Tesla and SpaceX Confirm Terafab: $16.8 Billion Texas Chip Plant

Tesla and SpaceX have locked in Grimes County, Texas as the site of Terafab, a vertically integrated chip plant with a $16.8 billion first phase, at least 3,000 jobs, and more than 100 million square feet of floor space. Governor Greg Abbott’s office confirmed the site on August 6, 2026, along with a $30 million Texas Enterprise Fund grant. Elon Musk has called it “the largest and most valuable building on Earth by far.” The stated goal is to produce over one terawatt of compute per year.

I have covered enough “world’s largest factory” announcements to be reflexively suspicious of the phrase, and I will get to the reasons for suspicion further down. But the paperwork here is real in a way that most of these press events are not. There is a named county, a named reservoir, a state incentive package with a dollar figure attached, and a governor’s office press release. That is further than most vapor-fabs ever get.

What Was Actually Announced

The August 6 announcement did three things. It named the site, it named the first-phase spend, and it attached state money to it.

The site itself is a nice piece of reuse. Gibbons Creek Reservoir used to cool a coal-fired power plant that shut down in 2018. SpaceX says Terafab will draw from that reservoir rather than pulling local groundwater, which is a direct response to a real fight happening across Texas right now over data centers and aquifers. Grimes County residents have been watching that fight closely. Whether the reservoir can actually carry a fab of this size through a bad drought year is a question nobody has answered publicly yet.

Why “Vertically Integrated” Is the Important Word

Almost every headline led with the dollar figure. The more interesting detail is the layout: logic, memory, advanced packaging, and testing all under one roof.

That is not how the industry works. A modern AI accelerator gets its logic die etched at TSMC in Taiwan, its high-bandwidth memory stacked by SK Hynix or Samsung in Korea, and the two married together in an advanced packaging step that has been the single worst bottleneck in the AI supply chain for two years running. Every one of those handoffs is a border crossing, a contract negotiation, and a queue you can get stuck behind.

Musk’s argument is that Tesla and SpaceX will need more silicon than the entire global supply chain can spare them. Optimus humanoids need cheap inference chips in enormous volume. Cybercabs need them too. SpaceX’s space-based data centers need high-power parts that survive orbit. Buying that capacity on the open market means bidding against Nvidia, Google, and Amazon for the same wafer starts, and losing. Tesla’s current chips come from Samsung and TSMC under existing contracts, and those contracts have ceilings.

The counterargument is that packaging under one roof is easy to draw and brutally hard to run. Intel has spent the better part of a decade and tens of billions of dollars learning that lesson in public.

What “One Terawatt of Compute” Means

This is the phrase doing the most rhetorical work, and it is worth unpacking because it is not a normal unit.

Chip fabs are usually measured in wafer starts per month. Terawatts measure power. What SpaceX appears to mean is a terawatt of aggregate power draw across the chips it ships in a year, which is a way of saying “roughly a thousand gigawatt-scale data centers’ worth of silicon annually.” For scale, total US electricity generating capacity sits somewhere around 1.2 terawatts. Nobody is claiming Terafab’s output will actually be plugged in all at once, and a lot of that compute is destined for orbit where the power budget works differently. But as a marketing unit it is deliberately enormous, and it should be read as ambition rather than a spec.

The Parts That Should Make You Cautious

Three things give me pause, and none of them are secret.

The number went down, not up. When this project first surfaced in March, the figure attached to it was $25 billion. The confirmed first phase is $16.8 billion. SpaceX filings point to something like $119 billion across all phases eventually, which is the number that gets quoted in excited coverage, but multi-phase totals are the softest kind of commitment in industrial construction. Phase one is what has money behind it.

SpaceX’s own IPO filing hedges. In its May filing, SpaceX described Terafab as a general framework rather than a binding arrangement, with intellectual property terms not yet finalized. Companies do not write that sentence about projects they consider locked. There is a visible gap between the podium language and the securities language, and when those two disagree, the securities language is the one with legal consequences.

Nobody has said clearly who actually runs the fab. Reporting in April suggested Tesla was not planning to operate the plant itself, with Intel lined up as the manufacturing partner. Intel has been described as contributing to Terafab, but neither company has spelled out what that means. Owning a building is not the same as owning a process node, and process expertise is the part you cannot buy with a Texas Enterprise Fund grant.

Does This Change Anything for People Buying Hardware?

Not for years, and possibly not ever in a direct way. Terafab is being built to make chips for Tesla and SpaceX, not for retail. You will not buy a Terafab GPU.

The indirect effect is the one worth tracking. If Musk’s companies stop competing for TSMC and Samsung capacity around the end of the decade, that frees up wafer starts and packaging slots for everyone else. Right now that competition is a real driver of the memory and storage price surge that has made building a PC miserable, and it is the same demand curve behind TSMC’s revenue jumping 68% on AI chips. Relief from Terafab, if it arrives, arrives in the 2030s.

The nearer-term signal is that a company already spending heavily on AI infrastructure has decided owning fabrication is cheaper than renting it. Musk’s supply chain has been visibly straining for a while, which is the same story behind the $60 billion order haul Super Micro booked partly on SpaceX business. When your suppliers become your bottleneck, you either pay more or you build.

FAQ

When will Terafab start producing chips?

No production timeline has been given. Construction has not been announced as started, and a greenfield leading-edge fab typically takes three to five years from groundbreaking to volume output. A facility of this size, with packaging and test integrated, would plausibly be at the long end of that range. Treat any date before 2030 as speculation.

Is Terafab bigger than TSMC’s Arizona plants?

On paper, dramatically. TSMC’s Arizona commitment is larger in total dollars but spread across multiple fabs and years, and no single building comes close to 100 million square feet. The honest comparison is that TSMC is shipping wafers in Arizona today and Terafab is a site announcement. Square footage is not output.

Will Terafab make chips for anyone besides Tesla and SpaceX?

Nothing announced suggests so. The described products are edge and inference silicon for Optimus and Cybercab, and higher-power parts for orbital data centers. A captive fab with spare capacity sometimes sells it, but that is a decision years away from being relevant.

What does Texas get out of the $30 million grant?

The commitment attached to the Texas Enterprise Fund award is at least 3,000 jobs, with hiring expected to draw from Grimes and Brazos counties. Terafab was also designated a qualified project under the state’s JETI program, which is the property-tax incentive framework Texas uses for large industrial builds. You can read the state’s own announcement at the Office of the Texas Governor.

The Bottom Line

Terafab is the most concrete version of a project that has been drifting between rumor and roadmap since spring. A county, a reservoir, a grant, and a governor’s signature move it out of pure speculation. What is still missing is a construction start, a process partner named on the record, and a production date. Until those exist, the right way to read the $16.8 billion is as a serious down payment on an idea, not as a fab.