Terafab: What Tesla and SpaceX Actually Committed in Texas

A large industrial building part-built, steel framing and a bare concrete core standing between two tower cranes — illustrative of the phase-one stage the Terafab commitment covers. Photo by wdreblow0 on Pixabay

BUSINESS · 11 Aug 2026 —

AUSTIN, 11 AUG 2026 — Tesla and SpaceX will put more than US$16.8 billion into the first phase of a semiconductor complex in Grimes County, Texas. Elon Musk, who runs both companies, called it "the largest and most valuable building on Earth by far" in a post on X on Thursday.

The name is Terafab. The plan is to do chip fabrication, memory, advanced packaging and testing on one site, and to feed two customers who happen to be the two companies paying for it.

What was committed and what was described

These are not the same thing, and the announcement mixes them freely.

Committed: $16.8 billion, described as phase one. A site near the Gibbons Creek Reservoir, about 100 miles from Tesla's Austin headquarters and 20 miles east of College Station. At least 3,000 jobs across engineering, technician and plant-operator roles. Industrial water drawn from the reservoir rather than local groundwater, which the companies volunteered.

Described: 100 million square feet of manufacturing space, roughly ten times Tesla's Gigafactory Texas. Chips for Optimus humanoid robots and Cybercab robotaxis, plus processors for satellites SpaceX intends to use as orbital data centres. TechCrunch reports SpaceX has suggested the multi-phase build could reach $119 billion, which is seven times the committed figure and attached to no schedule.

Texas has awarded incentives; TechCrunch describes them as millions of dollars in tax breaks without itemising them.

$16.8bnPhase one. A decision has been taken.

$119bnSuggested across later phases. No schedule.

100m sq ftDesign target, about ten times Gigafactory Texas.

Over 1 TW“Computing capacity”, with no unit defined.

The number that needs a second look

SpaceX and Tesla say their combined chip demand will exceed one terawatt of computing capacity.

"Computing capacity" measured in terawatts is not a standard unit, and the announcement does not define it. Read as electrical power, which is what the unit ordinarily means, it is a very large claim. The US Energy Information Administration expects a record 86 gigawatts of new utility-scale generating capacity to be added across the entire United States in 2026. One terawatt is more than eleven such record years of national build-out, for two companies.

It may not mean that. It may be a statement about installed silicon, or a rhetorical figure. The point is that a number of that size published without a unit definition cannot be checked, and a reader should treat it as ambition rather than a forecast.

Two figures that do not sit easily together

Put the floor area and the headcount side by side. One hundred million square feet, at least 3,000 jobs. That is about 33,000 square feet of building per employee.

Modern fabs are the most automated factories on earth, and cleanroom space is not office space, so a high ratio is expected. This one is still worth pausing on. It implies either a facility far more automated than anything currently operating, or a floor area that describes an eventual multi-phase footprint while the job number describes phase one, or a job number that will grow substantially and has been stated conservatively.

All three are plausible. The announcement does not say which, and the two numbers were presented together as though they described the same thing. Quoting a project's eventual scale next to its initial hiring figure invites a misleading comparison.

Why build the fab at all

The strategic logic is more coherent than the arithmetic. Tesla's autonomy programme and its robot programme both depend on inference silicon in volume, and SpaceX's orbital compute plan depends on parts that survive launch and radiation. Neither is a product a merchant foundry's roadmap is organised around.

Buying capacity on someone else's roadmap means queueing behind their larger customers, which for advanced nodes has meant queueing behind Apple and Nvidia. A captive fab converts that queue into a scheduling decision you make yourself.

The cost of that control is the reason almost nobody does it. Fabs are the most capital-intensive factories in industry, they take years to yield, and the economics assume volumes a single buyer rarely has. The companies that gave up owning fabs — AMD among them — did so because the maths stopped working, and the maths has not become friendlier since.

The structural experiment

Consolidating fabrication, memory, packaging and test on one site is the interesting engineering claim, more so than the floor area.

The industry is organised the other way. Wafers are fabricated in one country, packaged in another, tested in a third, and the logistics between those steps are a large share of both cost and lead time. Advanced packaging in particular has been the bottleneck of the AI hardware build-out, and it sits in a small number of facilities concentrated in Taiwan and South Korea.

If it worked, a single site that does all of it would represent a structural change in the industry. It would also concentrate every one of those risks in one Texas county, in a business where a contamination event or a single equipment failure can idle a line for weeks.

Reading the announcement

Phase-one figures are the only ones with a decision behind them. The $119 billion is a description of an intention held by a private company that is not obliged to explain it, and the 100 million square feet is a design target for a facility that has not been built.

The verifiable near-term markers are ordinary and unglamorous: land acquisition, the water agreement with the reservoir authority, grid interconnection filings, permits, and equipment orders. Those arrive on public record whether the project proceeds at the announced pace or not, describing its actual size long before any silicon is produced.

What to watch

Whether Terafab intends to serve outside customers. Everything announced points at internal demand, and a fab built only for its owners is a cost centre whose unit economics depend entirely on volumes that do not exist yet.

Which process node, because that decides whether this competes with TSMC's leading edge or occupies the mature-node space where a captive fab is a much more sensible proposition.

And the water. The companies pre-empted the groundwater question by naming the reservoir, which suggests they expect the argument. Semiconductor fabrication consumes water at industrial scale and Texas has been through repeated drought, so how much Terafab is entitled to draw, and what happens in a dry year, is a term worth seeing in writing.

Eva Chin’s Verdict Business & Enterprise

Separate the committed from the described and this gets much smaller, which is not the same as saying it is not significant. Committed: $16.8 billion, a site, a water source, three thousand jobs. Described: a hundred million square feet, a hundred and nineteen billion dollars across unscheduled phases, and more than a terawatt of something the announcement never defines. A reader who takes the second list as a plan has been handed the wrong document. The claim I would want examined is the terawatt. Read as electrical power it is more than eleven record years of new American generating capacity, for two companies, and no unit definition is offered anywhere. The floor-area-to-headcount ratio deserves the same scrutiny: thirty-three thousand square feet per employee is either an extraordinary automation claim or a sign that the two numbers describe different phases. Both of those are answerable questions, and until they are answered the interesting figure remains the one with a board decision behind it.

Sources & cross-checks

Primary: TechCrunch, 6 August 2026 — for the $16.8 billion figure being described as an initial investment; the Grimes County, Texas location outside Houston; more than 100 million square feet of manufacturing space; at least 3,000 jobs from Grimes and Brazos counties; chips optimised for edge computing and inference for Tesla's Optimus robots and Cybercabs plus high-power chips for SpaceX's space-based data centres; Elon Musk's “largest and most valuable building on Earth by far” posted on X; the company line that Terafab “will be epic in both its mission and in its sheer size”; state incentives described only as millions of dollars in tax breaks; and SpaceX's suggestion that spending could reach $119 billion across a multi-phase plan beyond the announced initial investment.
Corroborated: The Epoch Times, 6 August 2026, for the site sitting near Gibbons Creek Reservoir about 100 miles from Tesla's Austin headquarters and about 20 miles east of College Station; the 100 million square feet being roughly ten times Gigafactory Texas; the companies' undertaking to draw industrial water from the reservoir rather than local groundwater; the intent to consolidate chip manufacturing, memory production, advanced packaging and testing under one roof; SpaceX describing orbital computing as the “only truly scalable solution for the future of AI”; and the statement that the two companies' combined chip demand would exceed one terawatt of computing capacity.
Context: US Energy Information Administration figures reported by Enerdata, for the EIA forecast of a record 86 gigawatts of utility-scale generating capacity additions across the United States in 2026, used here only as a scale comparison for the terawatt claim.
What could not be established: ⚠ We could not establish the itemised value of the Texas incentives, the process node, the construction schedule, the water volumes involved or the terms of any reservoir agreement, whether Terafab intends to serve external customers, or what “one terawatt of computing capacity” denotes. We could not open the Quartz or Reuters accounts directly. We sought comment from neither Tesla nor SpaceX.
Method and limits: Compiled 11 August 2026. The separation of committed from described figures, the arithmetic giving roughly 33,000 square feet per employee and the three readings offered for it, the comparison of one terawatt against a record year of US capacity additions, the argument that a captive fab converts a queue into a scheduling decision, and the list of verifiable near-term markers are RECATOOLS analysis rather than claims made by either company. Nothing here is investment advice.