Terafab: Inside Musk's $55 Billion Austin Chip Gamble

Kind: captions Language: en Picture a stage in downtown Austin. March 2026. An old power plant, the Seaholm, all brick and history. The lights come up and Elon Musk says this. >> We either build the Terafab or we don't have the chips. We need the chips, so we're going to build the Terafab. >> And just like that, he announces the largest semiconductor factory ever proposed by, in his words, an absurd margin. >> 100 million square feet. About 10 times the main building at Giga Texas. >> A so-called terawatt of compute every year. 2 nanometer chips, logic and memory, packaging, test, all of it under one roof. >> And the company building it has never made a single chip in its life. So, shall we? >> Welcome to the show. I'm George. >> I'm Adam. >> And I'm Bella. Today, Terafab, the most ambitious chip factory in American history, or the most expensive press release. Possibly both. >> So, let's start with what was actually announced because the numbers almost don't pass. Adam, walk us through it. >> Right. The headline price in Austin was 20 to 25 billion dollars. But then in May, SpaceX quietly files a property tax application out in Grimes County. That's rural Texas, about 90 miles northeast of Austin. And the filing describes a fab costing 55 billion dollars in its first phases. Up to 119 billion if every expansion gets built. >> Wait. So, which is it? 25 or 119? >> Both, sort of. 25 is the splashy stage number. 119 is what's on the county paperwork. >> And the structure is unusual, too. This is a joint venture. Tesla, SpaceX, and xAI. That's Musk's three companies pooling their need for chips. And then later, Intel signs on as a partner to help with the actual fab technology. >> Which, honestly, is the first sane thing in the whole announcement. Bringing in Intel, a company that's actually built fabs. >> It's also a tell. If you could do the process science yourself, you wouldn't need Intel. >> Let's talk about the scale, though, because this is where it gets science fiction. Musk said, and I quote, "We couldn't possibly fit the terrafab on the Gigafactory campus. It will be far bigger than everything else combined there." >> The targets are genuinely staggering. 100,000 wafer starts a month, scaling toward 1 million. At full tilt, that's roughly 70% of everything TSMC makes globally today, out of one building. >> 70% of the biggest chip maker on Earth, from a company that has made 0% of anything so far. >> And the chips themselves? >> 100 to 200 billion custom AI and memory chips a year. And here's the part everyone did a double take on. Musk said roughly 80% of that compute is meant to go into orbit, onto satellites. >> Space data centers, because of course, the man builds rockets. He wants the compute up where the rockets go. >> And it's not just enormous, it's meant to do everything in one place. Logic chips and memory chips under one roof. Then the packaging, the testing, even the photomasks. Vertically integrated, top to bottom. >> Which almost nobody does anymore. The whole modern industry is specialized. One company designs, another fabricates, a third packages and tests. Musk wants to fold all of it into 10 modular buildings on one site. >> It's the classic Musk move. Look at how everyone else does it, then ask, "What if one company just did the entire thing?" Sometimes that's genius. Sometimes it's how you spend 5 years rediscovering exactly why everyone else split it up. >> Why a fab at all though? Why not just keep buying chips like everyone else? >> Because he says he can't get enough. His line was that the foundry supply only about 2% of what he'd need across all his companies. And, {quote} "There's a maximum rate at which they're comfortable expanding, and that rate is much less than we would like." >> So, the whole thing is a tantrum at the supply chain. "You won't sell me enough chips? Fine. I'll print my own." >> A $40 billion tantrum. But, you know, sometimes those work out. Now, Terafab doesn't land in a vacuum. Central Texas is already one of the busiest chip regions in America. Bella, set the scene. >> So, the anchor tenant out here is Samsung. They're building a huge complex up in Taylor, Texas. That's about 30 miles northeast of Austin. We're talking a roughly $37 billion investment. Two leading-edge fabs and R&D center, backed by up to $6.4 billion in federal Chips Act money and $250 million in Texas state incentives. >> And, it's nearly finished. The complex is over 90% complete, targeting 2-nanometer production starting this year and next. >> Here's the catch though. And, this is the part that should make every Terafab booster nervous. Samsung delayed that plant, pushed mass production back, and one report said completion was held back because, {quote} "There are no customers." >> A finished multi-billion dollar fab with nothing to make. >> That's the nightmare. You can build the most advanced factory the continent and have it sit there gleaming and empty because nobody ordered the chips. >> And yet the region keeps attracting them. Texas Instruments is expanding, even partnered with Apple on a new plant. There's an older Samsung fab in Austin that's run for decades. This is a generations deep cluster. >> Which is the real reason to build here. The suppliers, the engineers, the know-how, it's already in the ground. >> So why the sudden gold rush? That's policy, isn't it? >> It's almost entirely policy. The 2022 Chips and Science Act put about $52 billion on the table. $39 billion of that aimed straight at manufacturing. Plus a 25% tax credit on fab equipment. And it worked like a magnet. Companies have announced more than $600 billion in private commitments, 130 projects, 28 states. >> It's the biggest industrial policy bet in modern American history. >> But the ground has shifted under it, hasn't it? The politics in 2026 are different. >> Completely. The current administration prefers tariffs to subsidies. The threat is roughly 100% tariffs on foreign-made chips. Unless you build in the United States. They've already slapped a 25% tariff on certain advanced AI chips. >> So the logic flips. It used to be build here, we'll pay you. Now it's build here or pay us. >> And in that light, a giant American fab announcement, even a shaky one, is beautifully timed. >> Beautifully timed is doing a lot of work in that sentence, George. >> It usually is. All right, the hard question. Adam, you've been holding the engineering reality. Can this thing actually get built? >> Let me give you the competition first because it puts the price tag in perspective. TSMC, the company that makes most of the world's advanced chips, is spending 52 to 56 billion dollars in capital this year alone. Just to stay ahead. >> So, TSMC spends the entire Austin Terafab budget every single year as a routine expense. >> Exactly. And their American project, the Arizona fabs, is a reported 165 billion dollars for six factories. Their two-nanometer production in Arizona isn't expected until around 2029. That's the company that's best on Earth at this, and it's slow and brutal even for them. >> And the equipment? Tell them about the machines. >> One EUV lithography machine, the tool that prints the finest features, costs about 180 million dollars. The next generation, the high-NA tools, run around 400 million each. A leading-edge process has more than a thousand steps. Intel's own rule of thumb is that a typical fab packs around 1,200 multi-million dollar tools, costs about 10 billion, and takes three to five years. >> And Terafab wants to do all of that at the largest scale ever with a team that has built exactly zero fabs. Back up one second. Two nanometer. For everyone who doesn't live in a clean room, what does that even mean? >> Roughly how small the features on the chip are. Smaller features mean you pack more transistors into the same space, faster chips that sip less power. Two nanometer is the bleeding edge, only just entering production anywhere on Earth. It's the single hardest thing manufacturing humanity currently knows how to do. >> And Terafab wants to start there. Not work up to it. Start at the summit. >> From a base camp they have never once visited. >> And there's the treadmill problem. Even if Terafab somehow hit 2 nanometer by 2028, that's already 3 to 4 years behind where TSMC expects to be. They're talking about 1.4 nanometer by then. You could spend $50 billion and still ship a chip that's a generation behind. >> Which we have actually seen happen. Samsung, the world's number two foundry, has struggled to win 2 nanometer customers. And Musk's own next AI chip reportedly slipped 6 months because of Samsung's 2 nanometer yield problems. >> So he's been burned by exactly the thing he now wants to do himself. >> From a standing start. Yes. >> And building it is only half the battle, isn't it? You then have to make the thing actually work. >> That's the part outsiders always underestimate. Pouring the concrete is the easy part. The hard part is yield. Getting a high enough fraction of those thousand step wafers to come out as working chips instead of expensive sand. Companies with decades of experience still bleed money for years before a brand new node pays off. >> So even in the dream scenario where Terafab rises right on schedule, there's a long brutal valley between the lights are on and we're shipping good chips at a profit. And that valley has killed fabs run by people who genuinely knew what they were doing. >> And then there's power. This is the one that really stops me. >> This is the binding constraint. The Texas grid, ERCOT, is already setting records. The forecast for this summer is a peak above 92 gigawatts. That's enough for about 18 million homes. And the long-range forecast says demand could top 200 gigawatts before the decade is out, nearly double the peak from just 2 years ago. >> And the waiting list, the queue of big facilities wanting to plug into the Texas grid, has ballooned to something like 410,000 megawatts. Data centers, crypto mines, fabs, all showing up faster than anyone can build power lines. >> And into that, Musk floats a factory he describes with the word terawatt. >> Now, to be fair, a terawatt of compute is a compute metric, not a literal electrical draw. But the grid's all-time peak is under 1/10 of a terawatt. Even a fraction of that ambition is the kind of load that reshapes a regional grid. >> And this is Texas, the state that in February 2021 had its grid freeze and go dark in a winter storm. Can the grid take it is not a rhetorical question here. It's a memory. >> Plus the water. Fabs drink ultra-pure water by the millions of gallons a day in a part of the world that knows drought. >> And don't forget the people. There's a national shortage of semiconductor technicians, the ones who keep a billion-dollar tool alive at 3:00 in the morning. Every fab announcement comes stapled to a community college training pledge because we are not minting those workers fast enough to staff what's already being built. >> So even the labor is a ceiling. >> Electrons, water, and trained people. Those are the real limits, not ambition. >> So let's zoom out. Strip away the spectacle. Is there a real story underneath? >> There are two, and they're both true at once. Story one, America genuinely does need more chip making. Our share of global chip manufacturing fell from about 37% in 1990 to roughly 10 to 12% today. >> And we learned why that matters the hard way. The chip shortage in 2020 through 2022 idled auto assembly lines around the world. A country that can't make its own chips, it turns out, can't make its own cars. >> That fear is the engine behind every one of these announcements. So, the underlying goal completely sound. >> Sound goal, yes. Shaky vehicle. And I want to quote the skeptic here because he's an industry analyst, not a hater. Mark Lipacis looked at Terafab and said, "Quote, there is a 0% chance that the Terafab will move into production by 2027 or 2028." And he put it at an 80% chance the Terafab never happens at all. >> 80% never. That's a striking number from someone who knows the industry. >> And the track record backs the caution. Tesla's own 4680 battery cell, much hyped, hit only about 2% of its original volume goal after 5 and 1/2 years. A Musk manufacturing announcement and a Musk manufacturing result are very different objects. >> And the money's not even committed. Tesla's own finance chief admitted the 20 to 25 billion dollar cost isn't yet in the company's capital plan. Meanwhile, the car business is in a sales slump. So, you can read this, as Electric did, as a company trying to staple itself to the AI story. >> And yet, for one little county, the filing alone is enormous. Grimes County, the old Gibbons Creek Reservoir land, a rural place that would be utterly remade by even the first phase of a 55 billion dollar fab. >> For better and worse. Jobs and tax base on one side, water, traffic, and the risk of a half-built plant if it stalls on the other. Ask Taylor how it feels to host a giant that delayed. >> And to be fair to the dream for a moment, if it worked, it would matter enormously. A vertically integrated, American-owned, leading-edge fab at that scale is exactly the kind of capacity the whole CHIPS Act was trying to buy. And Musk has done the impossible on paper thing before. Reusable rockets were a punchline until suddenly they weren't. >> That's the fair version, yes. The graveyard of his missed timelines is real, but so is the list of things people swore he could never pull off. >> Both can be true. The man blows through deadlines and reshapes entire industries, usually at the same time. So, where does that actually leave us? The direction, more American fabs, more resilience, is real, and it's bipartisan. Whether Terafab specifically is the project that delivers it, that's the open question. >> And the honest answer right now? >> Probably not, but watch the filing, because the gap between a Musk announcement and a Musk factory is where this whole story actually lives. >> Terafab. A terawatt of ambition, a county hearing, and a very long way to go. The need is real. The factory, for now, is mostly a very large idea on a very large stage. >> We'll be watching the wafer starts, if they ever start. >> If, the biggest word in semiconductors. >> Thanks for listening. Take care of yourselves, and keep an eye on Central Texas. Something's getting built out there. The only question is what. This episode was generated with AI by Demades.