Terafab: Why Musk Needs His Own Chip Factory

Transcript

From the video captions, split by chapter. Timestamps link to that moment on YouTube. Corrections are marked inline.

0:00 Hook

Every AI model, every self-driving car and every humanoid robot runs on the same scarce thing: leading-edge chips. And one company, TSMC, now takes about seventy-two percent of the world's contract chipmaking revenue. So Elon Musk's companies decided to build their own chip factory. It's called Terafab, and its long-term goal is a terawatt of AI compute every year.

In March, Musk summed up the logic in one line: "We either build the Terafab or we don't have the chips, and we need the chips, so we build the Terafab." This is Bet The Decade: what Terafab is, why chips are so hard to make, the numbers, and an honest bull and bear case. This is analysis, not financial advice.

0:49 What Terafab is

Terafab was announced on March twenty-first, twenty twenty-six, in Austin, Texas. It is a joint project of Tesla and SpaceX, which now owns xAI. In April, Intel joined as a partner. It is meant to make two kinds of chips.

One is built for edge inference: the brain inside Tesla's cars and Optimus robots. The other is a high-power chip designed for AI satellites in orbit. The unusual part is the scope. The plan puts chip design, photomasks, logic, memory, advanced packaging and testing in one place, so engineers can change a design and see the result fast.

Why bother? Musk says his suppliers, like TSMC, Samsung and Micron, are only comfortable expanding at a certain rate, and that rate is far below what his companies want. SpaceX says the combined chip demand of SpaceX and Tesla is expected to exceed one terawatt of compute, which it calls significantly larger than today's global supply. For scale: on SpaceX's October earnings call, Musk said his AI business expects roughly three gigawatts of compute online by the end of this year, and about ten more gigawatts in twenty twenty-seven.

A terawatt is a thousand gigawatts.

2:11 How a fab works, and why it's hard

A chip starts life on a polished disc of silicon called a wafer. Many chips are built on each wafer at once, layer by layer. Each layer goes through a loop: deposit a thin film, coat it, print a pattern with light, etch it, clean it, and inspect it. Repeat that for dozens of layers, and an advanced wafer can go through more than a thousand process steps.

That takes time. A wafer typically spends around three months inside the fab, longer on the most advanced processes, before packaging and testing even begin. The hardest step is printing the finest layers. That takes extreme ultraviolet lithography, or EUV.

A powerful laser hits tiny droplets of molten tin, creating light with a wavelength of just thirteen and a half nanometers. Because EUV light is absorbed by almost everything, even air, the whole path runs in a vacuum. Mirrors instead of lenses guide the light onto a patterned mask, and from there down onto the wafer. Only one company in the world makes these machines: ASML, in the Netherlands.

The most advanced versions cost hundreds of millions of dollars each. Then there's yield: the share of chips on each wafer that actually work. A single tiny defect can kill a chip, and big AI chips have more area to catch one. So the same wafer can give you mostly good small chips, but far fewer good big ones.

New processes start with lower yields, and pushing them up takes months of tuning. Even Intel, with decades of experience, talks about yield improvement quarter by quarter. And the building itself is a machine. A fab runs around the clock, inside an ultra-clean room, and it needs huge, steady supplies of electricity and water.

That's why the leading edge is such a small club. Tesla already designs its own AI chips, like the ones on this self-driving computer, but outside foundries such as TSMC and Samsung manufacture them.

4:19 The plan and the numbers

Step one is a research fab at Tesla's Giga Texas site in Austin: roughly a three billion dollar facility built for a few thousand wafers a month. It will use Intel's 14A manufacturing process. That makes Terafab a major early customer for 14A, which Intel plans to bring to high-volume manufacturing in twenty twenty-eight. Intel's CEO Lip-Bu Tan, on Intel's April earnings call: "As we look to continue challenging the status quo, I can think of no better partner than Elon Musk." The main campus goes in Grimes County, east of College Station, Texas, at Gibbons Creek, where a power plant once stood.

Process water is planned to come from the Gibbons Creek Reservoir. And SpaceX says it will bring its own power: natural gas plants and very large battery arrays, instead of relying on the grid. The first phase is budgeted at about sixteen point eight billion dollars. State tax filings sketch four phases running into the mid twenty-thirties, adding up to about one hundred nineteen billion dollars.

Musk's stated output targets are one hundred thousand wafer starts a month at first, and an ambition of one million a month at full scale. And the clock. On SpaceX's earnings call on October twentieth, Musk said the first test chip, a simple SRAM memory chip, is aimed at around May or June of twenty twenty-seven, at the Austin research fab. Then he said: "Our aspiration is to reach volume production around Q3 of 2028 at Terafab at Gibbons Creek, the main Terafab location." He thanked ASML and other suppliers for accelerating equipment deliveries, and added: "These are obviously insanely fast time frames for chip production."

6:12 The competition

So who is Terafab up against? According to TrendForce, TSMC took a record seventy-two and a half percent of global foundry revenue in the second quarter of twenty twenty-six. Samsung was second, with under six percent. TSMC plans to spend sixty to sixty-four billion dollars on capital expenditure this year alone.

And it has committed at least one hundred sixty-five billion dollars to its fabs in Arizona. Put that next to Terafab's first phase: about sixteen point eight billion dollars, spread over roughly three years. Terafab is big for a newcomer, but TSMC spends several times that every single year. Samsung is a Tesla supplier, too.

In July twenty twenty-five, Tesla signed a sixteen and a half billion dollar deal for Samsung to make its AI6 chips at its plant in Taylor, Texas. And Intel is both partner and rival. Its foundry business lost two point four billion dollars in the first quarter of twenty twenty-six, as it races to win outside customers for 14A. Terafab helps Intel make that case.

So Terafab does not replace these companies any time soon. SpaceX's own prospectus says it expects to keep sourcing a significant portion of its compute hardware from third-party suppliers.

7:38 The bull case

Here's the bull case to twenty thirty and beyond. One: demand. If AI compute keeps growing toward Musk's targets, owning chip supply could matter more than any single chip design. Two: speed.

With design, masks, chips and packaging in one place, each design loop could get much shorter. Three: partners. Intel brings a manufacturing process, and ASML is accelerating deliveries. Four: room to grow.

Four phases through the mid twenty-thirties mean Terafab can scale step by step, if the early phases work. In the bull scenario, the Austin fab proves the process in twenty twenty-seven, Gibbons Creek ramps around the end of the decade, and by twenty thirty-five Terafab is a major source of chips for robots, cars and orbital data centers. That's a scenario, not a forecast.

8:33 The bear case

Now the honest bear case, with data. First, the timeline. Research firm SemiAnalysis estimates Phase One needs about twenty months to build, seventeen months to equip, and six months to ramp. That points to volume production around mid twenty thirty, nearly two years after Musk's aspiration.

Musk himself put it plainly on the call: "we don't know if we will actually achieve this or not". Second, the deal itself. SpaceX's prospectus says Terafab is in "the very early stages", that "neither Tesla nor Intel are obligated to remain a part of the project", and that there's no assurance of hitting its objectives "within the expected timeframes, or at all". Third, experience.

Tesla and SpaceX have never run a leading-edge chip fab, and Bloomberg has pointed to Musk's history of over-promising on goals and timelines. Fourth, scale. A terawatt a year is far beyond today's entire AI infrastructure. Local reporting described it as more than ten times all current AI capacity.

Fifth, local friction. Grimes County approved a one hundred percent tax abatement in a four-to-one vote, over strong objections from residents. And in September, SpaceX sued to block the release of project records. And the power plan leans on natural gas turbines, which brings its own cost and emissions questions.

10:01 The thesis

So here is the thesis. Terafab is a bet that, in the AI decade, chip supply, not chip design, becomes the real bottleneck, and that whoever controls supply controls the pace. The vision is huge, and the partners are real. The timeline is aggressive, and the risks are written into SpaceX's own filings.

Both can be true at the same time. Here's what to watch: the first test chip from Austin around mid twenty twenty-seven; Intel 14A reaching volume in twenty twenty-eight; EUV tools arriving on site; construction progress at Gibbons Creek; and whether Tesla and Intel sign definitive agreements. Is Terafab the boldest industrial bet of the decade, or a step too far? Tell me in the comments.

This is Bet The Decade. Not financial advice.