The $17 Billion Chip Factory That Terrifies TSMC

Kind: captions Language: en northwest of Houston. Engineering News Record reported that recent drone footage already shows Earth movers working the site. Musk has described the finished factory as the largest and most valuable building on Earth by far. For now, that's a promise from the man selling it. And the full plan is a lot bigger than the headline number. Filings with the state of Texas lay out a buildout running through 2036, adding up to about $119 billion. That's close to two full years of TSMC's spending at this year's pace poured into one site in rural Texas. Phase 1 is meant to start commercial production in 2028, and every phase after that only happens if Tesla and SpaceX decide to keep going. The idea itself started almost as an aside. At Tesla's shareholder meeting last November, Musk told investors the company might have to build a terrafab, which he described as like Giga but way bigger. By March, he was launching it on a live stream from the old Seahome power plant in downtown Austin as a joint project between Tesla and SpaceX. SpaceX had absorbed Musk's artificial intelligence company XAI in February. So, the rocket company now carries the AI ambitions, too. The plan puts logic chips, memory, [music] packaging, and testing all under one roof in Texas. Two kinds of chips come first. One is an inference chip for Tesla's cars and its Optimus humanoid robots. The other is a hardened chip built to survive in space for the data centers SpaceX wants to put in orbit. So why would a car company and a rocket company try to make their own chips in an industry where even Intel has struggled to keep up? Musk gave the answer himself on Tesla's earnings call in January. He said the company would have to build terraab to remove the probable constraint in 3 or 4 years. At the March launch, he put it even more simply. [music] We either build the terapab or we don't have the chips. And we need the chips. So, we build the Terrafab. Right now, Tesla's chips come from someone else's factories. Musk has said TSMC will make the AI5 chip, first in Taiwan, and then in Arizona. Samsung signed a contract to build the AI6 at its [music] new FAB in Taylor, Texas. On Tesla's call in July, Musk said the AI5 is hopefully in volume production around the middle of next year. That chip is meant to run the cars and the robots. So, every month it slips is a month Tesla's biggest bets wait on someone else's schedule. Until [music] then, Tesla is waiting in the same line as Apple, Nvidia, and buying capacity from the same handful of factories. And one of [music] those factories owns the line. Trendforce found that TSMC took a record 72.5% of all contract chipmaking revenue in the second quarter of this year. That's nearly three of every $4 that chip designers around the world paid someone else to build their chips. Samsung in second place took in less than a 12th of what TSMC did. TSMC's real edge isn't the machines because anyone with enough money can order the same tools from ASML, Applied Materials, and Tokyo Electron. Its edge is yield. The share of chips on each wafer that actually work when they come off the line. Decades of running those machines, fixing their problems, and serving customers in every corner [music] of the industry have turned that into the hardest thing in the industry to copy. Musk can't buy his way to the front of TSMC's queue and he can't download TSMC's experience, which is why his boldest idea for Terrafab is about the air inside the building. In a podcast interview with Peter Diamandis released in January, Musk went after one of the most expensive habits in chipm. I think they're getting clean rooms wrong, by the way, in these modern fabs, he said. Then he made a bet. I'm going to make a bet here that Tesla will have a 2 nanometer fab and I can eat a cheeseburger and smoke a cigar in the fab. 2 nanome is the same class of process TSMC only started mass-roducing [music] at the end of last year after years of development. So Musk is promising the most advanced kind of chip factory there is with the kind of air you'd find in a burger joint. His reasoning comes down to one line from the same interview. The wafers, [music] he said, should just maintain wafer isolation the entire time. To see why that matters, picture the inside of a normal advanced fab. [music] The entire production floor is a clean room. Filtered air pours down from the ceiling and out through the floor without stopping. And every worker on that floor wears a full bodysuit, a hood, and a mask. All of that exists to keep dust, skin flakes, and droplets away from the wafers. Musk's idea turns the logic around. If the wafer never touches room air, then the room itself can be ordinary, and you stop paying to keep an entire [music] factory floor that clean. The reaction from the industry press was not kind. Tom's Hardware pointed out that eating and smoking are banned in the cleanest clean rooms and that stray particles could end up on the mirrors inside an EUV machine. Its writer went further and suggested the remarks showed a limited understanding of how leading edge fabs actually [music] work inside a real fab. Though, Musk's idea looks a lot less strange because the industry already does most of it. Walk through any modern fab that processes the big 300 mm wafers and you won't see wafers sitting out in the open. They travel between machines inside sealed plastic pods called FOPs, short for front opening unified pods. Think of a sealed lunchbox that only opens once it's locked onto the machine. Each machine also has its own mini environment, a small enclosure with its own filtered air built around the spot where the pod opens. SEMI, the industry's standards body, first published the specification for these pods in 1997. That's a full decade before the first iPhone went on sale. And the industry asked Musk's question almost as soon as the pods arrived. Back in the late 90s, solidstate technology reported that sealed pods could let fabs run with a lower grade clean room around them, which could cut costs. The same report carried the warning that came with it. Those savings disappear the moment yields suffer. So, the cheeseburger bet is an old idea pushed further than anyone has dared to push it. Musk wants to keep the sealed boxes and drop most of the ultra clean room that surrounds them in today's fabs. The reason nobody has gone that far is simple. A wafer is not sealed every second of its life. It leaves the pod every time a machine processes it. And every machine has to be opened up by engineers for maintenance, sometimes for hours with its insides exposed to whatever is floating in the room. Those are the moments the clean room was built for. Musk's bet is that Terraab can make those moments clean enough on their own. A few thousand wafers a month. That's how Musk described the first real test [music] of the idea on Tesla's earnings call in April before the giant Grimes County site. Tesla is building a research fab at Giga Texas in Austin, which he called probably a $3 billionish initiative. Techbot put TSMC's total capacity at [music] about 1.4 million wafers a month. That means TSMC can [music] start more wafers in a single day than Musk's research fab is designed to handle in a month. Nobody expects the Austin Fab to compete on volume, though. Its job is to prove the process works. [music] And in July, Musk said Tesla had already placed equipment orders for it. That split matters. The small Austin Fab is where the cheeseburger theory meets real wafers, real particles, and real maintenance crews. The giant site in Grimes County is where the theory has to survive at full scale with thousands of workers and machines running around the clock. If the idea fails in Austin, [music] Tesla finds out before the big money is spent. The process itself won't be invented from scratch. In April, Intel announced it would bring its design, manufacturing, and packaging know-how to Terraab. Later that month, Musk said on Tesla's call, "We plan to use Intel's 14A process, which is state-of-the-art and in fact not yet totally complete." That's an honest admission and a revealing one. Intel's own road map points to high volume production of 14A in 2028, the same year Terrafab's first phase is supposed to start producing. 14A is also the node where Intel plans to use ASML's newest high NA lithography machines, which is a story for its own video. So, line up the contenders. TSMC has had its 2n process in volume production since late last year, and it expects a steep ramp through the rest of this year. Samsung is still bringing its Taylor Fab in Texas online with Tesla's AI6 as its anchor customer. Intel is finishing a process it has never run at volume with Terraab as the biggest named user of it so far. And Terraab has money, a guaranteed buyer in Tesla and SpaceX and a clean room theory nobody has proven at the leading edge on process maturity. TSMC is ahead by years. But wafer counts and process names are not the scoreboard that decides this fight. The scoreboard is cost per working chip. Tom's Hardware reported an estimate from the consulting firm International Business Strategies that a 2nanmter fab with 50,000 wafer starts a month costs about $28 billion. That's roughly 2/3 more than Terapab's entire first phase for a single fab. Some of that bill goes into the clean room, the air handling, and the giant support floors beneath it. That's the slice Musk wants to shrink. But a Fab's construction cost is only half the equation. The other half is how many chips on each wafer actually work. Think of a bakery that buys a cheaper oven. It saves money right up until the trays start coming out burnt. And then every burnt tray costs more than the oven ever saved. Chips work the same way. A single particle landing in the wrong place at the wrong moment can [music] kill a chip. And AI chips are some of the biggest slabs of silicon anyone makes, which means more surface for a particle to land on. So, a cheaper building only wins if the chips coming out of it work as often as TSMC's do. Clean room technology, a trade publication that covers this exact field, [music] made the point plainly. Sealed pods and mini environments do not replace clean rooms. [music] And leading edge wafers still need clean room level protection, at least around the spots where they're exposed. That leaves Musk two ways to win. One is automation. If robots handle every open tool moment and humans almost never stand next to an exposed machine, then the biggest source of particles, the people, mostly leaves the equation. Tesla builds its own factory automation. [music] And that's the part of this bet where it actually has experience. The other is design. Every enclosure around every machine would have to stay sealed even during repairs, which nobody has done across an entire leading edge fab. Musk hasn't publicly explained how Terrafab handles maintenance. Until he does, that's [music] the biggest open question in the whole plan. Now, flip it around. Say the bet works and Terapab makes chips with good yields in a building that costs a lot less than a normal fab. Three things change at once. The first is capital cost. Every fab Tesla and SpaceX build gets cheaper, which matters for a company whose finance chief told investors that Tesla's spending this year will be more than $25 billion. The second is speed. A simpler building goes up faster. And in a market where AI chip demand keeps outrunning supply, being able to add a fab in less time is worth more than almost anything, whoever adds capacity first ends up setting the pace for everyone downstream, from car makers to cloud companies. The third is control. Tesla and SpaceX would stop waiting in TSMC's queue for their car chips, robot chips, and orbital data center chips. They'd be building on their own timeline without competing against phone makers and cloud giants for the same production slots. Now look at the size of what Musk says he wants. He's talked about a terowatt [music] of computing power a year, most of it eventually in space. Reuters reported that analysts at Bernstein estimated [music] a goal like that would take somewhere between 5 and 13 trillion in capital spending. Even the low end equals about 80 years of TSMC's [music] spending at this year's pace. And that's the connection that makes the cheeseburger bet more than a joke. At that scale, no company can build fabs the way TSMC does and ever reach the target. The only way Musk's numbers come even close is if each fab gets dramatically cheaper and faster to build. So, the clean room idea sits at the center of the whole plan because it's the only version where the math could ever add up. If you want the honest read on who can actually make the chips the world runs on, subscribe to Tomorrow Engineered because every video here answers that question. The bet does come with a catch, though, and it starts with the machines. No amount of money makes ASML build its machines any faster. The consultancy Yol Group puts ASML's standard EUV machine at roughly $180 million each. Terrafab's entire firstphase budget would buy fewer than a hundred of them before a single wall goes up or a single etching or deposition tool gets ordered. And every other chip maker is in the same queue for those machines, including [music] TSMC. Once they arrive, each one has to be installed, calibrated, and qualified before a single production wafer goes through. And that takes months, even for teams that have done it many times. The second catch is the process. Intel's 14A isn't finished by Musk's own admission. And even a finished process doesn't arrive ready to go in a new building. Every fab has to tune its own machines, its own chemicals, and its own air. And the first wafers through any new line almost never come out at full yield. On Musk's version of a fab, the engineers would be tuning a process and testing a brand new clean room philosophy at the same time. If Intel slips, Terrafab [music] slips with it because the factory can't run without Intel's recipe. The third catch is utilization. A fab only makes money when its machines run day and night. TSMC keeps its lines full by serving customers across phones, PCs, cars, and data centers. So, when one market slows down, another usually picks up the slack. Terrafab is built mostly around demand from Tesla and SpaceX. If Optimus robots ship slower than planned [music] or orbital data centers stay on the drawing board, those machines sit idle while the costs keep coming. The fourth catch is on paper. When SpaceX filed to go public in May, [music] Electrek reported that the filing describes Terapab as a general framework for the future development of Terraab. The same filing says neither Tesla nor Intel is [music] obligated to remain part of the project and the financial terms and intellectual property rights still aren't settled. In September, SpaceX went to court against the Texas Attorney General and Grimes [music] County to keep Terapab records from being released to the public. Companies protect trade secrets all the time. So that alone proves nothing. But it does mean the outside world is judging a factory it mostly can't see, which is exactly why the headline [music] numbers deserve some caution. And the last catch is the one money can't fix at all. TSMC's real asset is its engineers. People who have spent whole careers learning how to squeeze yield out of each new process. Terrafab has to hire or train that [music] experience from scratch in rural Texas while Samsung, Intel, and TSMC's own Arizona fabs compete for the same people. Terraab can still work, but the cheeseburger bet has to win on yield, on equipment, on process, and on staffing all at once. What happened this month shows how seriously the industry is taking it. Earlier this month, word got out that TSMC was exploring ways to help run Terraab. Musk first called it just discussions, then shut the door within days, posting on X, "No, we will build and run the FAB. Let there be zero doubt about that." He added that maybe TSMC could subleasase part of Terapab if it wanted, but nothing more than that. Bloomberg reported the same day that Intel chief executive Lip Bhutan said Intel would stay involved. Think about what that episode says. The most dominant chipmaker in the world looked at a factory that might one day compete with it and its first move was to see whether it could run it. TSMC tends to protect its position by staying close to its biggest [music] customers wherever they build. In July, Reuters reported that TSMC added another hundred billion dollars to its Arizona plans, bringing its total commitment there to about 26 >> [music] [music] >> Heat. Heat. [music] >> [music] [music] >> Oh, [music] heat. Heat. [music] [music] Heat. [music] Heat. [music] Heat. [music] >> [music] >> Heat. Heat. [music] Heat. [music] [music] Heat. >> [music]