Elon Musk this past weekend in Austin unveiled what he called "the most epic chip-building exercise in history" — a joint venture between Tesla, SpaceX, and xAI named Terafab, designed to consolidate every stage of semiconductor production under one roof. The project will rise on the North Campus of Giga Texas in a building planned to dwarf one of the largest structures on Earth, with initial costs in the $20–25 billion range and a stated goal of producing more than one terawatt of AI compute capacity per year — the majority of which is ultimately destined for space.
[HIGHLIGHT]Terafab unifies chip design, fabrication, memory production, and packaging for three Musk-controlled companies: Tesla, SpaceX, and xAI. Initial budget is $20–25 billion. Morgan Stanley estimates total capital required at $35–45 billion. First chip output is projected for mid-2028 at the earliest. Most of the compute is targeted at space-based AI infrastructure — including SpaceX's orbital data centers and Starshield-related programs.[/HIGHLIGHT]
# The Announcement: Scale, Cost, and Timing
Speaking at the unveiling, Musk made the strategic case directly: "We either build the Terafab or we don't have the chips, and we need the chips, so we build the Terafab." He cited a structural ceiling on how quickly TSMC and Samsung — the world's two leading-edge foundries — are willing to expand: "There's a maximum rate at which they're comfortable expanding. That rate is much less than we would like."
[STATS]
Initial Capex | $20–25B
Total Capital Estimate (MS) | $35–45B
Process Node | 2-nanometer
First Chip Output | Mid-2028 (est.)
Annual Compute Target | 1+ Terawatt
[/STATS]
The timing of the announcement is conspicuous. SpaceX is widely reported to be preparing for an IPO as soon as this spring, and Tesla's 2026 capital expenditure guidance does not include Terafab — meaning the project will be funded out of new commitments across the three sponsoring companies rather than absorbed into existing budgets. The market read is straightforward: Terafab is a fresh capital story being positioned in front of public investors, particularly the SpaceX cohort.
# What Terafab Will Actually Build
Unlike a traditional foundry that serves dozens of external customers, Terafab is a captive fab — every chip it produces is destined for an in-house program. Two product families will dominate the line.
## Edge-Inference Processors
The first product is an edge-inference chip optimized for Tesla's autonomy stack: Full Self-Driving in passenger vehicles, the Optimus humanoid robot program, and the Robotaxi fleet. This is the volume engine. Morgan Stanley's Andrew Percoco notes that Giga Texas alone is provisioned for 10 million Optimus units per year, requiring approximately 20 million chips — roughly six times Tesla's current chip demand across its entire automotive business. Tesla's stated long-term Optimus target of 100 million annual units would push that requirement past 200 million chips per year, more than fifty times current consumption.
## Space-Hardened High-Power Variants
The second product is the variant most relevant to the space industry: a high-power chip hardened for the radiation, thermal, and reliability requirements of orbit. These chips will go into SpaceX satellites — including next-generation Starlink and Starshield platforms — orbital data centers, and xAI compute infrastructure that is increasingly being architected to live in space rather than in terrestrial data centers.
[QUOTE]"As Tesla expands its AI ambitions across full self-driving, Robotaxi, and Optimus, access to AI compute is increasingly emerging as a potential constraint." — Andrew Percoco, Morgan Stanley[/QUOTE]
# The 2-Nanometer Process Node
Terafab's stated process technology is 2-nanometer — the most advanced commercial chip geometry in the world today, and one that even TSMC is only beginning to bring into volume production. For context, 2nm is the leading edge of the EUV lithography roadmap, and the install base of 2nm-capable equipment globally is measured in dozens of tools, not hundreds.
That choice has enormous implications.
- **Capex risk:** A greenfield 2nm fab is among the most expensive industrial projects ever attempted. TSMC and Samsung's own 2nm fabs are individually budgeted in the $20+ billion range — and they have decades of incremental process learning to draw on.
- **Yield risk:** Process yield at 2nm is famously fragile. New entrants, even those backed by extraordinary engineering talent, typically need multiple generations to reach competitive yields.
- **Equipment availability:** ASML's most advanced High-NA EUV scanners, required for the most aggressive 2nm flows, are produced at a rate of fewer than ten units per year worldwide.
Morgan Stanley's note is blunt: building this kind of capacity from scratch is "a Herculean task" because advanced logic processes are built upon "decades of prior incremental gains."
# Why Now: The Compute Bottleneck
Musk framed the macro picture starkly. He estimated that current global AI compute output is roughly 20 gigawatts per year, and that the rest of the world's annual output is "about 2% of what his companies need." The Terafab pitch is not that the existing foundry industry is bad — it is that the existing foundry industry is too small for the scale of compute Musk's portfolio companies intend to deploy.
The strategic logic compounds across the three sponsors:
1. **Tesla** needs chips at 6× current demand within a few years to staff its Optimus rollout
2. **SpaceX** needs space-grade compute at scale to populate orbital data centers and the next generation of Starshield national security satellites
3. **xAI** needs continuous access to leading-edge silicon for training and inference at scales that strain even hyperscaler procurement
A single fab dedicated to internal demand resolves three different supply-chain risks at once.
# The Space Story: Why Most of the Compute Is Going Up
The most important number in the entire announcement, from a space-industry perspective, is Musk's stated goal: produce more than one terawatt of AI compute per year, "with most of that capacity ultimately destined for space."
That single sentence reframes Terafab as a space program as much as a chip program.
## Orbital Data Centers as a Growth Vector
SpaceX has been quietly architecting a future in which substantial portions of compute infrastructure live in orbit rather than on Earth. The drivers are well-understood: limitless solar power, no terrestrial cooling water requirements, no land use, no NIMBY opposition, and direct integration with the company's existing satellite manufacturing and launch capacity. What was missing was a credible path to the silicon volumes required.
Terafab supplies that path. A captive fab producing space-hardened 2nm chips at internal cost lets SpaceX scale orbital compute on its own timeline rather than competing for foundry slots against every other AI buyer on Earth.
## Implications for the National Security Customer Base
The U.S. Department of War's interest in resilient orbital infrastructure also benefits. Space-grade chips are historically a low-volume, high-cost specialty market — a structural problem for any defense program trying to deploy proliferated constellations at consumer-electronics cadence. A captive 2nm space-rated supply chain inside SpaceX would, at scale, fundamentally reset that economic.
# Skepticism, and Musk's Response
Musk has no operating background in semiconductor manufacturing, and he has a documented history of overpromising on goals and timelines. Morgan Stanley's note emphasizes that even under aggressive build-out assumptions, first chip output from Terafab is unlikely before mid-2028. That is consistent with the historical record of greenfield fabs from established players.
Musk leaned directly into that skepticism at the event, pointing to the original incumbent dismissals of electric cars and reusable rockets — both of which Tesla and SpaceX ultimately delivered. The argument is not that Terafab will be easy. The argument is that the same first-principles vertical integration that worked twice before will work a third time.
[QUOTE]"We either build the Terafab or we don't have the chips, and we need the chips, so we build the Terafab." — Elon Musk[/QUOTE]
# What This Means for the Broader Space Industry
For the wider commercial space sector, Terafab is consequential along three dimensions.
First, it confirms that **silicon is now a strategic input** for space operations on par with launch vehicles and propellants. Companies that depend on commodity-grade compute for satellite payloads will increasingly compete with one another — and with the broader AI industry — for foundry capacity.
Second, it accelerates the **orbital data center category** from concept to capex. If even a fraction of Terafab's terawatt target lands in space, the addressable market for in-space compute services, ground stations, and high-bandwidth optical links scales dramatically.
Third, it widens the **strategic gap between SpaceX and the rest of the launch industry**. Vertical integration into chip manufacturing is the kind of move that competitors cannot easily mirror — most launch companies do not have the cash flow, balance sheet capacity, or sister companies to underwrite a $35–45 billion captive fab.
# Investment Implications
Terafab is best understood as three bets layered on top of each other: a bet on Optimus volume, a bet on orbital compute, and a bet on Musk's ability to compress a 30-year semiconductor learning curve into half a decade. Even if the project hits its most optimistic timeline, the first chip rolls off the line in mid-2028 — meaning the financial impact is a 2029-and-beyond story.
For now, the announcement itself is the move. It signals to public-market investors, particularly those underwriting a possible SpaceX IPO, that Musk's portfolio companies intend to control their own AI substrate end-to-end. Whether the execution lives up to the announcement is a separate question — and one the industry will be watching for the rest of the decade.
[FAQ]
Q: What is Terafab?
A: Terafab is a joint venture between Tesla, SpaceX, and xAI announced by Elon Musk in March 2026 to build a fully integrated semiconductor facility — covering chip design, fabrication, memory, and packaging — on the North Campus of Giga Texas in Austin.
Q: How much will Terafab cost?
A: Initial costs are $20–25 billion. Morgan Stanley estimates that meaningful chip-making capacity will ultimately require $35–45 billion in total capital investment.
Q: What kind of chips will Terafab produce?
A: Two main product families: an edge-inference processor for Tesla Full Self-Driving, Optimus humanoid robots, and Robotaxi; and a high-power, radiation-hardened variant for SpaceX satellites, orbital data centers, and xAI infrastructure.
Q: When will the first chips ship?
A: Morgan Stanley projects first output no earlier than mid-2028, even under aggressive build-out assumptions.
Q: What process node will Terafab use?
A: 2-nanometer, the most advanced commercial process geometry in the world — and one that established players like TSMC are only beginning to bring into volume.
Q: Why is Musk building this in-house instead of using TSMC or Samsung?
A: Musk argues that TSMC and Samsung will not expand fast enough to meet the AI compute demand of Tesla, SpaceX, and xAI combined, and that internal capacity is the only way to remove silicon supply as a constraint on his roadmap.
Q: How does this affect SpaceX's IPO?
A: SpaceX is widely reported to be preparing an IPO as soon as spring 2026. Terafab is a fresh capital story positioned in front of public investors, and the orbital compute angle directly supports SpaceX's growth narrative for the IPO.
Q: How much of Terafab's output will go to space?
A: Musk stated that most of the targeted 1+ terawatt of annual AI compute capacity will ultimately be destined for space — primarily orbital data centers and SpaceX satellite programs.
[/FAQ]