Elon Musk's $25B Chip Factory Should Terrify TSMC

Kind: captions Language: en Something colossal is unfolding in the Texas plains. On the grounds of a dead coal power plant about an hour outside Houston, construction is starting on what Elon Musk calls the largest and most valuable building on Earth by far. A two-nanometer chip factory built from scratch by companies that have never manufactured a single chip. It's designed to produce 1 terawatt of AI compute per year. Built on the most advanced transistor technology we have, gate all around. The entry ticket is $25 billion. The filings already point to $119 billion. And analysts at Bernstein put the full vision at something closer to $5 trillion. This move is either genius or a very expensive trap because even the best chip makers on Earth with decades of experience and hundreds of billions invested still get it wrong. So, why would they even try? And more importantly, what do they know that everyone else doesn't? Terafab. What an interesting fab at so many levels because the strange thing about it is that the deeper you go into it, the less crazy it starts to look. This can actually be the move that rebalances the global supply chain. And the strangest part of all, most of these chips aren't even meant for this planet. Subscribe to the channel and let me explain. Back in 2001, nearly 30 companies could manufacture chips at the most advanced nodes. By the late 2000s, that was down to roughly a dozen. Today, only two to three remain. TSMC, Samsung, and Intel. Everyone else looked at the cost, looked at the physics, and walked away. Even the survivors struggled to keep up because at the leading edge, this is not just about building the factory. Essentially, you need five pillars, tools, raw materials, clean room, hundreds of machines operating at the absolute edge of physics, and one invisible layer that holds it all together, the process that turns sand into a thinking machine. Now, inside this system, some tools define everything. The most critical are EUV lithography machines, the machines that print chip patterns onto silicon at atomic scale. Each one weighs around 180 tons, contains roughly 100,000 parts, and ships in 40 freight containers on three Boeing 747s. And the way they work sounds like science fiction. Inside the machine, droplets of molten tin fall through a chamber, and a high-power laser blasts them 50,000 times per second into a plasma roughly 40 times hotter than the surface of the sun. That plasma radiates light with a wavelength of just 13.5 nanometers, which gets collected and steered by the smoothest mirrors humans have ever made. Scale one of them up to the size of a country, and its largest imperfection would be about a millimeter tall. All of that violence just to print a pattern. Each machine costs around $200 million. The new high-end A generation is closer to $400 million per unit, and a single advanced fab needs roughly 15 to 20 of them. And only one company on Earth, ASML, knows how to build them. Now, scale this to terrafab. The targets Musk laid out are not subtle. Start at 100,000 wafers per month, then scale to 1 million wafers per month, which, by his own math, is about 70% of everything TSMC produces today across all of its fabs combined. 100 to 200 billion chips per year, 1 terawatt of AI compute per year. Let's run some back-of-the-envelope calculations to feel what that actually means. Take the best Nvidia GPU and assume a fab running 30,000 wafers per month at 85% yield, meaning about 85% of the GPU dies per wafer actually work. And remember, wafers are circular, so some dies are always lost at the edges. Under these assumptions, a single TSMC class factory produces roughly 40 gigawatts of compute per year. Now, push that to terrafab targets. To reach 1 terawatt of compute per year, you would need to build the equivalent of 25 semiconductor fabs in one place. That means over 300 EUV machines. Here is the constraint. ASML shipped 48 EUV systems in all of 2025, and they're pushing towards 60 this year. So, this one project would swallow roughly 5 years of the entire planet's EUV production. That's tens of billions of dollars for a single tool class before you've bought anything else. Instead of spreading that capacity across Arizona, Texas, and Ohio, the idea is to compress all of it into a single site, over 100 million square feet of manufacturing space, which Musk says will be 50 times the size of the Pentagon when complete. But this is where it gets uncomfortable, because that is an extreme concentration of value. You're putting the equivalent of an entire semiconductor ecosystem into one physical location. And that creates a new kind of risk. One fire, one flood, one bad contamination event, and you're not losing a fab. You may wipe out the equivalent of multiple fabs at once. For scale, a single modern 2-nanometer fab costs around $28 billion, more than two nuclear aircraft carriers. Terafab wants to park the equivalent of 25 of them under one roof. This is a new kind of scale, where efficiency goes up, but fragility goes up with it. And then it gets even harder, because the plan is not just to make logic chips in that one place. Right now, even chips built in the US get shipped back across the Pacific for packaging. A wafer processed in Arizona still flies to Taiwan or Malaysia to be cut, stacked, and assembled, because America's first advanced packaging campus, Amkor's $7 billion site in Peoria, only broke ground in late 2025. This is a real bottleneck, and TSMC's CoWoS packaging, the technology every AI accelerator depends on, is sold out through 2026, with customers waiting a year or more in line. That's why Terafab is such a bold move. Here, they're trying to build the entire semiconductor stack, logic, memory, packaging, and testing, all of it pulled into one place. And that's something the industry has avoided for decades for a reason. Each of these steps is its own world. Manufacturing logic is one kind of art. Packaging adds another level of complexity, thermal, alignment, stacking. And these systems don't naturally coexist. Mixing them risks yield loss. Now, add memory, and this is where it breaks. There is an extreme shortage of high-bandwidth memory, which is critical for AI chips, and it totally makes sense to want your own supply. Just look at the numbers. SK Hynix told investors that its entire output for 2026, DRAM, NAND, and HBM, was already sold out before the year even started. Micron's HBM is fully committed under long-term contracts. Conventional DRAM contract prices jumped so hard that in the first quarter of 2026, they nearly doubled in a single quarter. And the new HBM four stacks that feed Nvidia's Rubin GPUs reportedly go for around $560 a piece, more than a 50% jump over the previous generation. And every single Rubin GPU swallows eight of them. Memory is now so scarce that it, not the GPU die, is often the thing that decides how many AI systems get built. So yes, owning your own memory supply sounds brilliant, but memory is a completely different type of factory. It uses different process flows, different tools, and produces far more dies per wafer than logic. The equipment is expensive, each machine costs tens of millions, and a single facility needs hundreds of them. And the problem is, these machines are already sold out, too, because SK Hynix, Samsung, and Micron are expanding as fast as they can. SK Hynix alone just committed around $38 billion to new memory plants. This leads to one outcome, backlogs everywhere. But the core problem is not backlogs and not capital. It's orchestration. You have to align lithography, etch, deposition, metrology, inspection, packaging, and test. And even if you get all of that right, the first wafers won't be good because a new factory doesn't start by printing money. It starts by printing defects. And remember, a single wafer takes three to four months to travel through the entire line. So when a flaw finally shows up at test, you don't have one bad wafer. You may have months of production silently carrying the same mistake. The complexity of this process is insane because lithography affects etch, etch affects deposition, deposition affects electrical behavior. Every step interacts. Every step adds variations. Every step can kill your yield. Learning to control that takes years. So, the real game here is not just building the factory, it's learning how to make atoms behave. When you look at Terafab through that lens, this is one of the hardest challenges in modern engineering, which makes one wonder why anyone would even attempt something like this. Despite the cost, despite the risks. And this is where things start to get really interesting. Right now, advanced node capacity is effectively sold out. TSMC's 3 nanometer lines are fully committed through 2026. It's brand new 2 nanometer node is ramping with four times more customer designs than 3 nanometers had at the same stage. And customers are booking capacity two to three years in advance. In November 2025, TSMC's CEO, C.C. Wei, stood on a stage and said the quiet part out loud. The capacity customers want is about three times more than what exists. His exact words, "Not enough. Not enough. Still not enough." Think about what that means. AI demand alone exceeds supply by a factor of three. Every wafer is a fight. You're competing with Apple, Nvidia, AMD, and Broadcom. And companies are wiring multi-billion dollar prepayments just to hold their place in line. And you can see who wins in this arrangement. TSMC closed 2025 with $122 billion in revenue. And in the second quarter of 2026, it posted a record $40 billion with a 67.7% gross margin. Two-thirds of every dollar flowing through the most important factories on Earth stays with the factory. So, everyone else waits, and waiting means delayed products and lost momentum. This is the real bottleneck because right now Tesla and SpaceX design their chips, but they don't manufacture them. And the moment you depend on someone else to manufacture your chips, three things happen. You pay their margins, you depend on their capacity, and your innovation cycle becomes very long. That's fine when chips are just components, but that world is gone. Now chips are the product. Autonomy is chips. AI is chips. Satellite communication is chips. Basically, the entire business collapses into compute. Musk said it himself when he first floated the idea at Tesla's shareholder meeting in November 2025. Even the best-case output from his suppliers was still not enough. So, in his words, "I think we may have to do a Tesla Terafab. It is like a Gigafab, but way bigger." And by the [snorts] time the project officially launched in Austin in March 2026, the framing had hardened into something almost existential. We either build the Terafab or we don't have the chips. And we need the chips, so we build the Terafab. And when the obvious question came, "Isn't this just a negotiating stick to wave at TSMC and Samsung?" Musk shut it down directly. The Terafab is not about getting leverage over suppliers. It's about making sure the chips exist at all. In his world, the suppliers can sell every wafer they will ever make anyway. The problem is that all of them together are still too slow for the scale he's planning. Now, that's the dramatic version. Let's check whether the boring economics actually support it because this is where it starts to hurt. A chip that costs, let's say, $1 to make might sell for $2. And they buy thousands of them in every car, every rocket, every Starlink terminal. Take a Tesla Model 3. Inside that car, you're looking at up to 3,000 chips, roughly $2,000 worth of silicon, and most of them are not AI chips. These are microcontrollers, power chips, sensors. Each one is cheap, but each one carries a margin of 40 to almost 60%. Texas Instruments runs at 58% gross margin NXP at 57%. If you add it all up, this is where the money disappears. And this number only grows. The average car carried about $500 worth of semiconductors in 2020. And analysts project around $1,400 by 2028. Cars are quietly turning into data centers on wheels. And every year a bigger slice of every car flows straight to the chip vendors. Tesla learned this lesson the hard way in 2021 when the chip shortage hit and the company had to design, validate, and rewrite firmware for 19 different substitute controllers in a single quarter just to keep its factories running. Chips nobody thinks about almost stopped the production lines of one of the most valuable companies in the world. Then you add the brain, the AI inference computers. Today's cars run AI 4, built for Tesla by Samsung. Next comes AI 5. These are expensive by design, built on advanced process nodes, and it's about $200 per car just for this AI silicon. And AI 5 is where you can see how serious this has become. Musk claims that by some metrics it's 40 times better than AI 4, about eight times the raw compute, nine times the memory. And that it delivers a third of the power draw of an Nvidia Blackwell at under a tenth of the cost, at least for Tesla's specific workloads. The first samples came out of the fab this spring, a half radical die with a dozen memory packages stacked around it. And yes, that memory comes from the same shortage-stricken market we just talked about. Volume production is planned for 2027. And here is the detail that matters for our story. Every single AI 5 will be made in America. Split between TSMC's fab in Arizona and Samsung's new fab in Taylor, Texas. For AI 6, Samsung and Tesla signed a 16.5 billion dollar manufacturing deal and Musk called its strategic importance hard to overstate promising to personally walk the production line to speed things up. So Tesla already went from buying chips to co-managing fabs. The terra fab is just the logical next step of a slope they're already sliding down. Now if you manage to move that cheap high volume manufacturing in-house you can save up to $1,000 per car which means up to 12% higher margin on the same product. Scale that. If robo-taxi reaches 10 million cars per year, that same dynamic turns into 5 billion dollars saved every single year. And to be fair, we should keep our feet on the ground here. As of this summer, Tesla's actual robo-taxi fleet in Austin is a couple dozen vehicles while Waymo operates thousands. So between today's reality and 10 million auto autonomous cars per year, there is a long long road. But even the skeptical version of this future, slower, later, smaller, still needs millions of AI computers per year. The direction is clear and the math doesn't need 10 million cars to start working because now move to robots and the volume there is a different league. Tesla is standing up its first Optimus production line in Fremont with a capacity of around 1 million robots per year and the next generation line at Giga Texas is targeted at 10 million per year long term. Each robot is stuffed with compute, sensors, and power electronics, around 10,000 unique components. This market can grow 10 to 100 times beyond automotive and today there are barely any purpose-built chips for it, mostly Tesla's internal stack and Nvidia's Jetson platform. So the market is wide open and if Tesla captures even 10 to 20% of it, you are looking at a multi-billion dollar impact. Add it all up, cars, robo taxis, robots, data centers, and you understand why Musk posted in November 2025, "Tesla will build more AI chips than everyone else combined." I'm not kidding. When you consider that volume and the economics of scale, Terafab starts to make more and more sense. But, here's the part no one expected. When the Terafab targets were finally announced, it turned out most of the output won't be AI chips for cars or robots at all. It will go into space chips, up to 80% of the wafers. Only 20% of this gigantic factory is meant for Earth. And if that sounds completely upside down, stay with me because this is where the whole story flips. So, where does all of that silicon end up? On Earth, we know how to build insanely powerful GPUs. But, the moment you leave the planet, silicon enters a completely different reality. Space is unforgiving. It doesn't fail your hardware all at once, it degrades it. Up there, high-energy particles rip through circuits, flip bits, corrupt calculations, and slowly wear down transistors. A single charged particle can flip a memory cell, and suddenly your computer is confidently doing math with corrupted numbers. There is also the slow poison, total ionizing dose, radiation accumulating in the chip over months and years, shifting how transistors behave until one day they simply stop switching. And there is the nightmare scenario, latch-up, where a particle strike effectively short circuits part of the chip. And if you don't cut the power within moments, it cooks itself. So, standard chips, as they are, won't survive everywhere in space. The traditional answer is radiation-hardened silicon, and the traditional answer is brutally expensive. Let me give you the numbers because they are hard to believe. The most famous space processor is the RAD750. It flew on the Curiosity and Perseverance rovers and on the James Webb Space Telescope. It runs at up to 200 MHz, roughly the performance of a late '90s desktop PC. It's built on technology from about 25 years ago and it costs around $200,000 per chip. A processor from the dial-up era for the price of a supercar. And that's the pattern across the whole industry. An ordinary commercial part might cost $35 and it's fully radiation hardened equivalent can run $47,000, more than a thousand times more. Even the milder radiation tolerant middle class sits at $500 to $5,000 per chip. Why so expensive? Because fewer of them work on each wafer. Because every single one goes through testing far beyond normal checks. You literally shoot them with particle beams in accelerators at places like Brookhaven's Space Radiation Lab to mimic years of cosmic exposure. And then comes regulation because rad-hard chips are treated like defense technology under export control law. All of this adds time, cost, and complexity and the lead time stretch to a year or more. This is why innovation in space computing moves painfully slowly. Designs take years, everything has to be proven, tested, certified, and by the time it flies, it's already a museum piece. But the space economy can't scale like that. At tens of thousands of dollars per chip, space economics collapse. So the only way out is control, controlling the manufacturing and redesigning the entire stack, breaking the cycle and shifting the cost curve from tens of thousands of dollars per chip down into the hundreds. And this is exactly the bet behind the D3, the chip unveiled alongside Terafab, a processor designed from scratch to run AI in orbit, radiation resilience included, at manufacturing volumes the space industry has never seen. And here is what I find fascinating. SpaceX has been quietly rehearsing this philosophy for 15 years. Their Dragon capsules don't fly $200,000 processors. They fly ordinary commercial chips arranged in voting pairs, multiple processors computing everything in parallel and constantly checking each other with dozens of processors spread across the vehicle. When radiation flips a bit in one, the others outvote it, the corrupted unit reboots and rejoins. As one of SpaceX's avionics leads put it, the parts aren't hardened. The design as a total system is hardened. That single sentence is the key to the entire 80%. You don't need every chip to be indestructible. You need the system to survive and you need chips cheap enough to be redundant. Do the math on the constellation SpaceX has filed for, tens of thousands of satellites each stuffed with processors, modems, beam formers, and laser link silicon. At rad hard prices, that fleet is financially impossible. At consumer prices with smart redundancy, it's a business. And it helps that low Earth orbit is actually the mild end of space. Google recently published data suggesting that over a five-year mission in low orbit, a chip might absorb around 750 rads of total dose, while the old-school rad 750 is rated for 200,000 to a million. In other words, for satellites close to Earth, you don't need a tank. You need a good umbrella and a lot of spares. Deep space with its galactic cosmic rays and the radiation belts is a different monster entirely. That's where fully hardened designs stay unavoidable and a big part of that cost sits in the packaging. Heavy shielding and careful isolation wrapped around the core logic because failure is not an option when your system is halfway to Mars. But the volume isn't in deep space. The volume is in orbit right above our heads. And now, the obvious question, what could possibly need so many chips in orbit that it justifies 80% of the largest factory on Earth? If you're enjoying this episode so far, subscribe to the channel because the answer is one of the wildest ideas in tech right now, data centers in space. In November 2025, Musk wrote that simply scaling up Starlink batches of three satellites, which already carry high-speed laser links, would work and I quote, "SpaceX will be doing this." He followed up with a claim that Starship could deliver 100 gigawatts per year of solar-powered AI satellites to high orbit within 4 to 5 years. He even sketched the end game, 100 terawatts per year, possible from a lunar base churning out solar-powered AI satellites from local materials. Classic Musk. But notice how each step is just the previous one scaled. And at the terrafab launch, he went further, predicting that within a few years, it will actually be cheaper to launch AI compute into space than to build a traditional data center on the ground. The logic sounds insane until you remember what's strangling data centers on Earth, power and cooling. In orbit, the sun shines almost continuously, there are no neighbors to annoy, no grid interconnection queue, no water needed, and you radiate your heat straight into the void. The economics flip the moment launch gets cheap enough. And making launch cheap is the one problem SpaceX has probably solved before. This is not just talk, by the way. The first hardware generation of this idea already flew. Last November, a startup called Starcloud launched a satellite carrying an Nvidia H100, the first data center GPU in orbit, and trained the first AI model in space a month later. They've since raised $170 million to chase a vision of gigawatt-scale orbital data centers. Google unveiled Project SunCatcher, planning to fly TPUs on prototype satellites in early 2027. By their math, a solar panel in the right orbit can be up to eight times more productive than the same panel on Earth, drinking sunlight almost continuously. Their proton beam test showed the TPU compute cores hold up surprisingly well, and the first thing to complain was, you guessed it, the memory. And SpaceX, in July 2026, a Starship finally deployed the first batch of Starlink I-5 3 test satellites on a suborbital trajectory, and the first orbital batch is expected in the coming weeks. Each V3 satellite is designed to push a terabit per second of downlink, roughly 10 times the previous generation, and a single Starship can carry 60 of them. 60 terabits of new capacity per launch. The scaffolding of an orbital computer economy assembling itself launch by launch. These spaceships are the future volume, but right now, today, most of SpaceX's silicon hunger is actually on the ground, hiding in plain sight in Starlink terminals. These dishes are very silicon hungry, and the numbers here are just fun. The very first Starlink dish contained around 80 custom beam-forming chips. SpaceX has since squeezed that down to about six much smarter chips steering roughly 1,200 antenna elements. And the estimated cost of building a terminal collapsed from around $3,000 to a few hundred dollars. These chips are designed by SpaceX but manufactured by STMicroelectronics in Europe. And ST revealed that over roughly a decade, it has shipped more than 5 billion chips for Starlink. With volumes on track to double by 2027. SpaceX's Texas factory now stamps out terminals at a rate of about 15,000 per day for a subscriber base that passed 12 million this summer and is climbing toward 13. Starlink pulled in over $11 billion in revenue in 2025 with forecasts approaching $16 billion this dollars this year. And every single dish, every satellite, every gateway is packed with custom silicon that SpaceX designs but cannot manufacture. Now add the D3 program for orbital data centers on top and suddenly Terafab makes a lot more sense. Especially when space-grade chips are also a matter of national security. Wrapped in defense regulations by default. If you manage to bring even part of that stack in house, you take control of the most critical layer of the system. You secure supply. And over time, you own the economics of an entirely new industry before anyone else even realizes it exists. So now we have the demand side. An explosion of chip hunger across cars, robots, and space. With that kind of demand, the solution looks obvious. Bring manufacturing in-house. But if it's so obvious, why is nobody else doing it? Because building advanced semiconductor capacity is one of the hardest things humans have ever attempted. But then again, so is landing rockets. One of the first real constraints is not engineering, it's availability. EUV machines are not just expensive, they are scarce. ASML's order backlog stands at almost 39 billion euros, about 2/3 of it EUV, and the machines take years to build. Meanwhile, Intel, Samsung, and TSMC have been placing orders years in advance, and TSMC alone operates the majority of all EUV machines ever made. So the real question is, did Terafab secure its machines already? Because if they didn't, the timeline breaks, not by months, but by years. And lithography is just one line item. A modern fab needs an entire catalog: deposition, etch, ion implantation, metrology, inspection, cleaning, polishing, packaging. That's hundreds of tools, most with 12 to 24-month lead times, then months to install. An EUV machine alone takes months just to reassemble on site, months to calibrate, and years before they are actually mastered. The official timeline says construction starts this year, first chips in late 2027, volume production no earlier than 2028. For reference, TSMC broke ground in Arizona in 2021 with a 30-year-old playbook for building fabs, and shipped its first volume wafers at the end of 2024, and that was widely celebrated as fast. Terafab is attempting something bigger than any single fab in history on a first try. So when you hear chips in late 2027, treat it the way you treat every Musk timeline, as a direction, not a date. But even if the machines arrive on time, you're still not safe, because Terafab is aiming straight at one one the most complex transistor architectures we've ever built, gate all around. This is the next big leap after FinFET. In older chips, the transistor channel sat flat, controlled from one side. FinFET lifted the channel up into a thin fin, wrapped the gate around three sides, and suddenly you had much better control. That device carried the entire last decade, every advanced phone, every GPU. It worked brilliantly until it didn't. As scaling pushed further, the fins had to become absurdly thin and tall, leakage crept back, control degraded, and the transistor had to be reinvented again. That's where gate all around comes in. You slice the channel into a vertical stack of nano sheets. Each sheet just a few nanometers thick, stacked three or four high, and you wrap the gate around all four sides of every sheet. Beautiful control, horrible manufacturing. Because now you've turned a relatively flat structure into a true 3D structure at atomic scale. Every layer thickness, every spacing, every edge profile has to be controlled with atomic precision across billions of devices, and at 2 nanometers, you're doing it with something like 25 to 30 EUV lithography layers per wafer out of a thousand or more total process steps, stretching over three to four months per wafer. And within this entire flow, there is one specifically dangerous moment. To build the nano sheets, you first grow a crystal sandwich of alternating silicon and silicon germanium layers. Then, deep into the process, you perform the channel release. You dissolve the silicon germanium out from between the silicon sheets, leaving them suspended in midair like the floors of a building with no walls, and then you fill every gap around them with gate material. If the tiny structures called interspacers are slightly off, if the gate doesn't perfectly wrap every sheet, defects explode. And these defects are the worst kind because they are buried deep inside a 3D structure that inspection tools can barely see into. Optical inspection is nearly blind here. Fabs are being forced into E-beam and even acoustic techniques hunting for flaws just 2 to 3 nanometers across hidden under layers of material. Very often you only find out something is wrong later during stress testing or worst of all in the field. And we don't have to speculate about how hard this is because the industry is living it right now. Samsung reached gate-all-around first back in 2022 and its early 3 nanometer yields were reportedly stuck between 10 and 20%. Eight out of 10 chips dead on arrival for quarters on end. Even now reports put Samsung's 2 nanometer yields around 50 to 60% and here's the twist that makes it personal for this story. Tesla's own AI 6 chip at Samsung's Texas fab has reportedly already slipped by about half a year precisely because of 2 nanometer yield struggles. TSMC, the best manufacturer on Earth, got its 2 nanometers into volume production in late 2025 with healthy yields. The company says defect density on this node is actually lower than 3 nanometers was at the same stage. But look at how they did it. Gate-all-around runs first in the Taiwan mother fabs where thousands of engineers have decades of accumulated process knowledge. The Arizona site started with older mature FinFET nodes and is ramping step-by-step copying a proven recipe into desert conditions and 2 nanometer class production in Arizona won't realistically arrive until 2027 or 2028 6 plus years after groundbreaking. They are not reinventing anything and it still takes that long. There is exactly one other from scratch 2 nanometer attempt on the planet, Rapidus in Japan, backed by roughly 19 billion dollars of government money and the combined will of an entire nation. They broke ground in 2023, got Japan's first EUV machine printing in 2025, produced their first working 2 nanometer test transistors in the summer of 2025, and are targeting mass production in 2027. And in this industry, that pace is considered miraculous. That is the real benchmark for starting from zero at the bleeding edge. Now, imagine doing it at Terafab scale. Stabilizing atomic level precision across thousands of steps in a brand new mega facility with a workforce that has never run a fab. For context, TSMC employs more than 80,000 people, a huge share of them process and yield engineers carrying decades of tribal knowledge, and even they treat 2 nanometers with something close to fear. That is a harsh starting point. So, how do you de-risk something like this? You bring in someone who has already paid the tuition. At the November shareholder meeting, Musk dropped the hint, "It's probably worth having discussions with Intel." Everyone treated it as a throwaway line. It wasn't. In April 2026, it became official. Intel signed on as a foundry partner for Terafab alongside Tesla, SpaceX, and XAI. And I think this is one of the most interesting pairings in modern industry because these are two completely opposite philosophies, Silicon Valley's oldest, most process-obsessed manufacturer, and the move-fast empire that blows up rockets on purpose to learn faster. But directionally, it's a very smart and pragmatic move because Intel brings exactly what TerraFab lacks. Intel has 18A, its 2 nanometer class node, running in high-volume production in Arizona since late 2025 with Panther Lake laptop chips launched at CES in January 2026. And 18A is not a me-too node. It pairs gate-all-around transistors with backside power delivery, feeding the transistors from underneath the wafer, a technology even TSMC won't ship until its A16 node arrives. Intel was also first in the world to ship silicon patterned with high-NA EUV, the $400 million next-generation machines. Intel has advanced packaging capability on US soil, and Musk has already said Tesla will use Intel's next node, 14A, by the time it is mature. The design kit for that node lands with customers this October. Intel gets something priceless in return, volume. Intel Foundry lost $13 billion in 2024 and another $10 billion in 2025. Outside customers brought in less than $300 million last quarter. The US government took a 10% stake in the company. Nvidia put in $5 billion and started co-developing chips with them, news that handed Intel stock its best day in nearly four decades. Intel is a national champion desperately searching for a whale customer, and TerraFab might be the biggest whale in history. One side has decades of manufacturing scars and empty capacity. The other has bottomless demand and no experience. It's either a perfect marriage or a spectacular culture clash, and most likely it's both at the same time. Either way, with Intel inside, the odds of Terafab producing real chips just went up meaningfully. But even a partner like Intel doesn't remove the core difficulty of running the most advanced process at unprecedented scale. And this is where a dangerous idea appears, straight from Musk himself. Does it even have to be so complicated? Maybe we are over-engineering the factory itself. This is where the cheeseburger comes in. In January 2026, on the Moonshots podcast, Musk said this, word for word, "I think they're getting clean rooms wrong in these modern fabs. I'm going to make a bet here that Tesla will have a 2-nanometer fab, and I can eat a cheeseburger and smoke a cigar in the fab." His reasoning, "Just keep the wafers isolated the entire time, and the room around them stops mattering." The semiconductor world collectively choked on its coffee. But let's take the idea seriously for a moment because it's not as dumb as it sounds, and then let's look at where it breaks. At first, it sounds like a great idea. Let's simplify. Let's cut the costs. Because cleanliness in a modern fab is at levels that are hard to comprehend. The most critical zones are thousands of times cleaner than a hospital operating room. Let me put a number on that. The air on a city street carries tens of millions of particles per cubic meter. The cleanest class at wafer level, ISO 1, allows 10. Not 10 million. 10 particles. The air is exchanged hundreds of times per hour through multi-stage filtration. Humans are the enemy. Even standing still in a full bunny suit, you shed around 100,000 particles per minute. Start walking, and it's a million. Building clean room space at this grade costs thousands of dollars per square foot. Now, imagine doing this at Terafab scale, at 100 million square feet. That is roughly 60 times more floor space than TSMC's biggest gigafab clean room. Keeping that entire volume ultra clean is not just insanely hard. It might be economically impossible. So, the idea is indeed tempting. Relax the room, protect the wafer, automate everything with robots, move faster, spend less. And here is the thing, the industry already half agrees with him. Inside a fab, wafers don't just lie around in the open. They travel in sealed containers called FOUPs. Each one a climate-controlled mini environment holding 25 wafers, purged with nitrogen to push out oxygen and moisture because even tiny changes in humidity can trigger unwanted chemical reactions on an unfinished wafer. Robotic overhead vehicles carry these pods along kilometers of ceiling rails. In the biggest fabs, that network exceeds 200 km of track with 10,000 robotic carriers, and human hands never touch a wafer. And they'd better not because a finished 2-nanometer wafer sells for around $30,000, meaning a single pod gliding over your head carries 3/4 of a million dollars in sliced sand. The air inside the pod is orders of magnitude cleaner than the room. So, Musk's intuition, wafer isolation the entire time, is a real thing. It exists. It's called a FOUP. So, if wafers are protected in this box, why does the surrounding environment still matter so much? Because the wafer cannot stay in the box. That's the part the cheeseburger theory skips. A 2-nanometer wafer goes through on the order of a thousand process steps, and at every major one, lithography, etch, deposition, cleaning, measurement, the wafer must leave the pod and enter the tool. Yes, the tool has its own mini environment, but the wafer is exposed inside machinery that is itself violent. Plasma chambers, moving robotic arms, valves opening and closing, chemicals outgassing. Even at parts per billion concentrations, stray molecules drifting through the air can change how a finished transistor behaves. The tools themselves shed particles. They flake. They get opened for maintenance and repairs. Every machine is simultaneously the thing building your chip and a source of contamination trying to kill it. And at 2 nanometers, the margin for error is basically zero. A particle far too small to see landing on a circuit is like an asteroid hitting your chip. One invisible speck can destroy thousands of transistors, and a handful of them in the wrong places can drag your yield below the line where the entire business stops making sense. Remember, yield is everything. The whole economics of a fab lives and dies by how many chips per wafer actually work. That's why the system is built in layers, Foup, mini environment, clean room, bunny suits, filtered air, each layer catching what the previous one missed. Remove one layer and the system doesn't get slightly worse. It gets unstable. So, can you build a cheaper, dirtier fab? Maybe, but not by deleting a layer and hoping. You'd have to go back to first principles and redesign the process, the tools, and the environment together with the toolmakers themselves. Companies like ASML and Applied Materials co-engineering machines that never expose the wafer at all. That is a fascinating research direction. It is also exactly the kind of thing that takes a decade, which is the one resource this project doesn't have. You can automate everything, but you still can't escape the physics. Verdict on the bet. I would not bet against Musk on manufacturing in general, but if that cigar ever gets lit inside a working 2 nanometer fab, it will be the most expensive cigar in human history. Now, all this complexity depends heavily on one question. What kind of technology are you actually running inside? Because Terafab is not one factory, even if we ignore memory, packaging, and testing. At the logic level alone, it splits into two. One is the glamorous part, a cutting-edge 2 nanometer logic fab, the one everyone talks about. The other is very different, a specialized factory for space-grade chips. And the moment you take that second part seriously, the entire picture shifts, because when you optimize for space, you stop optimizing for raw performance. The more you shrink a transistor, the more fragile it becomes. In those tiny dimensions, a single radiation event can take down an entire logic block. So, for the space share of the fab, instead of pushing forward, you partially step back, maybe one or two nodes, because in orbit, reliability beats gigahertz. And this is where a technology called silicon on insulator, SOI, becomes relevant. The idea is simple. You build the transistor on top of a thin insulating layer, instead of directly on bulk silicon. That buried insulator dramatically limits how much charge a particle strike can inject, which means far fewer bit flips, and crucially, it structurally suppresses latch-up, that self-destruct short circuit we talked about earlier. Here's the interesting part. SOI is not an alien technology. It's essentially standard CMOS with a different starting wafer. The substrate costs a few times more, but at the finished wafer level, the premium nearly evaporates because you save process steps elsewhere. You are not rebuilding the factory, you're extending it. And here is the beautiful detail. SpaceX already knows this world intimately because the chips inside Starlink terminals are built by ST on exactly this kind of technology. This is not a theoretical choice, it's the technology their entire consumer fleet already runs on. The alternative would be to keep standard CMOS everywhere and push all the radiation protection into the package, heavy shielding around the die. But that gets expensive fast and doesn't solve everything at the device level. So this becomes a strategic choice. Do you solve radiation in the transistor or later in the stack? And whichever way they go, that choice tells you what this factory really is. It's not a vanity project chasing the smallest number. It's an instrument of control, that control over the supply chain so you're not exposed to geopolitical shocks, export rules, or another company's order book for the exact silicon your cars, robots, and satellites cannot exist without. But once you take control seriously, you run into a much bigger, much more physical question. Where on earth do you actually build something like this? When you're choosing a location for an extremely expensive semiconductor fab, you look first at seismic stability, access to massive amounts of power and water, and proximity to skilled talent. And in August 2026, we finally learned the answer. Grimes County, Texas, about an hour northwest of Houston, on the site of the retired Gibbons Creek Coal Power Plant. >> [snorts] >> And I love this choice because it tells you they've been paying attention. A dead coal plant comes with the two things every mega project bleeds for. High voltage grid infrastructure already wired into the site, and its own reservoir, so the fab drinks surface water instead of draining the local groundwater. It's the same scavenger logic XAI used when it stuffed the world's largest AI supercomputer into an abandoned appliance factory in Memphis and imported second-hand gas turbines to power it. Don't build from zero, occupy the carcass of the old industrial world and pump new blood through it. First phase, $16.8 billion, at least 3,000 jobs, backed by a $30 million state grant and a 35-year local tax abatement. Meanwhile, an hour up the road, Tesla is building a small $3 billion research fab right at Giga Texas, about 1,000 wafers a month, tiny by industry standards, but with mask-making logic, memory, and packaging under one roof. In Musk's words, a place to try out new physics. That's the test kitchen. Grimes County is the restaurant chain. And zoom out, because suddenly Texas doesn't look crazy at all. If this were happening in some random location with no semiconductor ecosystem, the skepticism would be justified. But Samsung's new 2-nanometer fab in Taylor sits 30 miles from Tesla's Gigafactory, the same fab that will build Tesla's AI chips, and Samsung's total Texas commitment now tops $37 billion. That Taylor fab moved its equipment in this spring and is switching on this year. With Tesla silicon among the very first to run through it. Texas Instruments is pouring more than $60 billion into US fabs, with up to $40 billion at a single mega site in Sherman. NXP and others run fabs across Austin. The state runs its own $1.4 billion Texas chips fund on top of the federal one. Equipment makers and chemical suppliers have been settling around these anchors for years, because a fab doesn't run on machines alone. It needs raw silicon wafers, specialty chemicals, ultra-pure gases like neon and helium, and the logistics to deliver all of it every day without fail. That means supply chains, logistics, and most critically, poachable talent already exist. The grid operator is bracing for what's coming, too. Texas peaks at around 98 gigawatts today, and if every proposed data center and factory on the books materializes, demand requests push past 360 gigawatts by 2032. Texas is simply where America builds big right now. This goes far beyond Tesla, by the way. Over 90% of the world's most advanced chips are still manufactured in Asia, with the overwhelming share concentrated on one island whose political future is one of the most dangerous open questions on Earth. Rebuilding this capability on American soil is exactly what the Chips Act was designed to spark, and it explains why Washington isn't laughing at Terrafab. It's leaning in. A government that owns 10% of Intel, Terrafab's manufacturing partner, is now structurally invested in this succeeding. Strategically, the stars have aligned in a way that would have been unthinkable 5 years ago, which brings us to the money, because the money tells its own story. The official numbers climb like a rocket staging, 20 to 25 billion dollars for the pilot phase, per Tesla's CFO, then filing surface showing about 55 billion dollars proposed initially, and up to 119 billion dollars across four phases. And Bernstein's analysts, running the numbers on the full 1 terawatt vision, landed near 5 trillion dollars. With a T. For comparison, that's more than the entire semiconductor industry has invested in fabs in its history. TSMC, the biggest spender the chip world has ever produced, deploys about 60 billion dollars in a good year. At that pace, $5 trillion is most of a century of spending. So, either the vision shrinks or someone invents a radically cheaper way to build fabs, which by the way is exactly why the cheeseburger bet is not just a joke. Musk needs the cost of clean manufacturing to collapse or the math of his own project buries him. Nobody pays for that with car margins. And this is where the timing stops looking coincidental because think about what happened around this project. In February 2026, SpaceX absorbed XAI. In March, Terafab launched as a joint Tesla SpaceX XAI project. And in June, SpaceX went public in the largest IPO in human history, raising around $75 billion at a valuation near $1.75 trillion. Investors weren't just buying rockets and satellites. They were buying the story of a company that designs its own chips, manufactures its own chips, launches them into orbit on its own rockets, and sells the compute back down to Earth. Full stack sand to orbit. Whether or not the fab hits its dates, that story already financed itself. This is the most Musk move imaginable. Take the most expensive object ever proposed and turn it into the reason people hand you the money to build it. Interestingly, there is one more layer here because this factory would turn Musk's empire into something we haven't seen in half a century, a fully vertically integrated technology giant. This is how the old titans were built. Philips didn't just make electronics, it co-founded ASML, the lithography monopoly, and held a 27.5% stake in TSMC at its founding. Then spent decades selling all of it off piece by piece. IBM once made its own chips and its own fabs. By 2014, it was literally paying GlobalFoundries $1.5 billion to take its factories away. AMD survived only by amputating its fabs in 2009. For 40 years, the entire industry moved in one direction, specialize, disaggregate, go asset light, let TSMC carry the capital burden for everyone. And the fabless companies won huge. Nvidia and Apple became the most valuable companies on Earth without owning a single fab. Now, watch the cycle turn back. Apple designs its own silicon, Google builds TPUs, Amazon, Microsoft, all designing custom chips. OpenAI signed with Broadcom for 10 gigawatts of its own accelerators. Every giant is crawling back down the stack. But every one of them stops at the same line, the fab. They design, TSMC manufactures. Musk is the only one proposing to cross that final line, which, to be fair, is on brand. SpaceX already builds roughly 70% of each rocket in-house, and Tesla famously integrated everything from battery cells to giant castings whenever suppliers couldn't keep up. Crossing one more line is what this empire does. But crossing this particular line comes with a curse because the fab is the most unforgiving business on Earth. A chip factory only makes money when it's full. The depreciation clock ticks every second. These tools become obsolete in 5 to 7 years, whether you use them or not. If demand arrives late, if robo-taxi stays a couple dozen cars in Austin, if Optimus ramps slowly, if orbital data centers take five extra years, the costs don't wait. This is exactly what happened to Intel. They were the kings, they slipped on one process node, 10-nanometer promised for 2016, delivered in 2019, and the empire cracked. They were stuck carrying the full weight of their factories, while fabless rivals danced ahead, flexible and light. GlobalFoundries hit the same wall and simply gave up on the leading edge in 2018. Vertical integration gives you total control, and it removes your safety net, both at once, always. So, let's land this. On one side of the scale, physics, history, and every expert instinct in the industry, including Jensen Huang, whose reaction to Musk's fab ambitions was to remind everyone that what TSMC does is extremely hard, the engineering, the science, and as he put it, the artistry. Translation, good luck. The yield learning curves cannot be bought, the EUV queue cannot be skipped. The 2028 time