Musk's Terafab Is 10x Gigafactory Texas! 100M Sq Ft, 1TW AI Chips & $55B Texas Plan Confirmed

Kind: captions Language: en Elon Musk stood on a stage in Austin and said something that landed harder than anyone expected. >> We either build the terrafactory or we don't have the chips. And uh we need the chips, so we're going to build the terrafactory. >> Either the terrafactory gets built or the chips simply will not exist. That was the whole argument stripped down to a single sentence. And it carried a weight that most people in the room understood immediately. The world is running short on advanced semiconductors and the shortage is not the kind that fixes itself with a few extra shifts at an existing plant. It is structural. So, the proposed answer is not a bigger factory in the ordinary sense of the word. It is a facility so large that the usual comparison stop being useful. The number attached to the project is roughly 100 million square feet. That figure is difficult to picture, so it helps to place it next to something familiar. Gigafactory Texas, already one of the largest buildings on the planet by footprint, would fit inside the proposed terrafactory about 10 times over. 10 gigafactories sitting side by side operating as a single coordinated industrial site. Nothing in the semiconductor world has ever been attempted at that footprint. And very little in any other industry has either. The scale alone is what makes the project either visionary or reckless, depending on who is describing it. Then comes the production target, which is stranger still. The stated goal is 1 terawatt of AI computing hardware manufactured every year. Not sold over a decade. Not accumulated across a product line. Produced annually as an ongoing rate of output. Musk himself offered the comparison that makes the number sting, pointing out that the entire United States generates something in the neighborhood of half a terawatt of electricity. So, the factory would aim to build each year computing hardware measured against a scale that currently describes the power output of an entire country. That comparison is worth sitting with for a moment because it reveals what the project is really about. Chips are usually counted in units, in wafers, in nanometers. Measuring them in terawatts reframes them as infrastructure, closer to a power grid than to a consumer product. It implies a world where computing capacity is consumed the way electricity is consumed, continuously and at enormous volume by machines that never stop processing. Building toward that assumption requires a manufacturing base that does not exist yet anywhere on Earth, which is precisely the gap the Terafab is meant to fill. Whether the full terawatt target is ever reached remains an open question, and the honest answer is that nobody knows. Semiconductor fabrication is unforgiving. Yields take years to stabilize, equipment lead times stretch across multiple years, and every stage of the process demands precision at a scale that punishes shortcuts. Missing the target by a wide margin is a realistic outcome. But, the arithmetic works in an unusual way here because even a fraction of the stated goal would still produce one of the largest semiconductor manufacturing complexes ever constructed. That is the part that changes how the announcement should be read. A A that fails at 90% of its ambition and still ranks among the biggest industrial facilities in history is not a normal project. It sets a floor high enough that the downside scenario would reshape global chip supply on its own. The site in Texas is being planned around that logic. With capacity sized for demand that has not fully materialized yet. Aimed at customers and applications that are still being designed. The question is no longer whether the ambition sounds excessive. It is whether the rest of the industry is quietly preparing for it. The preparation showed up in the least glamorous place imaginable. A quarterly earnings call. During ASML's Q2 2026 results discussion, Chief Financial Officer Roger Dassen confirmed that the company's capacity expansion planning for 2027 and 2028 already accounts for demand coming from the Terafab in Texas. That sentence passed quickly and without drama, the way most things do on an earnings call. But it carried an implication that most listeners did not immediately register. A factory that has barely broken ground is already occupying space in the production schedule of the most critical supplier in the semiconductor world. The expansion itself is substantial. ASML announced plans to grow its manufacturing capacity for both EUV and DUV lithography systems by roughly 30% per year. Measured against its 2026 baseline of about 65 EUV machines and 130 DUV machines annually. Those numbers sound modest until the price and complexity of each unit enter the picture. A single EUV machine costs in the hundreds of millions of dollars, contains hundreds of thousands of components, and takes years to build, test, install, and calibrate. Increasing that output by 30% annually is an enormous industrial undertaking on its own. More revealing is how much of that future output is already spoken for. ASML disclosed that its 2027 EUV production is nearly fully booked with a meaningful portion of 2028 also reserved. Customers are committing billions of dollars to machines that will not be delivered for years, which tells you something about how the industry views the demand curve ahead. Nobody makes that kind of financial commitment on a hunch. They make it because they have run the numbers on what AI computing will require, and concluded that waiting is the more expensive option. The reason this matters comes down to a single uncomfortable fact about the semiconductor supply chain. ASML is the only company on Earth capable of manufacturing EUV lithography systems, the machines required to produce the most advanced AI processors in existence. There is no second supplier, no alternative vendor, no competing technology ready to substitute. Nvidia depends on it. TSMC depends on it. Intel and Samsung depend on it. Every advanced chip powering every AI system in use today traces back through equipment built in the Netherlands by one company. That monopoly position changes what the TeraFab mention actually means. ASML has no shortage of customers and no incentive to build capacity for a project it considers speculative. Its order book is already full. Its machines are already allocated, and its engineering resources are already stretched. Naming a factory in its forward planning is not a courtesy or a marketing gesture. It is a resource allocation decision made by a company that cannot afford to waste capacity on ambitions that will not materialize into orders. So, the calculation running underneath that earnings call is fairly direct. The only firm capable of supplying the machines at the top of the entire computing pyramid has looked at the Terafab, assessed it against every other claim on its production line, and decided it belongs in the schedule. That does not confirm the factory will hit its targets, and ASML made no such claim. What it does confirm is that the project has crossed from concept into the category of industrial demand that suppliers plan around, which raises an obvious question about what all those chips are supposed to be doing. The answer starts in orbit with a scenario that sounds like fiction until the hardware requirements are broken down. SpaceX operates thousands of Starlink satellites today, and the constellation keeps growing. Push that trajectory far enough forward, and a figure that gets discussed often among Musk's supporters appears. 1 million AI-enabled satellites circling the planet. It is a hypothetical and an aggressive one, but hypotheticals have a way of clarifying what the underlying math actually demands. Every one of those satellites would need onboard processing power, and not the modest kind. The workload each satellite would carry explains why. Autonomous routing decisions, so traffic finds the fastest path without waiting for instructions from the ground. Beam management, aiming signals precisely at moving users below. Collision avoidance in an increasingly crowded orbital environment. Laser communication links between satellites. Cyber security running continuously against threats that never sleep. Earth observation and edge inference performed on the spot rather than shipped somewhere else. Each of those functions is a computing problem. And running them simultaneously in real time requires processors far beyond what older spacecraft carried. That is the break from how satellites have traditionally worked. A conventional communication satellite is essentially a mirror in space. A signal arrives from the ground, gets amplified, and gets sent back down. The satellite makes almost no decisions and understands almost nothing about the data passing through it. All the intelligence lives in ground stations, and the hardware in orbit stays deliberately simple because simple hardware survives longer in an environment where repairs are impossible. An AI satellite inverts that arrangement completely. It processes data where the data is generated in orbit continuously without waiting for a round trip to Earth and back. That shift is what makes the constellation useful in ways the old design never could be. And it is also what makes the semiconductor demand explode. Imagine each next-generation satellite requiring just two advanced AI processors during manufacturing, plus a handful of replacement chips across its operational lifetime. Multiply that modest assumption across a million units, and the chip requirement climbs into the millions. The demand does not stop at the edge of the atmosphere, either. Satellites generating and processing that much data feed directly into ground infrastructure that also needs to expand. AI models have to be trained somewhere, and training happens in data centers packed with accelerators. Communication gateways handle traffic flowing down from orbit. User requests get processed. Orbital traffic gets coordinated across an entire constellation. Every one of those functions consumes computing hardware on Earth, purchased in volume and replaced on a cycle. What emerges is a reinforcing pattern, rather than a one-time purchase. Each satellite launched increases the load on terrestrial systems, which increases the need for more processors on the ground, which increases the value of having more intelligent satellites overhead. The constellation and the data centers pull each other upward, and the chip requirement compounds at both ends. Supplying that kind of demand through outside vendors at prices set by someone else, on timelines controlled by someone else, becomes a strategic problem long before it becomes a financial one. Solving that problem is what vertical integration is actually for, and Tesla's version of it is unusually broad. Instead of buying AI processors from external foundries at whatever price and schedule the market dictates, the company would manufacture much of that hardware itself. The Terafab output would flow into Optimus robots, into full self-driving computers in vehicles, into Dojo training clusters, into energy storage systems, and eventually into space applications. One factory feeding a product portfolio that has almost nothing in common on the surface, but shares the same underlying computing requirement. The dependency being removed here is worth understanding clearly. Today, essentially, every company designing advanced chips has to queue for capacity at a small number of foundries, primarily TSMC and Samsung. Those foundries allocate their most advanced production lines according to their own priorities. And the largest customers with the deepest relationships get served first. A company designing chips for autonomous driving competes for that capacity against smartphone makers, cloud providers, and everyone else chasing the same nanometer node. Owning the fabrication changes the terms of that negotiation by removing the negotiation. Production schedules get set internally. Capacity gets allocated according to which product line needs it most. Not according to which customer a foundry values most. When a design change is needed, it moves through a process the company controls rather than a queue it waits in. That kind of control is expensive to acquire and has historically been out of reach for anyone not already operating at enormous scale, which is exactly why so few companies have attempted it. The second advantage is subtler and possibly more valuable. Chips built for general sale have to serve many customers with different needs, which means they carry capabilities that any single buyer will never use. A processor designed exclusively for one AI ecosystem can drop everything irrelevant and optimize hard for the specific workloads it will actually run. Full self-driving processes camera data in particular patterns. Optimus handles real-world manipulation with its own computational profile. Dojo trains models under a different set of constraints entirely. Designing silicon around those known workloads, rather than around a hypothetical average customer, produces meaningful gains in efficiency, power consumption, and cost per operation. Apple demonstrated the principle in consumer devices. Building processors tailored to its own software, rather than buying general-purpose parts. The same logic applied to AI hardware across vehicles, robots, and training clusters would compound across every product simultaneously. Since each improvement in the chip lifts everything running on it. There is a harder edge to this as well. Semiconductor manufacturing is one of the most punishing businesses that exists. With capital requirements measured in tens of billions, and yield problems that have humbled companies with decades of experience. Intel struggled at process nodes it once dominated. Committing to build fabrication capacity in-house means accepting those risks directly, instead of paying someone else to absorb them. The reason a company would take on that exposure has less to do with the chips themselves, than with what those chips make possible once they are running inside a product. Here is where the whole argument turns, because the chip was never meant to be the product. It is the thing that makes other products generate money repeatedly. And that distinction separates the Terafab from every conventional semiconductor business. A traditional chip maker builds a processor, ships it, collects payment, and the transaction closes. The relationship ends at the loading dock. Building silicon to feed an ecosystem you also own means the processor keeps producing income for years after it leaves the factory floor. Robo-taxis demonstrate the difference most clearly. A car sold in the ordinary way earns revenue exactly once on the day the customer signs the paperwork, and everything afterward is warranty cost and service margin. An autonomous vehicle carrying passengers earns money every day it operates, on every trip without a driver taking a share. Morgan Stanley has estimated that autonomous mobility could eventually reach a multi-trillion-dollar global market. And each vehicle in that fleet requires progressively more capable processors to read camera data, decide in real time, and improve through software updates. Optimus follows the same structure with different arithmetic. Musk has said repeatedly that humanoid robots could eventually outnumber people and grow into the company's largest business, which is a claim that invites obvious skepticism. Scale it down aggressively, and the numbers still hold up. 10 million robots at an average price of $20,000 represents roughly $200 billion in hardware revenue. And that figure counts only the initial sale of the machine itself. The hardware is where the revenue starts, rather than where it ends. Businesses running robot fleets would subscribe to capability upgrades, fleet management software, predictive maintenance, industrial applications built for specific tasks, and enterprise support contracts. Smartphones created an entire app economy that eventually dwarfed the profit on the devices themselves. A robot capable of physical work in the real world opens a comparable software market. Except the subscriptions attached to labor rather than entertainment. Underneath both businesses runs a loop that gets stronger the longer it operates. Vehicles on the road generate driving data continuously. Robots working in warehouses and factories collect manipulation data that has never existed in usable quantities before. AI models train on both streams and improve.