Can Musk Actually Pull Off TERAFAB? The $25 Billion Question

Kind: captions Language: en On March 24th, 2026, Elon Musk walked into a defunct power plant in Austin, Texas, stepped on stage, and called what he was about to describe the most epic chip building exercise in history by far. $25 billion, one roof, one terawatt of computing power a year, announced by a man whose companies have never fabricated a single chip in their lives. Within hours, chip equipment stocks jumped. TSMC's analysts started running the numbers, and the whole industry asked the same question: Is he serious this time? Because this is the biggest private semiconductor bet in American history, from a car company and a rocket company that have never built a chip. Today, we separate what is real about Terafab from what is fantasy, because the answer is stranger than either side admits. The chip crisis Musk says is coming. To understand why Musk is doing this, you have to understand the problem he says he is facing. At Tesla's late 2025 earnings call, Musk warned investors that the external chip capacity his companies rely on, from TSMC, Samsung, and Micron, would hit a hard ceiling within 3 to 4 years. On the Terafab stage, he put it directly. He said he was grateful to the existing suppliers, but that there is a maximum rate at which they are comfortable expanding, and that rate is far slower than what his companies need. The numbers behind that are genuinely staggering. The single largest driver is the Optimus humanoid robot program, which is estimated to require around 20 million chips a year, roughly six times Tesla's entire current automotive chip demand. Stack that on top of the cybercab, full self-driving, and SpaceX's satellite ambitions, and Musk claims that all the chip fabrication on Earth today covers only about 2% of [music] what his companies will eventually need. And [snorts] the demand is not theoretical or years away. It is already biting. Tesla's autonomous ambitions, its robotaxi fleet, and its robot army all run on silicon the company does not control and cannot get enough of. Every chip TSMC sends to Nvidia or Apple is a chip Tesla cannot buy, and Musk knows the line only grows longer. His conclusion was blunt. We either build the Terafab or we don't have the chips, and we need the chips, [music] so we build the Terafab. And here's the important part. The shortage is real. Tesla's own AI 5 chip has already slipped to mid-2027. Samsung's 2-nanometer process fell behind, dragging the AI 6 with it. In January 2026, TSMC warned even giants like Nvidia and Broadcom about capacity constraints. The crisis Musk is describing is not invented. The only real question is whether his solution makes any sense. What Terafab actually is. So, what is he actually proposing to build? Terafab is planned as a vertically integrated mega facility on the north campus of Giga Texas in Austin, a joint venture between Tesla, SpaceX, and xAI, which SpaceX absorbed earlier in 2026. The plan is to put every single stage of chip production under one roof. Design, lithography, fabrication, memory, advanced packaging, and testing, all in one place. No other company on Earth currently does all of that at a single facility. It is built around two product lines. The first makes energy-efficient inference chips for Tesla's vehicles, the cybercab, and Optimus. The second makes a radiation-hardened chip designed specifically for SpaceX's orbital satellites and space-based AI. The initial target is 100,000 wafer starts per month, scaling toward 1 million. That full-scale number alone would represent roughly 70% of TSMC's entire current global output from one site, producing somewhere between 100 and 200 billion chips a year. To put that volume in perspective, the entire current global output of advanced AI chips sits [music] at roughly 20 gigawatts of compute. Terafab's target of 1 terawatt is 50 times that. Musk is not proposing to join the chip industry. He is proposing to build something larger than all of it combined. But here's the part most coverage glossed over and where it goes from ambitious to almost science fiction. Musk said 80% of Terafabs compute output would be aimed not at the ground, but at orbit. [music] His argument is that sunlight in space is about five times stronger than on Earth's surface and that dumping heat into the vacuum makes it possible to scale computing in ways you cannot on the ground. He claimed orbital AI compute could become cheaper than building it on Earth within two to three years. And then he said the quiet part out loud. We're starting a galactic civilization. Quick pause here. I'm doing a $100 Amazon gift card giveaway to celebrate this channel reaching 1,000 subscribers. No action required to enter. I'll randomly pick a winner once we hit the milestone. If you want to follow the journey or see the winner announcement, feel free to subscribe. Intel joins and what it actually means. Now back to it because three weeks later the entire story changed. On April 7th, 2026, [music] Intel announced it was joining Terafab alongside Tesla, SpaceX, and xAI. And the wording of Intel's own statement is the most revealing detail in this whole saga. Intel said its ability to design, fabricate, and package ultra-high-performance chips at scale would help accelerate Terafab's goal. Read that carefully. Design, fabricate, and package. Those are the three things that actually make a chip fab work, which points to the real story underneath the press releases. Tesla is not building a chip fab from scratch. Intel is building it for them. Musk later confirmed Tesla would use Intel's most advanced manufacturing process, a 1.4 nanometer class node. What is really happening is that Tesla and SpaceX are financing and anchoring an Intel foundry expansion in Austin and getting to call it Terafab in the announcements. And the fascinating twist is who needs this more. >> [music] >> Intel's CEO has openly said the company would exit chip manufacturing altogether if it failed to land an external customer for that 14A process. Tesla is that customer. For Intel, a company that has spent years struggling to prove its foundry business, this is the marquee win it has been desperate for. So, in a real sense, Tesla didn't just announce a fab, it threw Intel a lifeline. The scale of what they're attempting To grasp how enormous this actually is, look at what it normally takes to do this. TSMC spent around 165 billion dollars and decades to reach the 2-nanometer process. A single 2-nanometer fab costs roughly 28 billion dollars and takes at least 38 months just to build in the United States. The equipment alone is a nightmare of scale. A facility this size would need 50 or more EUV lithography machines from a company called ASML at around 380 million dollars each, which represents a meaningful slice of ASML's entire annual production. Those machines have lead times of 18 to 24 months. The procurement challenge by itself would strain the entire global supply chain. And then there is the budget problem, which is the real crack in the story. Analysts at Bernstein estimated that building enough capacity to deliver one full terawatt of annual compute would cost somewhere between 5 and 13 trillion dollars. And sure enough, by May 2026, public filings revealed the first phase alone was now pegged at 55 billion dollars with the full buildout potentially reaching 119 billion. The problems nobody is saying out loud. Beyond the money, there are problems that money cannot easily solve. Start with vibration. The proposed site sits next to Giga Texas stamping machines that shake the ground with every press cycle. The 2-nanometer photolithography that Terafab depends on requires near perfect stillness. Then there is the workforce. >> [music] >> A fab this size needs 8 to 10,000 specialized engineers in disciplines that take years to train. TSMC's Arizona plant already struggled so badly to recruit American engineers that it had to fly in hundreds of technicians from Taiwan, and Terafab's requirement is roughly five times larger in the same labor market. There is also the matter of yield. Intel's most advanced process currently runs at roughly 65% yield, meaning a third of every wafer is scrapped. The entire Terafab ramp quietly depends on that number climbing substantially, and Wall Street is not yet convinced it will. And underneath all of it is the simplest problem. Tesla has never run a single wafer through a fab. Morgan Stanley called the idea of building a leading-edge fab from scratch Herculean, and honestly, that word undersells it. What it means if it works. But here is why none of the skepticism makes this safe to ignore. Right now, Nvidia's grip on the AI industry depends almost entirely on there being no credible alternative at the cutting edge. Terafab, even as a distant threat, changes [music] that calculation. Every AI lab, every cloud provider, and every government currently dependent on Nvidia hardware would suddenly have another option on the table. That is not a normal market shift. That is the potential restructuring of one of the most powerful supply chains ever built. And even if Terafab never produces a single competitive chip, simply being a credible threat gives Musk enormous leverage to negotiate better pricing and priority from TSMC and Samsung. Then there is the orbital vision. If it works, the next generation of data centers would not sit on Earth at all. They would be in space, powered by unlimited solar energy, free from the land and power grid [music] limits that constrain every data center today. Musk describes satellites eventually assembled on the moon and launched into orbit to form the largest computing network [music] in history. As Goldman Sachs framed it, whoever controls AI compute defines the next era, and Terafab is the first serious private attempt to move that infrastructure off the planet entirely. [music] Musk has been wrong on timelines before. Those Battery Day targets are still sitting at 2% of the goal 5 years later, and Terafab may follow the same path. Whether Terafab delivers or becomes the next 4680 story, the world it announced already exists, and that changes everything about what comes next. Subscribe for the honest version of these stories instead of the hype. We do this every week.