Elon Musk wants to build the largest chip fab on Earth. The building was never the hard part.

Stand on the floor of a leading-edge fab and the first thing you notice is how little of it looks like the future. The rooms are beige and windowless. The air is scrubbed a thousand times cleaner than an operating theatre, moved so constantly that a single speck of dust settling on a wafer at the wrong moment is treated as an industrial accident. Machines the price of a wide-body jet stand in rows, drawing circuits finer than a virus using light that does not occur in nature and has to be coaxed into existence one pulse at a time. Nobody in the room is thinking about robots, or rockets, or a trillion of anything. They are thinking about yield — the share of chips on each wafer that actually work — and about the dozens of ways a process that ran clean yesterday can quietly stop running clean today.

That room is what Elon Musk is proposing to build in Grimes County, Texas, about an hour northwest of Houston, at a scale no one has attempted. On August 6, Tesla and SpaceX said they would jointly commit $16.8 billion to a first phase of a semiconductor complex called Terafab, a plant Musk described as "the largest and most valuable building on Earth by far." The renderings show more than 100 million square feet of manufacturing space. The stated ambition is to fabricate, package and test advanced logic and memory chips under a single roof — the silicon for Tesla's Optimus robots and Cybercabs, and the higher-power parts for the data centres SpaceX wants to fly in orbit. The number Musk's team likes to quote is capacity for "over one terawatt of compute per year."

It is an extraordinary announcement, and the temptation is to argue about the ambition — whether the world needs a building this size, whether orbital data centres are a real market or a pitch deck. Set that aside. The more useful question is the one the semiconductor industry asks about every new fab, and asks most sharply about the ones announced with the most confidence: not whether you can pour the concrete, but whether you can make the thing inside work. Terafab's hardest problems are not the ones in the press release. They are upstream of it, and most of them are not for sale.

The announcement and the filing

Start with the gap between what was said on a stage and what was written for lawyers. When SpaceX filed for its public offering in the spring, the same Terafab project appears in the paperwork in a very different register. There it is described, in the reported language of the filing, as a "general framework" — an arrangement with "no binding commitments" and "no obligation for either side to keep participating." The figure has moved, too: an earlier version of the project was floated in March at around $25 billion; the number announced this month is $16.8 billion for a first phase, with SpaceX documents pointing at a build-out that could eventually reach roughly $119 billion if every stage proceeds.

None of that means the project is fictional. Large industrial commitments are staged precisely this way, and a framework is how these things usually begin. But it matters, because a fab is the one kind of factory you cannot will into existence with a headline number. A carmaker can bring an assembly line up in a year. A leading-edge fab is a different category of object. TSMC's plant in Arizona — a company that has built more advanced fabs than anyone alive, dropping a known process into a new geography — took years longer and cost far more than the first announcements implied, and spent much of that time fighting the unglamorous problem of getting an American site to yield like a Taiwanese one. That is the base rate for people who already know how. Terafab proposes to clear the same bar as a first-time fab operator, at the largest scale ever attempted, across not one hard process but three.

What a terawatt of compute actually asks for

"Over one terawatt of compute per year" is a slogan, not a specification, and it is worth pulling apart because the slogan hides the work. Compute is not poured; it is patterned, atom by atom, onto silicon, and every step in that patterning is a discipline that took the incumbents decades to master and that no amount of capital compresses cleanly. A modern AI accelerator is not one chip. It is a large logic die, made on the most advanced process a foundry can run; a stack of high-bandwidth memory beside it, made by a different specialist on a different line; and an interposer that stitches the two together with wiring finer than anything on a circuit board. Terafab's pitch is to do all three in the same building. Each of those three is, on its own, among the hardest manufacturing problems on the planet.

Consider the chip that goes into a robot or a car. Tesla's current AI silicon is fabricated by TSMC and Samsung, the two companies on Earth that can reliably run a leading-edge logic process at volume. "Reliably" is the load-bearing word. The difference between a fab that yields eighty percent of its wafers and one that yields forty is the difference between a profitable line and a money furnace, and closing that gap is not a matter of buying better machines — everyone buys the same machines. It is a matter of process knowledge accumulated over years: the exact recipe of temperatures and chemistries and timings, the failure modes catalogued and designed around, the institutional memory of which tool drifts and when. That knowledge does not come in the crate with the equipment. It is the actual product of a fab, and it is the thing Terafab would be starting without.

The tools Musk cannot vertically integrate

Musk's companies are the best vertical integrators of the age. They make their own rocket engines, their own battery packs, their own motors, their own silicon designs. Vertical integration is the reflex, and the Terafab announcement reads as its logical extreme: if the memory shortage is squeezing you and the foundries are booked, build the foundry. But there is a hard limit to how far up the chip supply chain integration can reach, and it sits exactly where the real chokepoints are.

The machines that draw the finest circuits — extreme-ultraviolet lithography tools — are made by exactly one company in the world, the Dutch firm ASML, and they cost on the order of a few hundred million dollars each. There is no second source. You cannot integrate around them; you can only join the queue, behind TSMC and Samsung and Intel and Micron, all of whom have been reserving capacity for years. Below the machines sit the materials: the photoresists that translate light into pattern, made at pharmaceutical-grade purity by a short list of mostly Japanese suppliers; the ultra-pure gases and specialty chemicals; the blank silicon wafers themselves, roughly ninety percent of which come from five firms, none of them American. These are the obscure upstream names that never appear in a keynote, and they are the ones that decide whether a fab runs. A plant in Texas is downstream of every one of them.

Intel's reported involvement is the tell here, and it cuts both ways. Terafab's backers have said Intel will join the project and help, in the reported phrasing, "refactor" the fab technology — which is a reasonable move, because Intel is one of the few outfits on the planet with the process muscle to stand a leading-edge line up. But Intel is also the cautionary tale. Its own most advanced manufacturing process has spent the past two years searching for outside customers precisely because a leading-edge node is a stranded asset until someone other than its owner is willing to trust it with real volume. If Terafab leans on Intel's know-how, it inherits Intel's timeline and Intel's difficulty. If it does not, it is inventing that know-how from scratch. Neither version is fast.

A fab is not a building you announce. It is a yield number you earn — one contaminated tool, one requalified process, one clean wafer at a time. — On why the hardest part of Terafab isn't the concrete

Memory is not a side quest

The part of the plan that should draw the most scrutiny is the quiet word "memory." Terafab does not only propose to make logic; it proposes to make the memory that sits beside it, at a moment when the entire industry is convulsing over exactly that. For eighteen months this column has traced a single story through the supply chain: the three companies that make the world's memory — Samsung, SK Hynix and Micron — have been redirecting their wafers toward high-bandwidth memory for AI accelerators, because a bit of HBM sells at several times the margin of ordinary DRAM and eats several times the wafer area to make. The result is a shortage that has raised the price of the memory in phones and laptops and pushed even Apple to call the situation a hundred-year flood.

Building leading-edge DRAM is not a lighter version of building logic; it is a parallel decade of craft, with its own physics, its own yield demons, its own packaging art. The reason there are only three companies who can do it at the frontier is the same reason there are only a handful who can run leading-edge logic: the barrier is not the willingness to spend, it is the accumulated process knowledge, and money buys the second faster than the first. To stand up logic and memory and advanced packaging simultaneously, as a newcomer, is to take on three of the four hardest problems in physical manufacturing at once and to do it in the middle of the tightest memory market in fifteen years — the market you would presumably be building the plant to escape.

The desert math

Then there is the part of every fab that has nothing to do with silicon and everything to do with geography. A leading-edge fab is one of the thirstiest and most power-hungry industrial facilities humans build. It needs enormous volumes of ultra-pure water — water stripped of nearly every ion before it touches a wafer — and it needs it continuously, because a fab that loses its water or its power mid-process does not pause; it scraps everything in the line. Terafab's answer, so far, is to draw from the Gibbons Creek Reservoir, a body of water that once cooled a retired coal plant, rather than tap local groundwater. That is a sensible choice, and it is also the first line of a long negotiation, because a plant this size in a part of Texas that already manages its water carefully will be pricing that water against farms, towns and a grid that the state's own AI build-out is straining.

The jobs number frames the same tension from the other side. Terafab is projected to employ at least 3,000 people, most of them local hires, and it has already drawn a $30 million grant from the Texas Enterprise Fund. Three thousand jobs is a real number and a modest one for a building the size of a small city — a leading-edge fab is enormous capital and comparatively few workers, most of them highly specialised, and the specialised ones are exactly the people the whole industry is fighting to hire. Where the process engineers who know how to ramp a new node come from, in a Texas exurb an hour from Houston, is not a detail. It is the same chokepoint as the tools and the materials, wearing a different uniform: the scarce input here is people who have done this before, and there are not many of them, and they are not idle.

How much rides on how little

None of this is a prediction that Terafab fails. Musk's companies have made a habit of doing things the incumbents called impossible, and it would be a poor correspondent who bet against them on reflex. Intel's know-how is real; the capital is real; the strategic logic — that if you are going to consume compute on the scale of nations, you might want to own the means of making it — is sound, and it is the same logic driving OpenAI, Google, Amazon and Meta toward their own silicon. The build-out is genuine even if this particular building is, for now, a framework.

But it is worth being clear-eyed about what the hard part is, because the announcement invites you to look at the wrong thing. The wrong thing is the square footage, the terawatt, the title of largest building on Earth. The right thing is the yield number that will not appear in any press release for years, if it appears at all — the quiet figure on a screen in a beige room that decides whether all of this is a fab or a very expensive warehouse. That number does not care how big the building is or how confident the founder sounds. It is earned against the same short list of suppliers, the same single lithography vendor, the same handful of people who know the recipes, that every other chipmaker on Earth depends on. Terafab can be the largest structure ever raised and still come down, in the end, to how much rides on how little — one Dutch machine, one Japanese chemical, one contaminated tool, one clean wafer at a time.