Elon Musk took the stage at Austin’s Seaholm Historic Power Plant on March 21 and announced Terafab—what he calls “the most epic chip building exercise in history by far.”
The plan: a $20-25 billion semiconductor facility that will manufacture AI chips for Tesla vehicles, Optimus robots, and SpaceX’s orbital AI satellites. Production target: one terawatt of computing power annually. Timeline: AI5 chips in late 2026, volume production in 2027.
There’s one problem. The companies behind Terafab have never manufactured a single chip.
What Musk Announced
Terafab is a joint venture between Tesla, SpaceX, and xAI. The facility will be built on the North Campus of Giga Texas, adjacent to Tesla’s existing Austin manufacturing base.
The scope is ambitious. Terafab aims to consolidate every stage of semiconductor production under one roof: chip design, lithography, fabrication, memory production, advanced packaging, and testing. Most semiconductor companies outsource at least some of these steps.
Two chip families are planned:
AI5: Tesla’s fifth-generation inference chip, designed for Full Self-Driving, Cybercab robotaxis, and Optimus robots. Musk claims 40-50x more compute performance and 9x more memory than the current AI4.
D3: A radiation-hardened processor for SpaceX’s orbital AI satellite constellation. Built to withstand cosmic radiation and extreme temperatures in low Earth orbit.
Musk’s rationale: “We either build the Terafab or we don’t have the chips, and we need the chips, so we build the Terafab.”
The Numbers Problem
Terafab’s production targets are extraordinary. The facility aims for 100,000 wafer starts per month and 100-200 billion chips annually.
For context, that’s roughly 70% of TSMC’s current global output—from a single location operated by companies with zero semiconductor manufacturing experience.
The technology target is equally aggressive. Terafab is aiming for 2-nanometer process technology, the most advanced node entering commercial production. TSMC is only now beginning to ramp its own 2nm output, and it has spent decades and hundreds of billions of dollars building that capability.
A single 2nm fab costs approximately $28 billion and takes around 38 months to build in the U.S. Terafab’s $20-25 billion budget would need to cover not just fabrication but also memory production, advanced packaging, and testing facilities.
The Battery Day Parallel
Industry observers are drawing uncomfortable comparisons to Tesla’s Battery Day in September 2020.
At that event, Musk promised 4680 battery cells that would halve production costs and enable $25,000 Teslas. Five and a half years later, the 4680 program has achieved roughly 2% of its original volume goals.
The pattern is familiar: revolutionary technology, dramatic cost claims, aggressive timelines—followed by years of delays and scaled-back targets.
Semiconductor manufacturing is arguably harder than battery production. The precision requirements are extreme. As one commenter noted, the facility’s planned location next to stamping machines creates vibration problems—semiconductor manufacturing requires the absence of vibrations since transistor sizes are so small that even imperceptible motion can ruin the result.
Why Now?
The timing of this announcement raises questions.
Tesla’s automotive business has declined for consecutive years, with particular struggles in Europe and China. SpaceX is preparing for an IPO. xAI recently admitted its Grok infrastructure was “not built right,” leading to a reorganization that saw half its co-founders depart.
Building a chip factory solves none of these problems in the short term. But it does generate headlines and attach struggling ventures to the AI narrative.
Musk’s stated goal—80% of Terafab’s compute output directed toward space applications—suggests this may be more about SpaceX’s satellite constellation than Tesla’s vehicles. The D3 chip for orbital use has no existing commercial supplier, which is a genuine market gap.
What Could Work
Not everything about Terafab is implausible.
TrendForce analysts suggest advanced packaging may be the best entry point. Packaging chips is less technically demanding than fabricating them, and it’s a genuine bottleneck in the AI chip supply chain.
Tesla already has some chip design experience through its Dojo supercomputer and existing AI inference chips. Vertical integration isn’t inherently unreasonable for a company of Tesla’s scale.
And Musk does have a track record of eventually delivering on ambitious technical projects—SpaceX’s reusable rockets and Starlink satellite constellation both exceeded initial expectations, albeit on longer timelines than promised.
What This Means
Terafab represents either:
- A genuine attempt to solve chip supply constraints that could give Tesla and SpaceX long-term competitive advantages
- An expensive distraction from core business problems, announced for its publicity value
- Something in between—a legitimate project that will ultimately deliver far less than promised, years later than claimed
The semiconductor industry’s response has been skeptical. Companies don’t go from zero chip manufacturing experience to 2nm volume production in 18 months. The technical, operational, and capital requirements are extraordinary.
But Terafab doesn’t need to hit its stated targets to matter. Even a partial success—a working fab producing older-node chips, or an advanced packaging facility—would reduce Tesla and SpaceX’s dependence on external suppliers.
The question is whether investors and enthusiasts can distinguish between “we’re building chip manufacturing capability” and “we’re building the largest chip facility in history.” Musk’s history suggests they often can’t.
What to Watch
If Terafab follows the Battery Day pattern:
- The 2026 “small batch production” target will slip to 2027 or 2028
- The 2nm process node will be quietly downgraded to 5nm or 7nm
- Production volumes will be a fraction of initial claims
- The project will still be described as a success
If Terafab is different:
- Equipment orders from ASML, Applied Materials, and Lam Research will be announced
- Talent acquisitions from TSMC, Intel, and Samsung will be publicized
- Third-party verification of chip performance will emerge
- Timelines will hold or slip modestly
The next 12 months will reveal which scenario we’re in. Until then, treat the terawatt claims with the same skepticism you’d apply to any Musk timeline.