Elon Musk’s TERAFAB Could Reach $119 Billion — And It’s Just Getting Started

Kind: captions Language: en In Grimes County, Texas, Elon Musk is planning a factory unlike anything Tesla or SpaceX has built before. The first announced phase alone represents about 16.8 billion in planned investment, while the full Terraab concept calls for more than 100 million square ft of manufacturing space, bringing advanced logic chips, memory, packaging, testing, and even mask production into one enormous semiconductor complex. If 100 million square f feet were spread across a single level, it would equal roughly 3.6 square miles of factory floor. But the most unusual part is not simply the size. On October 7th, 2026, Musk clarified that Tesla and SpaceX themselves intend to build and operate the facility. That clarification changes the story completely. Tesla does not just want to design another AI chip and send the drawings to an established foundry. SpaceX does not simply want to keep ordering more processors for increasingly powerful computing systems. Musk wants companies known for electric cars, humanoid robots, reusable rockets, satellites, and communications networks to move into one of the most difficult manufacturing industries on Earth and take direct responsibility for making advanced silicon. TSMC, the world's largest dedicated chip foundry, could potentially subleasase part of Terrafab. and Intel remains involved, but Musk says neither company is supposed to simply take over the project. Tesla and SpaceX are meant to remain at the center of it. That makes Terrafab much more than another supplier agreement. It is a bet that AI chips are becoming too important for Musk's companies to treat them like ordinary purchased components. Tesla needs increasingly powerful inference processors for self-driving vehicles and Optimus robots. SpaceX is expanding computing infrastructure of its own, while its longerterm plans include high performance chips designed for space-based data centers. Together, the two companies say their future requirement could exceed 1 terowatt of compute, far beyond what they believe existing and currently planned semiconductor capacity can comfortably provide. That number is a forecast of future computing hardware demand. Not a measurement of electricity being consumed by Terapab today, but it explains the logic behind the project. For most people, a chip looks almost insignificant. A finished processor can fit in your hand. Yet, creating it requires an industrial system so complicated that only a small number of companies can manufacture the most advanced versions at high volume. A leading edge chip begins with an extremely pure silicon wafer. Across that wafer, the factory repeatedly adds materials, removes materials, changes their electrical properties, projects microscopic patterns, builds transistors, connects them with layers of metal, measures defects, and checks whether every stage remained inside incredibly narrow tolerances. This can happen hundreds or even thousands of times before the wafer leaves the front end of the fab. Then the finished dyes still need to be cut, packaged, connected to memory and other components, tested, and prepared for installation inside the machine that actually uses them. For an AI processor, that final packaging stage has become increasingly important because computing performance is no longer determined by one piece of silicon alone. High bandwidth memory, chip-to-chip connections, power delivery, cooling, and packaging can all become part of the performance equation. Terraab is designed around that larger system. SpaceX says the project is intended to place logic manufacturing, memory, advanced packaging, and testing inside one vertically integrated complex. Tesla job postings go even further, describing plans for lithography, mask production, and work across three distinct chip families, edge inference processors, space hardened data center AI chips, and high bandwidth memory. Other current Terrafab roles mention two nanometerclass process technology using gate all-around transistor architectures. These job descriptions do not prove those technologies are already in mass production but they show the technical direction Tesla is staffing for today. This is the first important reality check. Terrafab is planned infrastructure not a completed semiconductor factory. The Grimes County location has been announced. The first phase investment has been identified. Hiring is underway for process equipment integration and infrastructure roles. A smaller research fab is being developed at Gigafactory Texas. Intel is involved. Semiconductor equipment suppliers have been approached. But none of this means Terrafab is already producing advanced AI chips at scale. Reuters reported earlier in 2026 that Musk's team was targeting the start of silicon manufacturing around 2029, followed by a production ramp. A target is not the same thing as a finished factory. And getting from an empty industrial site to working silicon is where the story becomes much harder. Tesla understands mass manufacturing. SpaceX understands mass manufacturing too, though in a very different environment. Tesla has spent years redesigning vehicles and factories around large castings, battery packs, power electronics, robotics, and increasingly automated production. SpaceX brought rocket engines, structures, avionics, satellites, terminals, and many other systems deep inside its own manufacturing organization. Both companies have repeatedly used vertical integration to reduce dependence on suppliers and shorten the feedback loop between design and production. Semiconductor manufacturing is different because the product is created almost atom by atom. A car factory can tolerate dimensions measured in fractions of a millimeter across many components. A leading edge semiconductor process deals with structures measured in nanometers. Contamination invisible to a human eye can destroy a device. A microscopic process variation can reduce yield across a wafer. Equipment may cost tens or hundreds of millions of dollars. Yet owning the machine does not automatically mean knowing how to use it well enough for profitable production. This is the hidden bottleneck inside every advanced fab yield. Imagine a wafer carrying hundreds of expensive processor dyes. If most of those dyes work, the factory can spread the cost of the wafer across many useful chips. If a large percentage fail, every successful chip becomes much more expensive. A process can technically produce advanced silicon and still be commercially weak because too many chips are defective, too slow, too power hungry, or unable to meet reliability specifications. Yield is not solved with one invention. It is learned. Engineers study defects, tune deposition conditions, adjust etch steps, control materials, improve masks, optimize lithography, inspect wafers, and slowly discover which tiny variations produce failures later in the process. Then they repeat the process again and again until a technology becomes stable enough for high volume manufacturing. That knowledge is one reason semiconductor experience is so difficult to reproduce quickly. TSMC did not become dominant simply because it bought the right lithography machines. It built a manufacturing system capable of producing designs for companies such as Apple, Nvidia, AMD and many others while maintaining high volume reliability across generations of process technology. Intel has spent roughly six decades developing semiconductor manufacturing expertise. Even though it has also experienced painful process delays of its own, Samsung combines memory leadership with advanced logic manufacturing. Yet, even Samsung faces the extraordinary difficulty of competing for leading edge foundry customers. Tesla and SpaceX are entering that world from the outside. Musk's answer appears to be that they do not have to learn everything alone. Intel publicly confirmed in its first quarter 2026 materials that it had joined SpaceX, XAI and Tesla on Terapab. Intel said the companies were exploring ways to improve the economics and efficiency of semiconductor manufacturing while Intel was also developing its future 14A process and engaging outside customers. This gives Terrafab access to a company that already knows how to design, fabricate, package, and qualify advanced processors at industrial scale. But Intel is not expected to simply own and run the entire complex. Musk's October clarification was explicit. Tesla and SpaceX plan to build and operate Terraab themselves. TSMC could potentially lease part of it while Intel remains part of the project. That creates a very unusual model. Instead of one foundry operating a factory for unrelated customers, Terrafab could become a large industrial campus controlled by its biggest future chip consumers, while experienced semiconductor companies participate inside or alongside the broader system. There is already evidence that Musk's companies are trying to assemble the supplier network required to make that possible. Earlier this year, people working on Terrafab contacted semiconductor equipment companies including Applied Materials, Tokyo Electron, and Lamb Research, while also seeking support from Samsung. The requests covered the kinds of tools and materials a real fab needs: etchers, deposition systems, cleaning equipment, substrates, photo masks, and testing tools. Reuters reported that the team was pushing suppliers for rapid pricing and delivery information as part of the effort to reach silicon production around 2029. That supplier list reveals something important. Vertical integration does not mean independence from the semiconductor ecosystem. Even TSMC depends on equipment and materials from specialist companies around the world. Extreme ultraviolet lithography depends on ASML. Wafer processing tools come from companies such as Applied Materials, Lamb Research, and Tokyo Electron. Electronic design software comes from another specialized industry. Chemicals, gases, wafers, masks, metrology systems, pumps, valves, filters, and sensors all come from highly specialized suppliers. Terraab can bring more manufacturing under Tesla and SpaceX control, but it cannot recreate the global semiconductor supply chain inside one Texas building. The realistic goal is different. Own more of the strategic process, reduce some external bottlenecks and connect chip design more directly to the products that will consume the chips. That last part may be the most important potential advantage. Tesla has been designing its own AI processors for years because the computer inside a self-driving vehicle faces very different constraints from a generalpurpose cloud server. The vehicle needs high inference performance, but it also has strict limits on power, heat, physical space, and cost. Optimus creates another edge computing problem. A humanoid robot has to interpret cameras and other sensors, make decisions, and control motors in real time without depending completely on a distant data center. Tesla said in April 2026 that it had completed the final design of its next generation AI5 inference processor. The company also described the Terapab effort as a response to anticipated chip demand greater than existing and planned industry capacity, beginning with the Tesla owned research fab at Gigafactory Texas. That smaller facility matters because it gives Tesla somewhere to develop semiconductor processes before attempting the much larger Grimes County buildout. SpaceX creates a different set of requirements. A chip intended for a terrestrial AI data center can consume significant power because the facility can provide huge electrical and cooling systems. A processor intended for orbit faces a completely different environment. Every watt matters. Every kilogram matters. Heat has to be rejected in vacuum. Radiation can corrupt electronics. Reliability becomes critical because a technician cannot simply walk into a satellite and replace a failed processor. Current Terraab job postings explicitly mention space hardened data center AI chips for orbital satellites alongside edge processors and high bandwidth memory. That does not mean SpaceX has already deployed a giant orbital data center. And it should not be described that way. It means Terrafab's planned product portfolio is being designed around several radically different computing environments from the beginning. This creates a possible feedback loop that helps explain why Musk wants the FAB so close to the product companies. Imagine an Optimus engineering team discovers that a certain neural network workload is limited by memory bandwidth. Today, the chip team can redesign future silicon, but fabrication remains separated by external manufacturing schedules and commercial foundry relationships in a more vertically integrated system. Chip designers, process engineers, packaging teams, robot engineers, vehicle engineers, and data center teams could potentially work through design trade-offs much more directly. SpaceX already organizes rocket development around rapid feedback. A component is designed, manufactured, tested, broken, changed, and tested again. Tesla has used similar loops in vehicles, batteries, and manufacturing equipment. Terrafab appears to be an attempt to bring some of that behavior into silicon, but semiconductor fabrication does not move like rocket development. A rocket prototype can explode and produce valuable information in seconds. A software model can be updated overnight. A wafer may spend months moving through fabrication before engineers discover whether a design change produced the expected result. An error in a chip layout can require a new mask set. A change in one process step can affect devices many steps later. And a manufacturing shortcut that improves speed but destroys yield is not an improvement at all. That is the wall Terrafab has to cross. Rapid iteration without losing semiconductor discipline. If engineering stories like this are why you watch this channel, subscribing helps more people find them because the impressive part of Terapab is easy to see in a rendering. The part that will decide whether it works is almost invisible. Microscopic defects, process windows, equipment uptime, and whether each new wafer comes out slightly better than the one before it. There is another reason this project is different from a normal factory. Semiconductor fabs require an enormous hidden infrastructure system before the first transistor is manufactured. The clean room is only the center. Behind it are central utility plants, ultra pure water systems, chemical distribution networks, specialty gases, wastewater treatment, cooling equipment, vacuum systems, backup electricity, fire protection, material handling systems, and facilities designed for prevent contamination. Current Tesla hiring for Terraab includes infrastructure roles responsible for power generation, utility corridors, chemical and gas systems, logistics, central utility plants, and site planning stretching from the research fab in Austin to the permanent Grimes County complex. That is what building a chip factory actually means. Before Tesla can manufacture AI processors, it has to manufacture the environment in which AI processors can exist. Water is one example. Chip production needs extremely clean water for repeated washing and chemical processes. Ordinary municipal water is nowhere near pure enough for many manufacturing steps. So FABS create ultra pure water by removing dissolved minerals, particles, organic material, and other contamination to extraordinary levels. Then the site needs systems to collect and treat waste water after manufacturing. SpaceX says the Grimes County project intends to use water from Gibbons Creek reservoir rather than relying primarily on local groundwater while incorporating wastewater treatment and reuse. That planning does not guarantee environmental impacts disappear. Large semiconductor facilities still require enormous infrastructure and careful management of water, chemicals, energy, and industrial waste. But it shows how far the project extends beyond simply buying lithography machines. Electricity is another major constraint. A modern semiconductor fab runs enormous numbers of pumps, chillers, air handlers, heaters, plasma systems, deposition chambers, etchers, inspection tools, and clean room equipment continuously. Some tools cannot simply be switched off every evening like machinery in an ordinary workshop. Process stability matters and unplanned outages can destroy work already inside expensive equipment. This means terapab sits inside a broader physical AI buildout. Tesla and SpaceX need chips. Making those chips requires fabs. Fabs require power, water, chemicals, machines, factories, and thousands of trained people. The chips then go into robots, vehicles, data centers, and potentially satellites that need even more electricity and infrastructure. AI may look like software when someone types into a chatbot, but underneath the interface is an expanding industrial chain. SpaceX has described this directly in its corporate materials, arguing that future AI growth is increasingly constrained by the physical stack, semiconductor manufacturing, data center infrastructure, and power generation. Terrafab is intended to push the company's vertical integration downward into chip design and manufacturing so future shortages do not limit computing expansion. This helps explain why Musk is willing to enter an industry where even experienced companies regularly struggle. The risk of building terrafab may be enormous. But from his perspective, the risk of not building enough chip capacity could eventually be larger if Tesla, SpaceX, and associated AI projects really consume compute at the scale they expect. The present-day evidence shows how aggressive that demand already is. Reuters reported on October 6th that SpaceX was seeking roughly $40 billion in financing connected to the purchase of NVIDIA chips for its data centers. In other words, while Musk plans a future factory intended partly to reduce supply constraints, his companies are simultaneously spending enormous amounts of money buying computing hardware from the existing ecosystem. Terrafab does not replace Nvidia, TSMC, Samsung or other suppliers today. The current AI buildout still depends heavily on them. That creates a fascinating transition period. Musk is effectively trying to build tomorrow's internal supply while continuing to buy today's external supply at extraordinary scale. If Terrafab reaches high volume production, the balance could gradually shift. But even then, Tesla and SpaceX may continue using outside foundaries because chip demand can exceed one factory. Different processes suit different products and relying on multiple suppliers reduces risk. Tesla already demonstrates that approach. The company has worked with outside semiconductor partners rather than depending on one source. Terrafab can become another supply vector rather than instantly replacing every external relationship. And there is a strong argument for diversification. Imagine Terraab suffers a process delay while Tesla is preparing millions of new vehicles or robots. If external foundaries remain available, production can continue. If Tesla had moved everything inside too quickly, the internal fab could become a single point of failure. Vertical integration gives control, but concentration creates risk. There is also the question of economics. A semiconductor factory makes sense only when it stays busy. The equipment is too expensive to spend years sitting idle. A fab has to spread huge capital costs across large numbers of useful wafers. TSMC succeeds partly because customers across the technology industry continually fill its advanced production lines. Intel historically filled factories with its own processors and its foundry strategy now seeks more external customers. Terrafab's answer is supposed to be internal demand. Instead of asking whether enough unrelated chip designers will use the fab, Musk is betting that Tesla and SpaceX will consume enormous quantities themselves. Cybercabs, Optimus robots, autonomous vehicles, AI training and inference systems, communications infrastructure, and possible orbital computing could all become captive demand. That sounds powerful, but it is also one of the project's largest risks. Future demand is not guaranteed simply because it appears in a forecast. Optimus production could grow more slowly than expected. Autonomous driving deployment could face technical or regulatory delays. Space-based AI infrastructure could remain much smaller than Musk imagines. More efficient chips could reduce the number of processors required for certain workloads. External foundaries could expand faster than expected. If any of those things happen, a gigantic internal fab could have more capacity than the companies actually need. Remember the size discussed earlier, more than 100 million square ft planned. Scale is useful only when there is enough work to fill it. Otherwise, scale becomes overhead. This is where the phased structure matters. The $16.8 billion figure refers to the announced first phase, while additional future phases could take investment much higher. Public incentive filings cited by regional development organizations have described a potential four-phase project reaching as much as roughly $119 billion in combined investment. But that is a possible long-term buildout, not money already spent and not a guarantee that every phase will happen. The distinction between firstphase commitment and ultimate project potential is crucial. That phased approach gives Tesla and SpaceX a way to learn before committing the entire envision scale. Build initial capacity, install equipment, develop processes, improve yield, see how rapidly internal demand grows, then decide how aggressively to expand. The research fab in Austin may serve a similar purpose at a smaller scale. SpaceX says Tesla broke ground on the research fab in April 2026 and describes it as the precursor to Terraab. If that facility can develop process knowledge, train engineers, and reveal problems before the Grimes County plant reaches full scale, mistakes become less expensive. This idea has a precedent inside Tesla. Manufacturing development rarely goes directly from a laboratory idea to perfect mass production. Battery cells, drive units, castings, and vehicle platforms all pass through pilot phases where processes are discovered, modified, and sometimes abandoned. Semiconductor manufacturing is the same principle amplified dramatically. The challenge is that Tesla cannot simply learn slowly forever. Semiconductor technology keeps moving. By the time one manufacturing node is mature, competitors are preparing the next. A new fab must learn while the industry itself advances. Tesla job listings suggest terapab teams are already being recruited across lithography, etch, deposition, epitexi, metal processing, implantation, polishing, measurement, integration, and high volume manufacturing readiness. That is the vocabulary of a real semiconductor process organization rather than a conceptual chip project. Still, hiring engineers is proof of preparation, not proof of production. The most important validation will come later when those teams begin moving wafers through actual tools. Then the yield curve starts. At first, perhaps many chips fail. Engineers identify a deposition defect. They adjust the process. Another problem appears during lithography. They solve it and discover an issue in metal interconnects. Then packaging exposes a thermal problem. Test data goes back into design. Each solution raises the percentage of useful chips. That is the slow, unglamorous process behind a successful fab. The public will see an enormous building. Engineers will see thousands of variables, and this is where Intel's continued involvement may matter most. In its April earnings call, Intel said its 14A development was progressing and that customer feedback was helping shape the technology. Intel also explicitly named its partnership with SpaceX, XAI, and Tesla around Terraab. A company trying to learn chipmaking from scratch would have a mountain to climb. A company trying to learn while working with one of America's most experienced semiconductor manufacturers has a very different starting point. TSMC's possible role creates another intriguing possibility. Musk has said TSMC will not operate the entire complex but may subleasase part of it. If that happens, Terrafab could place one of the world's most experienced foundry companies physically inside a campus controlled by Tesla and SpaceX. That arrangement is only a possibility today, not a finalized operating structure, but it illustrates how unconventional the project could become. There is a larger industrial story here too. Advanced semiconductor production remains geographically concentrated, particularly in East Asia. The United States has been spending heavily to attract new fabs from TSMC, Samsung, Intel, Micron, and others. Terrafab would add a different model to that trend. Not an established chipmaker expanding American capacity, but major chip customers attempting to become manufacturers themselves. That is a much more radical form of supply chain security. Instead of asking which foundry can guarantee enough wafers for us, the question becomes why don't we own part of the wafer capacity. Tesla has used that logic before. It built charging infrastructure when the public charging network was not ready. It moved deeper into battery manufacturing when cell supply became strategic. It built a lithium refinery in Texas to move farther upstream into battery materials. SpaceX built its own engines because buying the propulsion system it needed was not practical. Starlink required SpaceX to manufacture satellites and user terminals at enormous scale rather than depend entirely on traditional aerospace suppliers. Terraab pushes that same instinct to perhaps its most difficult end point. But there is a difference between building what suppliers cannot provide and trying to outperform suppliers who are already extremely good at their job. TSMC exists specifically to manufacture chips. Samsung and Intel have spent decades developing fabrication technology. Terrafab therefore needs to offer more than independence. It needs some combination of better economics, faster iteration, more secure capacity or chips unusually optimized for Musk's own hardware. Otherwise, buying from specialists remains easier. This is why the project's eventual success should not be judged by whether the building is completed. A semiconductor factory can be physically finished and still fail commercially. The real measurements are yield, cost per useful chip, performance, power efficiency, production volume, equipment uptime, schedule, and the rate at which one chip generation can replace the last. Those measurements are still in the future. For now, Terrafab has passed from an abstract Musk statement into a more concrete industrial program. A location has been selected in Grimes County. Texas has announced the first phase investment and 3,000 expected jobs. The research fab is under development in Austin. Intel has publicly confirmed its involvement. Tesla is hiring semiconductor process and infrastructure specialists. Equipment suppliers have been approached. Musk has clarified that Tesla and SpaceX intend to operate the complex themselves. What has not happened is equally important. Terrafab has not started high volume silicon manufacturing. Its future 100 million square ft buildout is not complete. TSMC has not been announced as the operator. The 1 terowatt compute requirement is a future demand forecast, not present production. The potential long-term investment figures are not all committed capital today. And the 2029 silicon date reported earlier this year remains a target. That reality does not weaken the story. It makes the real story clearer. Musk is attempting to apply Tesla and SpaceX's vertical integration philosophy to an industry where the manufacturing barriers are measured at nanometer scale. He is betting that AI growth will make chips strategic enough to justify billions of dollars in factories, a new semiconductor workforce, partnerships with companies like Intel, relationships with global equipment suppliers, and potentially manufacturing arrangements with TSMC. If the bet works, the payoff would be unusual. A Tesla engineer could eventually design an AI processor for a robot, hand it into an internal semiconductor ecosystem, manufacture it, package it alongside internally produced high bandwidth memory, test it, put it into Optimus, collect real world performance data, and send that information back into the next chip generation. SpaceX could follow a parallel loop for AI hardware built around the constraints of large computing systems and future orbital platforms. The advantage would not simply be owning a fab. It would be owning more of the feedback loop. And remember the scale from the beginning. More than 100 million square ft are planned because Musk is not talking about a boutique chip operation producing experimental wafers. The stated ambition is industrial scale large enough to address a future compute requirement Tesla and SpaceX believe existing supply cannot satisfy. That is also why the project could fail spectacularly if the assumptions underneath it are wrong. Chip demand might not grow as predicted. Yields might take too long to improve. Costs could rise. Advanced equipment could arrive late. The process technology might lag competitors. Recruiting enough semiconductor talent could become difficult. TSMC, Intel, Samsung, and other foundaries could expand so aggressively that outside supply becomes easier and cheaper than expected. Or the opposite could happen. AI demand could continue rising faster than chip capacity, making guaranteed internal supply extraordinarily valuable. Optimus and autonomous vehicles could require hundreds of millions of edge processors. SpaceX's computing operations could become enormous. Advanced packaging and memory could remain major industry bottlenecks. Under that scenario, the decision to start building manufacturing capacity years earlier could look very different. Nobody knows which future arrives. That is why Terraab should be understood as an industrial wager rather than an inevitable revolution. The project is not impressive because Tesla and SpaceX have already proven they can manufacture leading edge semiconductors. They have not. It is impressive because two companies outside the traditional foundry industry are preparing to test whether the methods that helped them integrate cars, rockets, batteries, satellites, and software can extend into a field where microscopic manufacturing precision determines everything. Cars made Tesla famous. Rockets made SpaceX famous. But both increasingly run on the same invisible resource, computation. A self-driving car is partly a computer on wheels. Optimus is a computer inside a moving body. Starlink is a network of computers in orbit. AI data centers are industrial systems built around computing hardware. As those machines become more intelligent, chips stop looking like ordinary components and start looking like the foundation beneath the entire product strategy. That is the strategic idea behind Terapab. Musk does not merely want more processors. He wants a direct path from silicon to machine. For Tesla, that machine might be a cyber cab navigating through a city or an Optimus robot manipulating objects inside a factory. For SpaceX, it might eventually be computing infrastructure operating on Earth or in orbit. Different machines, different environments, but the same underlying dependence on enormous volumes of increasingly specialized silicon. The question is whether Tesla and SpaceX can turn that dependence into manufacturing capability. By the end of this decade, we may begin to get the first meaningful answer. If the current target holds, silicon manufacturing could begin around 2029. Then the less dramatic but far more important milestones start. usable wafers, improving yield, validated packaging, qualified chips, higher volume, and actual processors entering real Tesla and SpaceX products. That is when Terrafab stops being judged as a construction project and starts being judged as a semiconductor factory. Until then, the cleanest description is also the simplest. Terrafab is a huge planned Texas chip complex being developed by Tesla and SpaceX with a 16.8 8 billion first phase, Intel involvement, potential participation from TSMC, and an ambition to vertically integrate logic, memory, packaging, and testing at extraordinary scale. It is real as a project, but its manufacturing promise remains to be proven, and that makes the next image worth watching for surprisingly ordinary. Not Elon Musk standing beside a rendering. Not another giant investment number. Not another announcement about a partner, a silicon wafer moving through the factory. Because if Tesla and SpaceX eventually learn to turn that wafer into high yield, economically competitive AI chips at the scale they are planning, Terraab will represent something much larger than one more semiconductor plant. It will mean Musk's vertical integration strategy has moved all the way from finished machines to the microscopic devices that give those machines intelligence. If these deep dives into the factories behind future technology are useful, subscribe for the next one. And if you were running Tesla or SpaceX, would you spend billions building your own chip factories for long-term control, or keep relying on specialists like TSMC, Samsung, and Intel and focus your money on the products that use the chips?