Terafab: The Chip Factory That Keeps Changing Size

Kind: captions Language: en 7 months before unveiling the most ambitious chip factory ever proposed, Tesla quietly killed its own chip supercomput project. That's the strange timing at the center of this story. And it's worth untangling what's actually been built from what's still just a slide in a keynote. In March 2026, Elon Musk stood inside a decommissioned power plant in downtown Austin and announced Terrafab, a joint venture between Tesla, SpaceX, and XAI meant to put every stage of chip production under one roof. The plan called for 1 terowatt of computing power produced every year, split across two fabs on a site as large as 100 million square ft. One fab would build AI chips for Tesla's cars and its Optimus robot. The other would supply chips for SpaceX's space-based data centers. In April, Intel signed on to help design, fabricate, and package chips at that scale, its first major outside foundry commitment in years. But rewind to August 2025, and Tesla's own public position was the opposite of ambitious. That's when it shut down Dojo, the in-house supercomputer it had built specifically to train its self-driving AI. Musk explained the decision plainly. It didn't make sense to split resources across two different chip architectures. Dojo's lead engineer left the company and Tesla said it would lean harder on outside partners Nvidia and AMD for compute, Samsung for manufacturing while folding its own chip ambitions into a single design called AI6. So, 7 months before promising the largest chip factory on Earth, Tesla was telling the world that building chip infrastructure in-house wasn't worth the resources. And here's the thing, none of Tesla's near-term chips are actually coming from Terrafab at all. The AI5 chip, Tesla's next inference chip, reached tape out in April 2026, promising 40 to 50 times more compute and nine times more memory than its predecessor. It's being built at Samsung's existing FAB in Taylor, Texas, a plant that existed long before Terraab was ever announced. AI6, the training chip meant to replace Dojo, is targeting tape out by the end of 2026 with mass production on Samsung's 2nm process arriving in 2027, part of a deal worth more than $15 billion. Also completely unrelated to Terraab. So every chip actually going into a car or a robot over the next two years is coming from an existing partner's existing factory. Terraab is something else entirely, a separate much longer bet and the numbers describing that bet keep moving. The original pitch was a 20 to$25 billion facility. A financial filing in May put the potential total investment as high as $119 billion. Then in August, the actual site was confirmed, and it wasn't Austin at all. It's in Grimes County, roughly an hour northwest of Houston, with a first construction phase priced at $16.8 billion, targeted for completion in 2028. Full buildout toward that 1 terowatt figure isn't even projected until 2029 at the earliest. Whatever terapab becomes, it clearly isn't the thing announced on that stage in March. It's in a different place, at a different scale, on a slower timeline. Analysts aren't just skeptical because Musk has a history of missed deadlines, though that history is real. They're skeptical because of the physical math. Morgan Stanley estimated that meaningful chipmaking capacity alone would require somewhere between 35 and $45 billion in capital, well above the phase Tesla has actually committed to to hit the full 1 terowatt target. As originally stated, some analysts calculated Terraab would need to represent the equivalent of roughly 140 separate foundaries at a combined cost approaching $5 trillion. Then there's the equipment problem. The most advanced chip production requires extreme ultraviolet lithography machines made by exactly one company on Earth, ASML, whose order book is already full for years out. No amount of money fixes a supply chain when there's only one supplier and it's already sold out. Analyst Mingchi Qua flagged three specific risks. The scope is too large. The timeline is too tight. And the specialized workforce needed to staff a semiconductor megaactory doesn't currently exist at that scale in Texas. One trade publication went further and called the whole concept a pipe dream. To be fair to Musk, he addressed this directly, pointing out that critics said the exact same thing about reusable rockets and mass market electric cars, both of which he delivered against genuine expert skepticism. That track record is real, too, which is part of why investors haven't simply dismissed this outright. So, strip away the keynote language. And here's the honest picture. A real joint venture exists. A real construction site has been confirmed. A real $16 billion first phase is underway. A real chip company, Intel, has signed on. But the number that made headlines, one terowatt of comput a year, would require an industrial buildout on a scale the entire world's chip industry doesn't currently have the equipment, workforce, or capital to support. Terraab might still turn into something enormous. But right now, it's a construction project in Grimes County that's already redefined its own scope three times in 5 months. And it hasn't broken ground on the part that was supposed to make it historic.