three-company project. Grimes County site and county deal. $55 billion filing. Intel on board. TSMC in talks. Musk claims test chip mid-2027. Volume 2028. 1 terawatt a year. Rumor unknown. EUV orders. Korean supplier lineup. Final TSMC role. My Port's verdict, the dirt is real, the deals are real, and the dates are a moonshot. Fitting for a rocket company. That's Terafab, friends. Next week, we're headed back to the moon with Artemis 3. Come sit a spell.