Elon Musk's $16.8B Terafab: He Only Signed for $5 Billion

Kind: captions Language: en Elon Musk says his companies will soon need more chips than the entire planet can make. Not more than TSMC, not more than Samsung, more than all of them combined. Then he said he would fix that himself and put a number [music] on it. In August of 2026, the state of Texas confirmed the first check, >> [music] >> 16.8 billion dollars for a factory going up on the site of a dead coal plant [music] an hour outside Houston. Three months earlier, Musk's own company had told federal securities regulators [music] that the first phase would cost 55 billion. Same project, same year, 38 billion dollars apart. Nobody involved has explained the difference. The project [music] is called Terafab and the argument over what it costs turns out to be the smallest of the arguments about what it actually is. Start with the name because it comes from the target. Musk says Terafab will eventually produce, [music] in his words, over 1 terawatt of compute per year. That isn't a unit. Nobody in the semiconductor industry [music] measures a factory in terawatts. Fabs are measured in wafer starts per month. How many polished silicon discs enter the line, each one carrying thousands of chips. >> [music] >> So, translate it. The plant opens at 100,000 wafer starts a month. The ceiling is 1 [music] million. When the project was announced, that top number was put at around 70% of everything TSMC makes today, the company behind nearly every leading edge chip in your life, 100 to 200 billion chips a year from a company that has never manufactured one. To hold a line that size, they are proposing a single building of 100 million square feet, roughly 36 times the office space inside the Empire State Building. Musk's own description was blunt, the largest and most valuable building on Earth by far. Which is a strange ambition for a car maker and a rocket company given that neither of them has ever run a fab and the one they are describing would be the most complex industrial facility ever attempted. So, the question isn't whether it's ambitious, it's what makes it necessary. It isn't a shortage of manufacturers willing to take the work. That part is already handled. Tesla designs its own silicon, the AI 5 and AI 6 processors meant to run full self-driving and the Optimus robot. It has committed roughly 16 and 1/2 billion dollars to Samsung [music] to build them on a contract running through 2033. TSMC produces for Tesla as well. The problem is [music] the other customer on the list. The one no foundry on Earth will pour concrete for. Terafab is planned as two fabs, each making essentially one chip design. One for the road, the processors going into cars and humanoid robots. And one for orbit. A radiation-hardened chip built to survive constant bombardment and higher operating temperatures. Intended for SpaceX's [music] planned data centers in space. No foundry is going to sink 20 billion dollars into capacity for computers that will one day be flown into orbit on a rocket that is still being tested. So, the logic runs backward from there. This isn't a car company that decided to make chips. It's a company that intends to put data centers in space, worked out how much silicon that takes, and found that the supply chain it needs doesn't exist. You can't buy your way out of that. And starting from zero, you can't build your way out either. Which is why the most consequential name attached to Terafab is neither Tesla nor SpaceX. In April of 2026, Intel joined. Intel invented the business of making advanced chips in America. [music] For decades, it was the business. It has now signed on as the junior partner in someone else's factory, contributing process technology, packaging, and expertise, [music] rather than writing the check. The dependency runs the other way, too. Intel's entire comeback rests on a next-generation manufacturing process called 14A, which enters risk production in late 2027. As of August 2026, [music] the only publicly confirmed customer for it is Tesla. The numbers underneath that are stark. Intel's foundry arm booked $293 million from outside customers in a single quarter against an operating loss of 2.1 billion. In August, it sold $15 billion of stock to fund capacity for customers who, with one exception, have not publicly committed [music] to anything. So, Intel is staking its manufacturing future on a company that has never made a chip, and that company is staking its silicon future on a manufacturer that spent years [music] losing the technology lead. Each side's plan requires the other to be right. And even if they both are, there's a problem [music] sitting in front of them that neither has spent much time addressing. It has nothing to do with money. You cannot beat TSMC by buying the same machines. Every leading-edge chip on Earth is patterned by lithography equipment from one Dutch company, ASML. Its newest high numerical aperture machines run around $380 million each. TSMC buys them. Samsung buys them. Intel buys them. Terrafab would buy them. Which means the tool is not the advantage. The advantage is yield. The unglamorous, decades-deep knowledge of how to get working chips out the far end instead of expensive glass frisbees. TSMC has been compounding that knowledge since the late 1980s. Buy the same machines, and you are not a competitor. You are a worse version of the incumbent, carrying all of the cost and none of the learning. Which leaves a new entrant exactly one move. Don't buy the machine. Change it. In August of 2026, someone [music] posted about lithography light sources. Musk replied with four letters, FEL FTW, free electron laser for the win. To see why that reply matters, you need to know how the light gets made right now. Inside an ASML machine, molten tin droplets are fired into a vacuum chamber, and a laser strikes each one mid-flight, vaporizing it into plasma. That plasma glows at 13.5 [music] nanometers, the wavelength small enough to draw a modern transistor. It works. It is also grotesquely wasteful. Roughly 1/10 of 1% of the electricity going in comes out as usable light. In practice, about 4.4 megawatts of input to produce a single kilowatt of extreme ultraviolet. A free electron laser gets there differently. Accelerate electrons to near the speed of light, run them through an array of alternating magnets, and they emit light at whatever wavelength you tune them to. No tin, no debris, and on paper, that same kilowatt of output for around 0.7 megawatts in. So, why has nobody [music] done it? Because a free electron laser is a particle accelerator. It does not fit inside a normal semiconductor plant. For 30 years, that has been enough to keep it a physics lab curiosity instead of a manufacturing tool. Now, go back to the building, 100 million square feet, five times the largest structure standing anywhere on Earth today. The size stops sounding like ego. The caveat is serious, and it belongs here. No commercial free electron laser light source exists. The most credible effort, a company called Xlight, is chaired by Pat Gelsinger, the chief executive Intel pushed out, and it is targeting a working prototype in 2028, a prototype, not a production tool. That makes it the least proven piece of the plan, and the only piece that explains how Terafab could be anything more than an expensive imitation. Musk isn't proposing to out-execute TSMC at TSMC's own game. He's proposing to change the game's most expensive rule, which introduces a constraint [music] of a completely different kind. Particle accelerators and semiconductor fabs share one appetite, power. Terafab needs somewhere between 400 and 700 MW continuously. That is the round-the-clock draw of a few hundred thousand homes with almost no tolerance for interruption. None of it is coming from the Texas grid. SpaceX told Grimes County officials the site will power itself using natural gas turbines and battery arrays. Not Tesla solar panels. There's a real engineering argument for that. Fabs need extraordinarily stable electricity, and a voltage sag lasting milliseconds can ruin a wafer worth tens of thousands of dollars. Turbines deliver that directly. Utility-scale solar needs inverters and a grid connection that in Texas can take 3 to 5 years, and first chips are targeted for late 2027. There is also this. The turbines come from APR Energy, a company Musk reportedly acquired around May 2026 for roughly a billion dollars. He owns about a fifth of Tesla. He owns essentially all of APR. The rest of the site follows the same instinct. Water drawn from Gibbons Creek Reservoir, [music] rather than local wells. Its own fire station. Its own wastewater plant. 500-ft setbacks from [music] the nearest homes and churches. It isn't a factory. It's a company town with a particle accelerator in it. The people who live next to it [music] noticed. Around 900 Grimes County residents signed a petition demanding written protections, water safeguards, and clawbacks if the project ever shrinks. Two school boards approved the tax [music] agreements five votes to two over sustained public objection. Everything about that site is built for permanence, which is what makes the paperwork so difficult to explain. Follow the number. The idea surfaced at a Tesla shareholder meeting in late 2025. By March, it was described as roughly $20 billion. By April, with Intel attached, 25. By May, the county application listed 55 billion >> [music] >> for the initial phase and 119 billion in total. In August, the governor's office announced a first phase of 16.8. Sextupled, then cut by 2/3 inside a single year. Every step of it documented in a real filing or a real press [music] release. Now look at the one document anyone has actually signed. The Grimes County agreement, executed in June, binds SpaceX to a minimum investment of $5 billion by 2030 and 1,800 jobs by 2035. Hit 90% of those targets and there is no penalty. And SpaceX can terminate the whole thing at any time for any reason [music] on 30 days notice. Announced, up to $119 billion. Signed, 5 billion and an exit door. The federal filings are colder. >> [music] >> SpaceX's own prospectus calls Terafab a general framework for future development. It states that neither Tesla nor Intel is obligated to remain part of the project. It warns, "There is no assurance the project will be completed in the expected time frames." >> [music] >> Then adds three more words, "or at all." And Tesla, the partner whose robots and cars justify half the demand, has not meaningfully described Terafab in its own federal filings at all. Which makes it tempting to call [music] the whole thing vapor. That would be the wrong read. For a $10 million payment and a signature he can withdraw [music] in a month, Musk has bought an option, the right, but not the obligation, to build the largest factory on Earth. It has already paid for itself. Intel now has a named anchor customer for the process its survival depends on. Texas attached a $30 million grant and a decade of tax abatement. And every foundry selling to Musk, Samsung, TSMC, anyone bidding on the next contract is now negotiating with a customer who has publicly [music] and expensively demonstrated that he is willing to walk. In an industry where capacity is allocated years ahead, and the largest customers get served first, a credible threat to build your own supply is worth more than most companies' entire capital budget. But an option is only leverage while people believe you might exercise it. And there is one detail that cuts against the story being told out loud. The spending schedule Terafab actually filed is back-loaded. The heavy money lands between 2028 and 2037, well after the dates in the press releases. So, the number that matters isn't 16.8 billion. It's three dates. Intel's 14A design kit due in late 2026. Nothing gets designed until it ships, >> [music] >> and nothing gets built until it's designed. First chip production targeted for late 2027. And 5 billion dollars of steel, concrete, and equipment in the ground by 2030. The first commitment a court [music] could actually enforce. If those three land, this stops being a story about Elon Musk. Because the world economy runs on an assumption almost nobody states [music] out loud, that advanced computing power is something you buy from a handful of companies made in a small number of buildings, most of them on a single island that a much larger country has spent [music] decades claiming as its own. Every AI lab, every car maker, every defense program, every phone in every pocket sits downstream of that arrangement. It has held for 40 years because building the alternative was understood to be impossible for anyone not already inside the club. Terafab is a test of whether that is still [music] true. Whether a company that needs enough compute can simply decide to make it, and whether the physics of chip making can be rewritten by someone with no legacy process to protect. If it works, vertical integration comes back to an industry that spent two generations outsourcing itself. And the foundry model gets a competitor it never planned for. If it collapses, it will still have proved the more uncomfortable point, that the threat alone was worth billions, [music] and that the most concentrated supply chain on Earth can be moved by a building that doesn't exist yet. Terrifab does not have to break TSMC to change what TSMC can charge and who it has to say yes to. The building is the argument. The signature is the hedge. And the only number that will ever settle it is the [music] one poured in concrete, an hour outside Houston on the grave of a coal plant.