Inside Giga Texas: Tesla Secretly Builds a $25B, 2nm AI Chip Plant

Kind: captions Language: en A drone flew over Giga Texas and captured something no one saw coming. Four massive construction projects underway at the same time. Not one after another, not in any sequence, but completely in parallel. What is being secretly built inside? It's Terafab, a 25 billion dollar 2 nanometer AI chip factory co-built by Tesla, SpaceX, xAI, and Intel. Two independent production lines, one for Optimus and FSD, one for Starlink satellites. First silicon is targeting using Intel's 14A process node, a technology that does not yet exist on the market. Why is Tesla compelled to make its own chips? And is 2027 even achievable? Today, we analyze every step. Real numbers, real risks. Let's dive right in. >> [music] >> Elon Musk stood before tens of thousands of people at the Cyber Rodeo event, the grand opening of Giga Texas. The production line was still unfinished. Workers were still racing through every last detail. The very first Model Y rolled out through the factory gates in a state that, honestly, was just barely done. Nobody thought this factory would become what we are seeing today. Just 1 year later, Giga Texas had reached 5,000 vehicles per week. Then that number climbed to 10,000, breaking the production record across Tesla's entire factory network. The current target is now being pushed toward 12,500 vehicles per week. And it isn't just the Model Y, the Cybertruck also began rolling off the line in late 2023, making Giga Texas the only factory in Tesla's system running two completely different product lines simultaneously. In October 2025, the milestone of 500,000 vehicles was recorded, Model Y and Cybertruck combined. During that same period, the 4680 battery cell line right on the factory campus had produced 10 million cells, a figure that only a few years ago still existed as a goal on paper. But here is the strangest thing. Most large factories in the world, after 2 to 3 years of operation, reach a stable point. Construction done, workforce filled, running at a steady rhythm. Giga Texas never did. The cranes never left the construction site. The machinery never stopped. And the question is, what is Tesla building that it cannot stop? When the press covered Terafab, they called it Tesla's chip factory. Simple, clean, easy to understand. But that is not the full truth. And that very act of simplification has caused many people to completely misunderstand the scale of what is happening. Terafab is not one chip factory. Terafab is two chip factories placed side by side on the same campus with two entirely independent missions. The first factory specializes in producing chips for Tesla's ground-based ecosystem. FSD chips for self-driving vehicles, control chips for the cybercab, the robo-taxi model Tesla is deploying in Houston and Austin. And most critically, AI chips for Optimus, which will directly determine the thinking and response capabilities of every robot during every second of operation. The second factory serves a completely different objective. It produces high-performance chips for SpaceX, specifically for the next generation of Starlink satellites. Not Starlink as we know it today, but a generation of satellites far smarter, processing far more data, and requiring chips capable of withstanding cosmic radiation. Something no conventional supplier on the market can currently deliver. In addition, XAI, the AI company recently acquired by SpaceX, will also use this factory's output to run the infrastructure for Grok. Why split into two factories instead of combining them into one? Because chips for robots and chips for satellites have entirely different technical requirements. Chips for Optimus must be optimized for real-time processing speed in uncontrolled environments, where the robot must make decisions in a split second. Chips for Starlink must be optimized for durability and the ability to withstand the harsh conditions of outer space. A shared production line is not an option, and Tesla understood that from day one. The floor area dedicated exclusively to the chip R&D and manufacturing zone, 2 million square feet for chips alone. Not counting the vehicle factory, not the battery factory, not the robot factory. And the first phase will begin with 3,000 wafers per month. A number that sounds small compared to TSMC, but it is a deliberate decision. To understand why, we need to revisit a lesson Tesla paid a very steep price to learn. 2017, Tesla was trying to produce the Model 3 at the Fremont, California factory. Elon Musk announced they would reach 5,000 vehicles per week. The plan sounded perfect, maximum automation, robots doing everything, humans only supervising. On paper, everything was logical. In reality, the production line nearly ground to a complete halt. The robots were too complex, had too many failure points, too difficult to control when running simultaneously. Tesla had to stop, dismantle part of the automation system, and bring humans back to manually handle the steps the robots could not. Elon Musk at that time slept on the floor of the Fremont factory for weeks on end. Not to create an image, but to personally deal with every problem that arose each day. That was the most costly lesson in Tesla's history. And this time, with Optimus and Terafab, they are not repeating it. Instead of pushing everything straight into Texas at massive scale right away, Tesla chose a different path. In Q2 2026, they began converting the entire Model S and Model X production lines at Fremont into Optimus robot assembly lines. With a designed capacity of 1 million robots per year, Fremont will be where Tesla learns how to assemble robots at industrial scale, learns how to detect defects, learns how to fine-tune processes, learns how to train workers to operate alongside robots, before multiplying all of that by 10 in Texas. And the early results have already shown clear signals. As of January 2026, Tesla had approximately 300 Optimus Gen 3 robots operating in actual production environments inside its factories. Not exhibitions, not press demos. These robots are genuinely working and collecting real data from real manufacturing environments. Every hour an Optimus robot operates inside the factory is an hour of invaluable data recorded. Data on how the robot grasps components, data on how it reacts when encountering unexpected situations, data on which points the robot most frequently makes errors. All of that data will be used to improve the next generation. And that is precisely why starting small, learning thoroughly, and only then scaling up is a far wiser decision than charging straight into million robot production from day one. This is not marketing thinking. This is pragmatic engineering thinking. And that distinction matters enormously when you are trying to build something nobody has ever done before. In April 2026, a specific type of equipment appeared at the North Campus of Giga Texas. Not a standard crane, not a standard excavator. It was Geo Pier equipment, a specialized ground reinforcement system that Tesla had previously used during earlier expansion phases of this same factory. Geo Pier operates on a distinctive principle, drilling deep into the earth, then compressing rock and hard material into vertical columns, creating a foundation with a degree of rigidity and stability far beyond what conventional ground can provide. The cost is significantly higher than standard concrete placement. The construction time is longer. And Tesla still chose it because there was no other option. The reason lies in a reality many people do not recognize. A humanoid robot assembly line has entirely different technical requirements from a vehicle assembly line. When assembling a car, deviations of a few millimeters at many stages are still acceptable. There are inspection and adjustment steps afterward. But the hands of an Optimus robot are designed with 22 degrees of freedom, with dexterity approaching that of human fingers. To assemble those components to the required precision, the equipment on the production line must operate in an environment free from vibration, settling, or misalignment, even by a fraction of a millimeter. If the floor surface vibrates, even from something as minor as a truck passing outside, or from machinery operating elsewhere within the factory, errors accumulate across each component. And the end result is robots failing in large numbers before they ever leave the assembly line. At a target of millions of robots per year, even a small error rate translates to tens of thousands of scrapped units and billions of dollars wasted. That is why Geo Pier is there. And that is also why the total construction cost for this expansion phase alone, according to filings submitted to Travis County, falls between $5 billion and $10 billion for infrastructure alone. Total expanded floor space, 5.2 million square feet, equivalent to 90 football fields placed side by side. Target completion for the main structural framework, end of 2026. But Tesla had previously built the entirety of Giga Texas from bare ground to an operational factory in under 2 years, and nobody believed that until it happened. Of everything being built at Giga Texas, there is one thing that I think gets talked about the least, yet is the smartest part of the entire overall design. In April 2026, Tesla activated the first phase of Cortex 2.0, an AI supercomputer cluster placed directly at Giga Texas. Scale, 230,000 H100 equivalent GPUs. Power consumption at full capacity, 500 megawatts. That figure of 500 megawatts, Tesla is consuming enough electricity to power a city of 400,000 people. Not to manufacture vehicles, not to operate robots, but to teach robots how to exist in the world of human beings. But here is the detail that is truly worth noting, and that few people pay attention to. Cortex 2.0 is not located in California. It is not located in Nevada, where many of Tesla's larger data centers are. It is placed right at Giga Texas, just a few hundred meters from the Optimus factory. That physical distance is not accidental. It creates an operational loop that no robot manufacturer in the world today is capable of replicating. 300 Optimus robots working on the factory floor, collecting real data on how they move, grasp, and respond to their surrounding environment. That data is fed directly into Cortex 2.0 for training. The updated AI model is pushed back to the robots. The robots perform better. They collect more data. And that loop continues nearly in real time all on the same campus. No need to transport data across thousands of kilometers, no latency, no information lost in transmission. This is an intentional design from day one, not a coincidence. And when you place Cortex 2.0 alongside Terafab, alongside the Optimus factory, alongside the vehicle production lines, you begin to see a larger picture taking shape. Giga Texas is no longer a collection of separate factories. It is becoming a closed-loop ecosystem where everything feeds and nurtures everything else, growing together on the same piece of land. When Terafab was announced with a $25 billion budget, many people thought that was a massive number. And it is massive by normal standards. But Bernstein Research, one of the most respected financial analysis firms on Wall Street, ran the numbers and arrived at a different figure. For Terafab to reach its ultimate goal of 1 terawatt of computing capacity per year, the actual estimated cost will fall somewhere between $5 trillion and $13 trillion, not billion, trillion. The gap between the current budget and the final target is approximately 200 times over. But to be fair, $25 billion dollars the budget for the prototype phase, not the entire project. Elon Musk has also made this clear. Giga Texas is only the starting point. A fully realized terrafab will require an entirely new location at a scale approaching 100 million square feet, approximately 12 times the size of Samsung's chip plant in Texas. TSMC took 30 years and hundreds of billions of dollars in continuous investment to reach its current position. Nobody builds the world's largest chip factory in a single phase. The real question is not whether 25 billion dollars is enough, but whether Tesla has sufficient resources to go the long distance. This is the technical challenge that no amount of billions of dollars can resolve immediately. To manufacture 2 nanometer chips, EUV machines are absolutely required. Extreme ultraviolet lithography, equipment that uses extreme ultraviolet light to etch ultrafine circuits onto silicon wafers. Globally, only one company manufactures these machines, ASML of the Netherlands. A single EUV machine costs approximately 200 million dollars, and ASML's current order backlog is fully booked through the end of 2027. Even if Tesla has the money, the land, and the engineers, they still have to stand in line and wait. There is no way around this constraint through speed or ingenuity. This is a limitation embedded in the global supply chain, and nobody can change it in the short term. Giga Texas currently consumes 556 million gallons of water per year, a figure that has increased 60% in just 2 years, making Gigafactory is the third largest water customer of the Austin water system. A standard chip factory requires an additional 365 to 700 million gallons of water per year, added on top of an already very large baseline. The problem is that Austin sits in a climate zone prone to drought. Texas has experienced multiple severe drought episodes over the past decade. And local environmental groups have formally written to the Travis County Commissioners Court demanding legally binding mechanisms should Tesla fail to meet its environmental commitments. This is not a problem that can be solved by money or technology in the short term. This is a natural constraint, and it may become the most unexpected and most unpredictable bottleneck of the entire project. In 2020, Apple announced it would drop Intel chips, a partnership of more than 15 years, and shift to producing its own chips under the name M1. The reaction from the tech industry at the time? Skepticism. Many questioned why Apple would abandon Intel, the world's largest chip maker at the time. The risk was too great, the cost too high. 6 months after the M1 launched, the MacBook Air delivered more than double the performance of its Intel predecessor while consuming only half the power. Battery life extended from 8 hours to 18 hours, price unchanged. Apple never looked back. The reason the M1 succeeded is that Apple designed the chip for exactly one purpose, running MacBooks and iPhones, not designed for hundreds of different customers as Intel and Qualcomm must be. Every transistor, every circuit was optimized for the Apple ecosystem and only Apple. Tesla is betting Terafab will do the same, but at a scale many times larger. Before Terafab, Tesla bought chips from Samsung and TSMC, paid market prices, depended on others' delivery schedules, and had no ability to intervene in chip design at the fabrication level. With Terafab, for the first time, Tesla can optimize its chips all the way down to the silicon level. Chips for Optimus will be designed and manufactured solely to run Optimus. No overhead, no redundant features, no unnecessary cost. Tesla estimates that in-house chip production could reduce semiconductor component costs by 40% and when that figure is multiplied against a target of millions of robots per year, that is tens of billions of dollars in savings annually. This is not ambition built for a good story. This is a pure business case. Gigafactories in 2022 was a car factory. In 2023, it added the 4680 battery line. In 2024, it added the Cybertruck. From 2026 to 2027, adding an Optimus robot factory, a 2-nanometer AI chip factory, and an AI supercomputer consuming 500 megawatts of electricity. No factory anywhere in the world does all of these things at the same location. Apple designs chips, but contracts TSMC to manufacture them. Nvidia designs chips, but does not make robots. Google trains AI, but does not produce hardware at this scale. Gigafactory in 2026 is the only place in the world where chip design, chip manufacturing, AI training, robot assembly, and vehicle manufacturing all take place at the same location within the same timeline. There is no official name for this model. There is no precedent to reference. There is no textbook written on how to operate something like this. Numbers like 10 million robots, 1 terawatt of computing, 25 billion dollars. They are not PowerPoint slides in a presentation. They are legal filings submitted to Travis County. They are Geo Pier equipment on the construction site. They are 230,000 GPUs consuming 500 megawatts of electricity. They are Intel, a 50-year-old company, putting its name on this project because they believe it has a genuine technical foundation. Those things are not marketing. Those things are evidence. But evidence that the project is happening, not evidence that it will succeed on schedule. EUV is still on a waiting list. Water in Austin is still an unanswered question. The gap from 25 billion dollars to 5 trillion dollars is still an extraordinarily long journey. The history of technology shows one thing repeating itself. The most doubted projects are often the ones that ultimately change everything. But that same history shows that not every bold ambition ends in success. And the distance between those two outcomes usually lives in the small details that nobody paid attention to from the start. 25 billion, 2 nanometers, 500 MW. And all of it is happening right now. I started the Tech Revolution channel not to talk to experts, but so that even the most ordinary people can understand the things shaping the world. If this video helped you see more clearly what is happening, share it with someone who should also know this. If there is any point in my analysis you feel was not thorough enough, say so directly in the comments. I read every single one. Like, subscribe, and see you in the next video.