How a Car Company Plans to Build the Biggest Chip Factory Ever
Description
Elon Musk says every chip factory on Earth, added together, makes about 2% of the AI chips his companies will need. So in March 2026 he launched Terafab: a plan for Tesla and SpaceX (which absorbed xAI in February) to make their own chips in Texas, with design, masks, logic, memory and packaging under one roof. Here is what it takes, what it costs, the case for and against, and what exists as of November 2026.
Chapters
0:00 Intro
1:10 The announcement
2:55 Why he says he needs it
5:41 Inside a chip factory
7:57 The one-company machine
9:48 Water, power and time
12:08 Who makes the world's chips
14:10 The money
16:15 The site
19:00 Partners and rivals
21:06 The case against
23:28 The case for
25:21 What is real so far
26:59 In short
Sources
- Tom's Hardware, launch: https://www.tomshardware.com/tech-industry/elon-musk-formally-launches-20-billion-terafab-chip-project
- Electrek: https://electrek.co/2026/03/22/tesla-spacex-terafab-chip-factory-ai-desperation/
- Wikipedia, Terafab: http
Transcript
Read auto-generated transcript (4598 words)
Kind: captions Language: en Every chip factory on Earth added together makes about 2% of the chips Alon Musk says his companies will need. So in March 2026, he announced that they would build their own. He called it Terraab. It is meant to be the largest chip factory ever built, 100 million square ft under a single roof on the site of an old coal plant in rural Texas. Its goal is enough chips every year to draw a terowatt of power. By Musk's own count, that is about 50 times all the AI chips the world makes today. The price tag has grown from 20 billion to as much as 119 billion depending on who is counting and which phases they count and people who follow the chip industry closely are split. One calls it a step change. Another gives it an 80% chance of never happening. So what is Terraab really? Why would a car company and a rocket company want to make their own chips? What does it take to build a modern chip factory? And how much of this one actually exists? This is the whole story. The promises, the physics, the money, and the arguments on both sides. A week earlier, he had posted that the project would launch in 7 days. On the 21st of March 2026, Musk launched Terrafab inside the Seome power plant, an old concrete power station in downtown Austin that no longer makes power. The event was streamed live on X. Tesla and SpaceX would build the factory together and SpaceX now owned Musk's AI company Xi, which it had absorbed in February. He called it the most epic chip building exercise in history by far. But he gave few details that night. No construction dates, no production date and no name technology partner. And he gave the logic in one line. We either build the terapab or we don't have the chips and we need the chips. The idea is to put every step under one roof. Designing the chip, making the masks that print it, making the logic chips and the memory chips, then packaging them and testing them. Normally those steps happen at different companies, often in different countries. Musk said no other site in the world combines them all. His pitch is speed. Make a chip, test it, fix the mask, and try again without shipping wafers around the world between each attempt. It would make two kinds of chips. One small and efficient for Tesla's cars and its humanoid robot Optimus. the other more powerful and hardened for space, for satellites that would run artificial intelligence in orbit. He even said that millions of Optimus robots would one day help build and run the factory. "We're starting a galactic civilization," he told the crowd. "Even by Musk's standards, it was a big night. Today, Tesla designs its own chips, but it does not make them. It pays TSMC in Taiwan and Samsung in South Korea to make them. In July 2025, Tesla signed a 16.5 billion dollar deal with Samsung to make a future chip, the i6, at a new Samsung factory in Taylor, Texas. Musk said the strategic importance of that deal was hard to overstate. Relying on outside factories has cost Tesla time. According to electric, it reported that Tesla's i5 chip had slipped to mid 2027 and that the i6 was delayed by about six months because Samsung's newest process ran late. In August, Tesla put it bluntly. Both Tesla and SpaceX will need far more chips than current and future global production can supply. Musk wants chips for four hungry customers at once. cars that drive themselves, robots that walk and work, XAI's chatbot and its giant data centers, and Spac X's satellites. At the launch, he put a number on the gap. All the world's AI chips made in a year, he said, add up to about 20 gawatt of power. That, he said, is about 2% of what his companies will eventually need. His target is a terowatt a year. A terrowatt is a,000 gawatt, 50 times the 20 he says the world makes now. He also said the factory should turn out 100 to 200 billion custom chips a year. SpaceX's own filing put a terowatt at about a,000 times the output of its two flagship data centers combined and most of it by his own estimates would not stay on the ground. At the launch he said about 80% would go to satellites. In August, he guessed about 75% for what he called AI spacecraft and 25% for Optimus robots. SpaceX's own stock market, filing Tide Terraab to orbital AI data centers starting as early as 2028. In January, SpaceX had asked the US regulator, the FCC, for permission to launch up to 1 million satellites, an orbital data center powered by the sun. Musk argues that sunlight in space is about five times stronger than on the ground and that computing in orbit could soon become cheaper than on Earth. In late October, he went further. Terrestrial compute will be peanuts compared to orbital compute. He expects more computing to be launched into orbit each month than built on the ground by around the end of 2028. Many experts doubt that. Electrax Fred Lambert wrote that it has essentially no connection to near-term business reality. To judge whether Terrafab can work, it helps to know what a chip factory, a FAB, actually does. Everything starts with a wafer, a dis of ultra pure silicon 300 mm across, about the size of a dinner plate. Hundreds of chips are built on each wafer at once, layer by layer, through more than a thousand process steps. Light prints a pattern. Chemicals etch it away. Metals and insulators are laid down. Then it all repeats dozens of times until billions of tiny switches, transistors are wired together. Chip generations are named in nanome. Terrafab is aiming for 2 nanome, the newest generation in volume production. But here is a surprise. The name 2 nanome has no relation to any actual physical feature on the chip. It is mainly a marketing label for a new generation that is denser, faster, and more efficient. TSMC started making its two nanometer chips in volume at the end of 2025. Intel's rival process called 18A started around the same time. At these sizes, a speck of dust is a boulder, so chips are made in clean rooms. The cleanest class allows just 10 tiny particles in a cubic meter of air. Workers wear full bodysuits and the air is filtered all the time. A typical modern fab has a clean room of about 40,000 square meters holding around 2,000 machines, each fed by about 50 pipes and cables. Factories like this are measured in wafer starts. How many fresh wafers go in each month? A typical leading edge logic fab handles about 20,000 a month. According to Tom's Hardware, TSMC, the biggest chip maker of all, shipped about 15 million wafers in 2025, its best year ever. At the launch, Electric reported a terapab target of 100,000 wafer starts a month to begin with, and 1 million at full scale. Electric put that at roughly 70% of TSMC's current output. By our own arithmetic, 1 million wafers a month is the work of about 50 typical leading edge fabs. The hardest step is printing the pattern. For the finest layers, the leading chip makers all rely on one kind of machine, an extreme ultraviolet or EUV lithography scanner. Only one company in the world makes them, a SML in the Netherlands. Reuters describes it as roughly the size of a school bus. It weighs about 150 tons and costs around $250 million. The newest high versions cost about $400 million each. According to chip journalist Mark Leetis, they are so costly that by his account TSMC has no plans to use them for its next few generations of chips. Inside, droplets of molten tin fly past a laser about 50,000 times a second. Each one is blasted into a plasma that gives off extreme ultraviolet light. That light is bounced off mirrors smoother than those in space telescopes and onto the wafer. Meanwhile, the table holding the wafer floats on magnets, racing back and forth so each chip can be printed in turn. One machine ships in about 40 containers and flies on jumbo cargo jets and a SML can only build a few dozen of them a year for the whole world. Analysts expect that number to grow, but so is the demand. That matters for Terrafab. Tom's hardware estimates that a 3 nanometer class fabs 80 to 100 lithography scanners of all kinds and every chipmaker on Earth is already waiting in the same line. In October, journalist Tim Culpen reported that Terrafab's first EUV machine is due in mid 2027, possibly before its building is ready. Neither a SML nor Terrafab has confirmed it. FABs are also thirsty. TSMC's site in Arizona uses about 4.75 million gallons of water a day and recycles about 65% of it. It is building a plant to raise that to 90% or more by 2028. And they are hungry for power. TSMC's electricity use nearly doubled between 2017 and 2022 to about 21,000 gawatt hours a year. One projection had the company alone using more than a tenth of all of Taiwan's electricity. Then there is the building itself. The engineering firm Excite says a leading edge fab takes around 83,000 tons of steel, 5,600 m of wiring, and 30 to 40 million hours of work. Intel has described a typical fab as taking 3 to 5 years to complete. Exite puts a leading edge fab at more than 20 billion with 4 to6 billion for the building alone. It also takes people. A leading edge fab needs 4 to 7,000 construction workers at its peak. According to Tom's Hardware, TSMC expects its Arizona expansion to create 40,000 construction jobs, and its first Arizona fab employs about 3,000 people once running. Excite also compared how long fabs take from design to finished building. In Taiwan, about 19 months. In the United States, about 38. twice as long, mostly because of permits and because sites are not worked around the clock and finishing the building is only the start. TSMC's chairman Ci has said a new fab takes 2 to 3 years to build, then one to two more years to reach full output. That second stretch is about yield. The share of chips on each wafer that actually work. New processes often start with many chips failing. Samsung's 2 nanometer yields were reported at 50 to 60% into early 2026. Intel's newest process was reported at around 65% in April and rising. Every failed chip still costs the full price to make. That is why experience matters so much in this industry. Which brings us to who does this today. In the second quarter of 2026, TSMC took 72.5% of all the money paid to contract chip makers. According to Trend Force, Samsung was a distant second at 5.9%. Intel is trying to win outside customers, too, but its foundry business lost more than 10 billion in 2025. and TSMC based in Taiwan is estimated to make more than 90% of the world's most advanced chips. Having so much of the most advanced production in one place worries many governments. The United States, among others, has paid chip makers to build factories at home. In America, that effort is the Chips and Science Act signed in August 2022. It set aside 52.7 billion, including $39 billion to help build factories. The biggest result is in Arizona. In May 2020, TSMC announced a 12 billion fab there to start production in 2024. It began production around the start of 2025. After TSMC brought in about 500 workers from Taiwan to help finish the first phase, its plans have since grown to six fabs and 165 billion. 2 nanometer production there is not expected until 2029, but it works. This year, Apple said it would buy more than 100 million chips made at TSMC's Arizona site. Intel's Ohio project was announced in 2022 as a $20 billion plan set to open in 2025. It has been pushed back to 2030 and 2031. And Samsung's Texas FAB was delayed for lack of customers until Tesla's deal arrived. These are companies that have built fabs for decades. That is the bar Terrafab has to clear. So what will Terapab cost? It depends on the date of the headline. At launch, reports said about 20 billion. Electric put it at 20 to2 billion. In April, Musk said Tesla's first step, a small research fab in Austin, would be a $3 billion project, making a few thousand wafers a month. In May, a county filing in Texas listed $55 billion for the first phases and up to 119 billion if every phase is built. In August, the state of Texas confirmed a first phase of $16.8 billion. For its part, Tesla's finance chief had said in March that the full cost was not yet in Tesla's spending plan for 2026, which was already above $20 billion. State filings later laid out four phases. Phase 1 runs from 2026 to 2028. Three more would follow, reaching 2036. Each later phase needs its own decision to go ahead. How much would the factory Musk described at launch really cost? Bank of America's analysts estimated that 100,000 wafers, a month of capacity alone, could need more than $60 billion. Chip journalist Mark Lapetis wrote in May that a new leading edge fab costs well above the popular $20 billion figure. For scale, the whole chips act set aside $52.7 billion and the full Terraab plan approaches the $165 billion T SMC plans for Arizona. Analysts at Bernstein went much further, reaching a full terowatt a year, they estimated could take up to 358 fabs and about $5 trillion. They called it a very rough back of the envelope calculation. For comparison, TSMC, the biggest chip maker in the world, expects to spend 60 to 64 billion on new equipment and factories in all of 2026. The main plant will not be in Austin. It is planned for Grimes County, about 90 mi northeast of Austin and about an hour from Houston. It is a rural county. The nearest town, Anderson, has fewer than 300 people. The company also looked at sites in Arizona, New Mexico, and Nevada. The site wraps around Gibbons Creek Reservoir, where a coal fired power station ran from 1982 to 2018. The zone set aside for the project covers more than 22,000 acres. The plan calls for 100 million square ft under a single roof. A local TV station put that at more than 1,700 football fields stretching for over 3 m. Musk wrote that Terraab Texas will be the largest and most valuable building on Earth by far. It will not run on Tesla's solar panels. SpaceX said it would bring its own power, natural gas plants, and very large battery arrays. It plans to buy $2.8 billion worth of gas turbines over 3 years. Water would come from the reservoir, not from local groundwater, with treatment and reuse on site. The company must also maintain six county roads used by its construction trucks and repair any damage they do. To win the project, Texas offered a $30 million grant. Grimes County approved a full property tax break for up to 10 years in return for payments of $20 million a year. The county deal requires at least $5 billion of investment and 1,800 jobs. The state's announcement promised 3,000. And in its application, the company told Texas that without the tax limits, Arizona would offer a better return. Two local school districts signed tax agreements, too. One superintendent called it a defining moment and promised to manage it with students at the center. A local citizens group has questioned the long-term cost of those tax breaks and how much time the public had to study the maps before the vote. Not everyone nearby is thrilled. Danny Cross has lived on 31 acres beside the reservoir for 22 years. He told local station KBTX he is not against Terraab, but he has put his land up for sale. "We were building what we thought was our forever house," he said. Another neighbor worried about water and power bills. A local construction worker saw it differently. "It's coming, so get ready." In September, after two residents asked to see Terrafab records, SpaceX sued the Texas Attorney General and the county to keep some of them private, including costs and design details. Musk's companies had never run a chip factory. So, in April, they found a partner. On the 7th of April, Intel said it was proud to join the Terraab project to help refactor silicon fab technology. Its shares rose more than 2%. Intel chief executive Lip Bhutan called Terapab a step change in how silicon logic memory and packaging will get built in the future. 2 weeks later, Musk said Terapab planned to use Intel's newest process called 14A which he said was not yet totally complete. For Intel, it was a welcome win. Its foundry business had lost money and the US government had become its largest shareholder. Analyst Gilura called the deal an important step in Intel's restructuring. Borrowing a process is not new. Global foundaries once licensed one from Samsung and Japan's new chipmaker Rapidus licensed its two nanometer design from IBM. Then in October, the plot twisted. Journalist Tim Culpan reported that TSMC was exploring a role in helping to run Terraab. Musk replied, "Just discussions, but something may come of it." Around the same time, he reportedly admitted he still had a lot to learn about making chips. 4 days later, he was firmer. "No, we will build and run the fab. Let there be zero doubt about that. Maybe TSMC subleasases part of the terapab if they want," he added, but nothing more than that. TSMC confirmed the talks in mid-occtober and said it had sent teams to Texas, but chairman Ci set a condition. Any fab making TSMC chips would have to be run by TSMC itself. Then on the 2nd of November, the Taiwanese publication Diggy Times reported that Musk may be softening with control and joint investment on the table. Neither side has confirmed that. Intel for its part says it is still working on the project. Now the arguments. First the skeptics in their own words. Mark Lapetis, editor of the chip newsletter semi-ecosystem called Terrafab a PowerPoint presentation. He gave it a 0% chance of production by 2028, a 20% chance of a pilot line by 2029, and an 80% chance that it never happens. his reasons. Musk's companies have no experience building fabs, years of process development and yield problems ahead, and by the time it ran, rivals would be 3 to four years further along. He also argued that a fab earns no revenue on its own and should be the last priority for a car maker, he says, is losing market share. His advice, if Musk goes ahead, hire a seasoned foundry chief, partner with experienced chip research groups, and buy memory chips instead of making them. Bank of America's analysts called large-scale production before the end of the decade almost impossible and said Terapab's chips could cost 30 to 50% more to make than TSMC's. Tom's Hardwar's Anton Shillov ran the numbers on a full terowatt and estimated it would take a SML about 70 years to supply the scanners for the logic chips alone. While Elon Musk tends to raise enough money for his projects, he wrote, "Time is something money cannot buy. People are another limit." Schill estimated a full terowatt would need more than 300,000 skilled fab workers. TSMC, the largest chip maker, had about 84,000 employees at the end of 2024. Bank of America also pointed to missing pieces. The design software and libraries of proven circuit blocks that established foundaries have built up over decades. Skeptics also point to Musk's timelines. In 2019, he said driverless Tesla roboaxis would be on the road the next year. A small robo taxi service started in Austin in June 2025 with about 10 to 12 cars. At Tesla's battery day in 2020, he set big goals for a new battery cell. Electric reports Tesla is at roughly 2% of that original volume goal. And its editor called the Terraab launch battery day on steroids. Now the other side also in their own words. Musk's core argument is supply. He says TSMC, Samsung and Micron can only expand at a rate much less than we would like and industry trackers such as Trend Force report that the most advanced capacity at 3 nanome is in acute shortage. The demand is real. TSMC's revenue rose 51% in a year to 46.7 billion in the third quarter of 2026. Wedbush analyst Dan Ies wrote that while the timeline is uncertain, Terrafab will accelerate the company's ambitious AI path and that chip supply is the greatest constraint. Texas officials are enthusiastic. Governor Greg Abbott said Texas is where big ideas grow even bigger. The Grimes County judge and the regional business partnership said the area is ready to supply the workers and infrastructure. Wedbush's Ives also sees a bigger picture. He called Terrafab the first step toward a merger of Tesla and SpaceX, which he predicts for 2027. That is his forecast, not a company plan. Supporters also point to the one roof design. A chip could go from mask to tested part without leaving the building, and they point to Musk's record at SpaceX. By January 2026, it had launched its 11,000th Starlink satellite. And in June, it completed the largest stock market listing in history. Even Tom's hardware, one of the harshest critics of the terowatt goal, wrote that a partial version, making some of Tesla's and SpaceX's own chips, could be feasible. That may be the real question. Not whether Terrafab replaces TSMC, but whether it can make enough chips soon enough to matter. So, as of early November 2026, what actually exists in Austin? Tesla broke ground in April on its research fab at Giga Texas. By late September, aerial footage showed its foundation slab being poured. In Grimes County, land clearing began in August and drone footage in October showed extensive earthwork. Tesla is hiring its own construction managers for the site and SpaceX has paid the county its $10 million upfront payment. State filings list the main construction start as the 1st of December 2026 with completion by the end of 2028. SpaceX has said limited operations could begin toward the end of 2027. Everything else is still a plan or a promise. the 100 million square ft, the terowatt, Intel's process, TSMC's role, and chips in orbit. Even SpaceX's own stock market filing in May, which ran to more than 300 pages, said only that it had agreed with Tesla on a general framework, and that timelines, milestones, and spending have not yet been determined. Musk has also said Tesla's next chip, the i5, would come in small batches this year and in volume in 2027. Whether that happens on time will be an early test. What to watch next? The December construction start, whether TSMC ends up running part of the plant, the first EUV machine in 2027, and the first chips from the Austin Research Fab. In short, Musk says the world's chip factories make about 2% of the chips his companies need. So Tesla and SpaceX plan to build their own. Terrafab would put design, masks, logic, memory, and packaging under one roof in Texas with a first phase of 16.8 billion and a full plan of up to 119 billion. Making advanced chips takes machines only one company builds, huge amounts of water and power, and years of work to get yields up. Even the experts have taken longer than planned in America. Supporters see a bold fix for a real shortage. Skeptics see a timeline that money and physics will not allow. So far, there is a foundation slab in Austin and cleared land in Grimes County. So, what do you think? Will Terraab make its first chips by the end of 2028? Tell us in the comments and subscribe to Brain Light for more stories behind the headlines.