Terafab, the chip plant of Tesla and SpaceX: what holds up in the terawatt promise

Three Words in the Middle of the Countryside

Open country beside the Gibbons Creek reservoir. A little over a hundred kilometers from Austin, less from Houston. Wide sky, flat land, not a building on the horizon.

Greg Abbott steps up to the microphones and says what will rise there: "a vertically integrated semiconductor fabrication facility." In plain terms: a chip factory that does everything in-house, from design to finished part.

Under those words sits a promise SpaceX put on paper, short enough to stick: one terawatt of computing hardware every year. A thousand billion watts of computing machines, turned out in twelve months. A terawatt a year — a unit of power, hung on a production calendar.

Three words — one terawatt a year — and beneath them, holding them up, the figures from the document SpaceX filed in May 2026: $16.8 billion at the start, $119 billion across every phase of the prototype. By August, that $16.8 billion turns out to be only the first phase. The $119 billion ceiling stays where it was.

The three words remain. Once written down, they demand something far heavier than the paper they sit on: machines, buildings, power, time. One machine at a time, one year at a time. That's where the real reckoning begins — the one paid after the signature.

The world's only door

To know whether a chip plant can exist, you don't start with the money. You start with an object: the machine that prints the circuit design onto the wafer. Only one company on earth builds it, the Dutch firm ASML. The entire supply chain waits on that one machine. ASML holds 100% of the EUV market, and with it roughly 83% of worldwide sales of lithography machines. On the other side of the counter, the customers can be counted on one hand: five semiconductor manufacturers use EUV in volume production today.

Each machine holds more than a hundred thousand parts. A leading-edge plant installs between nine and eighteen of them, and they rank among the heaviest capital-spending items any chipmaker faces. Nobody buys just one: they buy a dozen, and only then go looking for somewhere to put them.

A single price, the kind that looks tidy in a slide deck, doesn't exist. Standard EUV machines — the Low-NA generation — sell for between 150 and 200 million euros. The new generation, High-NA, runs to four hundred million. Under the same label sit different machines entirely: setting one price list beside the other is like comparing the price of a car with that of a truck.

The real constraint, though, isn't there. A two-hundred-million-euro machine is within reach of anyone who can put sixteen billion euros into a construction site — that buys eighty machines at pure list price. The trouble starts after the signature. There is already a line in front of the supplier's door, and that line moves at the pace the Dutch factory assembles a hundred thousand parts at a time, not at the pace checks get signed.

Five Years of the Whole World

In 2025 ASML shipped 48 EUV systems, for 9.6 billion euros in revenue. Not to one customer: to every customer on the planet put together. That is what the entire world knows how to build in twelve months, and it is counted in two digits. For 2026 the target is at least sixty systems. The chief executive has pointed to the capacity to ship about sixty-five in the Low-NA version alone, against forty-four the year before. The curve is climbing. It climbs about ten machines at a time.

On the other side sits a figure that circulates without anyone ever having put their name to it: a plant capable of a terawatt a year would need more than three hundred machines. Twenty-five semiconductor fabs gathered under a single roof. That is where the number stops being abstract. Three hundred machines, at the stated rate of sixty a year, add up to five years of the entire world's production diverted to one plant. At the actual 2025 pace, the years stretch to a little more than six.

This is not a question of a large factory. It is a question of a factory that, to exist in the form promised, would have to swallow everything the world knows how to build for half a decade. Phones, cars, data centers: the rest of the industry would meanwhile have to get in line and wait. Here the promise runs into the first wall that answers to no one who signs it.

The line, for that matter, already exists today. TSMC's 3-nanometer capacity is booked eighteen to twenty-four months out, and some accounts have it committed through 2028. The 3- and 5-nanometer lines are expected to run at 100% occupancy in 2026. Inside that queue sits one customer who outweighs all the others: market reconstructions put Apple in control of more than half the initial 2-nanometer allocation for 2026 and 2027, leaving the rest to Nvidia, AMD and Qualcomm. Whoever arrives later does not negotiate on price. They negotiate for a spot.

The field, meanwhile, has narrowed almost to the point of closing. On advanced logic nodes only three manufacturers remain active: TSMC, Samsung and Intel. Of the 61 new fabs counted by the U.S. Congressional research service, just fourteen are expected below seven nanometers.

Then there is a geography that weighs as much as the numbers. More than 90% of advanced chips come out of a single region. In the first quarter of 2026 the world's four leading foundries — TSMC at 73%, Samsung at 7%, SMIC at 5%, UMC at 4% — were all headquartered in Asia. It is the map that turns a Texas plant into a matter of industrial policy before it is ever a matter of the balance sheet.

Il chip che costa quanto una berlina

In March 2026, Elon Musk laid out where Terafab's computing output should end up: 80 percent to satellites carrying artificial intelligence in orbit, 20 percent to applications on the ground. Four-fifths of the silicon, in other words, is meant to leave the atmosphere. If that is indeed the destination, the plant's economics look entirely different.

The reason lies in the price per unit. A Class V rad-hard processor costs between $10,000 and over $200,000 apiece. The price of a luxury sedan, for an object that fits in the palm of a hand. It is a different industry, with different laws.

NASA measures the same thing a level up. Bringing a mission's electronics up to space standard costs between $1.5 million and $4.5 million, depending on whether the parts can be bought already qualified or need to be upgraded. That figure doesn't buy a chip. It buys the permission to use one up there.

At the other end of the scale sits mass-market electronics, the kind that lives on volume and dies on cost. A Starlink terminal cost roughly $3,000 to build in 2020; today it costs less than $600. And more than twenty thousand roll out each day, bound for customers in over a hundred and fifty countries. The silicon inside has to be cheap: that's the condition for the box to exist at all.

That cheapness, multiplied, becomes a mountain. STMicroelectronics has shipped more than 7.5 billion chipsets for Starlink, on top of the five billion antenna modules counted through the end of 2025. Billions of parts born outside the factory gates — and that is exactly the outside thread a plant built to do everything in-house aims to cut, pulling production back within its own walls.

On the car, the priciest silicon is the power silicon. The Model 3's inverter carries 24 modules, each with two 650-volt silicon carbide MOSFET transistors. When they debuted, those components cost ten times as much as conventional silicon; today, about double. The gap has narrowed, though not closed entirely: in a car sold by the millions, that "about double" weighs on every unit that rolls off the line.

On one side, a processor worth as much as a German sedan. On the other, an antenna that has to come in under six hundred dollars, whole. Two industries at opposite poles, and the bet is holding them inside the same factory. The relationship between those two worlds is the key to everything.

Time does not accept checks

There is one constraint that billions do not overcome and a partner does not get around: the years of construction. For the first module of Fab 21, near Phoenix, TSMC took about five years from excavation to the start of production. In Asia, the same job took two. The site's third fab, the one destined for the N2 and A16 nodes, broke ground in April 2025. Production at 2 nanometers in Arizona is expected in 2029; at 3 nanometers, in the second half of 2027. These are numbers on the calendar of a company that has done this work for decades.

The strain left its mark. "After a painful learning curve, we have finally connected most of the dots and know which local contractors we can work with," a TSMC executive said of the American site. Labor, rising costs, cultural barriers. The delay never came from physics. It came from people, from schedules, from the firms that had to be found in a countryside that had never seen a cleanroom. Terafab has yet to travel that curve.

Yet there is a way of working that, in the house of Musk, has already shown a different pace. Seven weeks passed between the March announcement and shovels in the ground at Giga Texas: a compression that suggests a site chosen and prepared before anyone stepped onto a stage to talk about it. It had already happened at Giga Berlin. In cases like these, the stopwatch starts once most of the race has already been run while nobody was watching.

The stated shortcut is called Intel. On April 7, 2026 the company entered the project as a manufacturing partner, to produce the processors destined for the group's robotics and data-center ambitions. Its dowry is advanced packaging, with the EMIB and Foveros 3D technologies. It shortens one item on the calendar. Only one.

Intel's role remains that of a technological and manufacturing partner: in the company's public filings, as of August 2026, there is no equity stake, no capital commitment and no contract value. And the full-scale plan is tied to a process that has yet to arrive, the 14A. If that one slips, Terafab slips with it. A risk that adds to the others, not one that replaces them.

The document SpaceX filed says it in the language lawyers use once the marketing has already gone out the front door: with Tesla there exists "a general framework" for future development, and timing, milestones and capital expenditure "have not yet been determined." Neither Tesla nor Intel is obliged to stay in the project, and the definitive agreements may never be concluded.

Wall Street ha già scelto da che parte stare

There is a number in this story that waits for no press release and asks no one's permission. It updates every evening at the Wall Street close, while out in the countryside nothing has moved yet.

Between March 2 and August 3, 2026, ASML stock climbed from $1,316 to $1,844. Over 40% in five months. The company that builds the machines which print chips is worth far more today than it was in spring, and it is worth more precisely because those machines remain few, and remain ASML's own.

TSMC, the supplier the project's logic says should be displaced, rose over the same stretch from $337 to $426. Together the two numbers say one thing: whoever holds the chokepoint has gained value, and so has whoever uses it first. The market has bought the bottleneck, and left its removal sitting where it was.

Tesla has moved the opposite way. From $371.75 on March 2 to $342.27 on August 3, with a peak of $436 in May and a fall to $311 in July. Five months of announcements, and the premium a century-defining factory should have carried with it never arrived.

A stock moves for a thousand reasons, and almost none of them have to do with a plant that does not yet exist. But three things will settle how this ends, and they will become public long before the first chip does. The orders for EUV machines, which ASML reports every quarter: those bookings are not made in silence. The node that actually ends up installed on the line. And the date the first wafer leaves the plant.

Three numbers that will arrive on their own, one at a time, and that will read themselves.

The evening before, in a county hall

On June 3, 2026, the Grimes County commissioners voted four to one for a 100% tax abatement.

Two months later, on August 5, a county hall in Texas filled with hundreds of residents. On the table lay millions of dollars in tax breaks and very little else. Twenty-four hours stood between them and the August 6 announcement.

School superintendent Sarah Borowicz rose to defend the deal. "There are defining moments in the life of a school district, and this is one of those moments."

Among the people in the seats was Robert Rose, from College Station. He raised his hand and said what he thought: homes costing more, traffic, water running out. John Boyd, a site-selection consultant, answered him: "There are no certainties in life, and there certainly are none in economic development." The room broke up that way. A concrete question, a philosophical answer.

Why a plant of one's own becomes a temptation hard to resist is explained, better than any argument, by an episode from that same year. Tesla has a $16.5 billion deal with Samsung to make AI6 chips in Taylor, and that deal runs through 2033. In 2026 a delayed prototype lot on Samsung's 2-nanometer process pushed mass production back about six months, and the blow landed on every customer of that node. When the line belongs to someone else and it slips, there is no recourse. You wait.

Putting up a building and dating a production run remain two different trades, and the second is the slower one. In the Grimes County campaign there is, for now, a figure and a plot of land. That evening, in the hall, people had asked about homes, traffic, water. It is the countryside around that waits to learn when.