Terafab Explained: From Origins to Now

Why it’s suddenly everywhere

1. 01

   On October 7, Musk publicly shut down rumors that TSMC would build or run Terafab, posting on X that "we will build and run the fab — let there be ZERO doubt about that," and offering TSMC only a possible sublease of part of the site.

2. 02

   Intel CEO Lip-Bu Tan confirmed the same week that Intel remains involved, reaffirming that the full-scale fab is planned around Intel’s 14A (1.4 nm-class) process.

3. 03

   In August, Tesla and SpaceX committed an initial $16.8 billion to build Terafab in Grimes County, Texas — with county filings sketching a path toward $119 billion across all phases.

4. 04

   The planned 100-million-square-foot facility would be larger than the Pentagon, Apple Park, Mall of America and Giga Texas combined, and is scheduled to start construction December 1, 2026.

The timeline

1. 2026-03-21Announced at a defunct power plant

   Musk unveils Terafab at Austin’s Seaholm Power Plant, framing it as the answer to a chip industry that "cannot expand quickly enough" for Tesla, SpaceX and xAI.

2. 2026-04Intel joins; prototype fab breaks ground

   Intel signs on to contribute manufacturing expertise; Tesla breaks ground on a prototype fab at Giga Texas to iterate chip designs without shipping wafers between sites.

3. 2026-05The 14A bet and the $119B envelope

   Musk says the full-scale fab will use Intel’s 14A process; county filings reveal an initial outlay near $55 billion and up to $119 billion across phases. SpaceX’s IPO registration cautions the deal is still a framework.

4. 2026-06Grimes County says yes

   Commissioners approve tax abatements for the former Gibbons Creek coal plant site; local debate over housing, water and transparency begins in earnest.

5. 2026-08$16.8B first phase

   Tesla and SpaceX commit $16.8 billion to phase one at Grimes County, promising at least 3,000 jobs and chips for Optimus robots, Cybercabs and space-based data centers.

6. 2026-09Construction dates filed

   Texas filings list a December 1, 2026 construction start and completion by the end of 2028 — powered by on-site natural gas and batteries, not Tesla solar.

7. 2026-10-07The TSMC clarification

   After weekend talk of TSMC involvement, Musk insists Terafab will be built and run by his own companies, floating only a sublease; Intel reaffirms its 14A role the same day.

The backstory

A fab born from a shortage that hasn’t happened yet

Terafab’s origin story is a projection. Musk’s argument, made on stage at a decommissioned Austin power plant in March 2026, is that every fab on Earth together produces only about 2% of the chips his companies will eventually need — for Tesla’s vehicles and Optimus robots, for xAI’s training clusters, and for SpaceX’s planned orbital data centers. His conclusion was characteristically binary: either build the fab, or don’t have the chips.

The scale target is what turned a supply-chain decision into a global story. Terafab aims for more than one terawatt of AI compute capacity per year — roughly 50 times the entire semiconductor industry’s current annual output of AI chips, measured by the power that compute consumes. That works out to 100–200 billion custom AI and memory chips a year at full scale, with about 80% earmarked for solar-powered AI satellites in low Earth orbit. The facility itself is planned at over 100 million square feet on the site of a retired coal plant in rural Grimes County, using the old plant’s reservoir for cooling.

Just as unusual is the structure. Instead of the industry’s decades-old division of labor — design here, foundry there, packaging somewhere else — Terafab wants chip design, wafer fabrication, memory, advanced packaging and testing under one roof. Musk has said current fabs supply only a fraction of what’s needed because the handoffs are slow; the prototype fab at Giga Texas exists precisely to let engineers “make a chip, test it, revise the mask, and repeat” without wafers ever leaving the building. It’s the same vertical-integration instinct behind Tesla’s factories and SpaceX’s rockets — the same Tesla–SpaceX engineering axis that put thrusters on the new Roadster — applied to the one component Musk has always had to buy from others. It also explains why his companies sit on the demand side of the AI memory supercycle rather than the supply side.

Why the experts are skeptical

The semiconductor industry’s response has been a raised eyebrow. Analyst Anshel Sag of Moor Insights & Strategy told BBC Chinese that the binding constraint isn’t money but equipment: leading-edge lithography tools have 18–24 month lead times, ASML’s order book is full for years, and “there’s no shortcut.” Taiwan-based semiconductor expert Dr. Yueh-Chung Chu pointed out that Musk’s talked-about idea of replacing cleanrooms with vacuum chambers would be extraordinarily expensive to build, since the chambers must withstand atmospheric pressure — and that the project seriously underestimates fab construction costs.

There’s also a timing paradox. The full-scale fab is betting on Intel’s 14A node, a process that doesn’t yet exist in volume production; by the time Terafab ramps, 2 nm may no longer be the frontier. Add a roughly 30,000-engineer shortfall in the US semiconductor workforce — TSMC’s Arizona fabs had to import hundreds of engineers from Taiwan — and the skepticism writes itself. On the other side of the ledger, Musk has a documented habit of being dismissed right up until the Falcon 9 lands.

The TSMC week

This week’s surge of attention came from a more prosaic question: who would actually run the thing? Musk had mentioned discussions with TSMC, and speculation snowballed into reports that the Taiwanese foundry — the one company on Earth that indisputably knows how to run advanced fabs — might build and operate Terafab. On October 7, Musk killed the idea in public: his companies will build and run the fab, “ZERO doubt,” and TSMC could at most sublease part of the site.

As Tom’s Hardware laid out, the offer is strange from every angle. TSMC has its own aggressive expansion plans and no need for leased space; a fab qualified for Intel’s 14A can’t casually host TSMC’s different process flow; Intel’s IP-protection requirements would make sharing a building contractually thorny; and if two chipmakers share one power and water plant, liability for a downtime event affecting hundreds of millions of dollars of wafers gets very messy. The same day, Intel CEO Lip-Bu Tan confirmed Intel is staying in — which read as the real news: the 14A partnership, not TSMC, is the technical foundation.

Meanwhile, in Grimes County

Far from the semiconductor discourse, the fab is already a local fact. Grimes County approved tax abatements in June; SpaceX has promised at least 1,800 permanent jobs (3,000 across the two counties in later company statements) and committed to drawing cooling water from the Gibbons Creek Reservoir rather than groundwater. But the county currently has 13 rental listings and about 110 homes for sale, and local realtors — briefed by colleagues in Abilene who lived through the Stargate data center buildout — are warning that the construction workforce alone will overflow into neighboring Brazos County. Hundreds of residents packed a county meeting in August over tax breaks and transparency. The school districts, at least, are on board: one superintendent called it a defining moment for her district.

Key facts

Announced

Mar 21, 2026

First-phase investment

$16.8B

Full build-out (filings)

up to $119B

Planned size

100M+ sq ft

Compute target

1 TW / year

Site

Grimes County, TX

Process bet

Intel 14A

Construction start

Dec 1, 2026

What people are saying

Common questions

Will TSMC run Terafab?

No. Musk explicitly rejected that on October 7, 2026: his companies will build and operate the fab. He floated the idea that TSMC could sublease part of the site, but industry analysts note that TSMC has its own expansion program and that sharing a fab qualified for Intel’s 14A process would be technically and contractually fraught.

Why does Musk want to build a fab at all?

His claim is arithmetic: he says existing fabs worldwide supply only about 2% of the chips Tesla, SpaceX and xAI will need for robots, robotaxis, AI training and orbital data centers. Vertical integration — design, wafers, memory, packaging and testing in one building — is meant to cut iteration time from months to weeks.

When will Terafab actually make chips?

The Giga Texas prototype fab broke ground in April 2026 and targets small-batch AI5 chip production in 2026, volume in 2027. The full-scale Grimes County plant is filed to start construction December 1, 2026 and finish by end of 2028 — though skeptics note tool lead times alone make that aggressive.

Is the 1-terawatt target realistic?

Almost nobody outside Musk’s companies thinks so on the stated timeline. One terawatt is about 50 times the industry’s current annual AI-chip output and twice the average output of the entire US grid; analysts estimate the full vision would cost hundreds of billions of dollars, not the $16.8B first phase. Most observers treat it as an aspiration that anchors a more modest real project.