Terafab vs. TSMC: The Gap Between a Factory Announcement and Working Chips

Athelguard analysis · November 1, 2026

Elon Musk’s Terafab deserves the scrutiny any major semiconductor project would receive. The important question is how its capital, buildings and partnerships become a dependable supply of useful chips. Construction progress alone cannot answer that.

Our assessment: a partner-supported research line and, eventually, production for Musk’s own companies are credible objectives. Economical volume output on the proposed timetable remains unproven. Matching TSMC’s breadth, manufacturing consistency and customer relationships would be a much larger undertaking.

The project is planned for Texas, while Taiwan-based TSMC has discussed potentially using Terafab capacity. Those discussions had no announced conclusion in TSMC’s October 15 update. A collaboration could change the competitive story considerably. TSMC Q3 earnings transcript, page 9.

What is confirmed, planned and still unproven

Confirmed announcements and reported progress: SpaceX and Tesla identified Grimes County, Texas, for Terafab. SpaceX says Tesla broke ground on a smaller research precursor in Austin in April. Tesla’s October update reports continued Austin Fab construction and plans to begin equipment commissioning in 2027. That establishes an active development effort, rather than just a concept. SpaceX announcement, August 6; Tesla Q3 update, Supporting Infrastructure.

Planned scale: Texas’s August announcement puts first-phase investment above $16.8 billion. That is a phase-one figure, not the all-in price of the entire vision or proof that the money has already been spent. The project’s website describes 100 million square feet and the integration of logic, memory and advanced packaging. Manufacturing floor area should not be read as cleanroom area or productive capacity. Texas Governor’s announcement; Terafab project website.

The project’s 1 TW/year output ambition also needs careful handling. It is a power-based description of proposed compute-hardware output, not the factory’s electricity demand or a wafer count. A useful comparison with TSMC needs assumptions about chip power, die size, process, product mix and yield. Terafab project website; SpaceX Q3 transcript, page 9.

Not established by the public evidence reviewed: a settled operating structure, qualified commercial products, sustained manufacturing yields, competitive unit costs, or a demonstrated path to the full advertised scale. These are the evidence gaps that matter.

The clock starts before the first useful product

On SpaceX’s October 20 earnings call, Musk described an internal target for a functional SRAM-type test chip in Austin in May or June 2027, followed by possible volume production at Gibbons Creek in the third or fourth quarter of 2028. He explicitly attached substantial uncertainty and called them “very aggressive dates.” These are management targets, not delivery commitments. SpaceX Q3 transcript, page 9.

A working test chip would be meaningful progress. It would show that part of the manufacturing flow works. A product still has to meet its performance, power and reliability requirements, then do so repeatedly across production lots. Treating a test structure as proof of an economical AI processor would skip several crucial milestones.

TSMC’s own Arizona expansion illustrates the gap: its second fab’s structure was completed in 2025, while N3 volume production is targeted for the second half of 2027. That is a useful benchmark, not a universal minimum schedule. Even an experienced operator distinguishes a finished building from productive capacity. TSMC Arizona project update.

Why the building is only part of the challenge

Chipmaking requires tightly controlled, repeated processing across many layers. ASML describes demanding alignment precision, controlled cleanroom conditions and different lithography systems for different layers. Buying sophisticated equipment is necessary, but integrating the complete manufacturing flow is the central challenge. ASML’s manufacturing explainer.

The decisive measure is yield: how much output meets the required specification. A plant can process wafers while producing too few usable chips to justify its cost. More floor space does not cure defects, unstable processes or equipment downtime. An accelerated build can still leave a long learning curve inside the completed building. Intel’s explanation of yield.

Equipment availability is another constraint. On October 14, ASML said its planned supply for 2027 and 2028 would not fully meet customer demand, despite expansion plans. That does not establish Terafab’s allocation, but it makes confirmed delivery and installation slots important evidence. Ability to pay and ability to obtain tools on schedule are different tests. ASML Q3 results.

Nor does fabrication finish the job. Silicon must be assembled, packaged and tested before shipment. Intel’s manufacturing overview describes electrical, thermal and functional checks after fabrication. For a project promising logic, memory and packaging together, the assessment question is whether these interdependent operations mature together, rather than whether they share an address. Intel’s assembly and test explainer.

Our practical questions would include the availability of experienced process engineers, reliable power, ultrapure water, specialty materials, equipment servicing and qualified packaging. A project-specific answer requires delivery schedules and operating plans; a large investment announcement cannot substitute for them.

TSMC’s advantage extends beyond its factories

TSMC was founded in 1987. In 2025, it manufactured more than 12,000 products for 534 customers, according to its annual report. Its advantage reflects accumulated process knowledge, production experience and a broad customer ecosystem, not simply the size of its buildings. TSMC 2025 annual report.

That distinction creates two different tests. A captive fab supplying a limited set of internal products can succeed without matching everything TSMC offers. Becoming a general-purpose foundry means supporting outside customers with dependable processes, design support, confidentiality and delivery discipline across changing product requirements.

Internal demand could help Terafab plan capacity and justify investment in supply security. It would also concentrate the business case in its owners’ product forecasts. The relevant commercial question is how much capacity remains useful if those forecasts arrive later, or in a different mix, than expected.

The operating agreement could change the assessment

Who owns the building and who controls chip production are different questions.

In October, TSMC CEO C.C. Wei confirmed that teams had evaluated Texas and discussed using Terafab capacity. He also said a TSMC fab would have to be operated by TSMC. Separately, Reuters reported Musk’s position that his companies would build and run Terafab, with TSMC potentially subleasing space. These statements do not establish an agreed arrangement or necessarily rule out separate operations at one site. TSMC Q3 transcript, page 9; Reuters report, October 7.

Our inference is straightforward: capacity operated by an established foundry would be a different proposition from independently recreating that foundry’s capabilities. A partnership could improve feasibility while leaving Terafab dependent on external process expertise. The final process, intellectual-property and operating agreements therefore deserve more attention than the project’s headline size.

Intel confirmed its own Terafab partnership in April and expressed optimism about improving manufacturing efficiency. That gives the project a potential route to established expertise, but does not disclose a complete process-license contract or validate future yields. Intel is also an interested commercial participant. Intel Q1 prepared remarks, page 2.

What the industry reaction tells us

Asked about Terafab in April, Wei emphasized technology, manufacturing execution, trust and service. He described two to three years to build a fab and another one to two years to ramp it, saying “there are no shortcuts.” That is an experienced incumbent’s assessment, not a physical law or independent verdict. It nevertheless explains why the proposed 2028 ramp warrants skepticism. TSMC Q1 transcript, page 9.

The public reactions leave room for both improvement and execution risk. Intel sees an opportunity; TSMC stresses operating discipline; Musk himself acknowledges uncertainty. None of those positions is evidence that the final outcome is settled.

The milestones that should determine confidence

Timing, investment, yield and output carry different risks. A delayed ramp can increase costs even as it reduces near-term production. Each claim deserves its own evidence rather than a blanket confidence adjustment.

We would instead increase confidence as these milestones become verifiable:

1. Process and operator: named technology partners, clear responsibilities and usable process rights.
2. Production readiness: commissioned utilities, installed tools and a staffed operating team.
3. Qualified silicon: useful products meeting their intended specifications, beyond a first test chip.
4. Repeatable economics: sustained yields, packaged shipments and credible cost per usable chip.
5. Evidence-backed expansion: subsequent phases supported by funding and demonstrated demand.

Athelguard’s assessment: Terafab could create useful domestic capacity without becoming a replacement for TSMC. Its strongest near-term case is a focused, partner-supported manufacturing effort. The grander claims deserve conditional confidence until capital and floor space become qualified processes, good chips and repeatable shipments. That standard leaves room for genuine success without treating ambition as an operating result.

Evaluating a technology company, vendor or major infrastructure claim? Explore Athelguard’s independent technical assessments to identify the evidence that should come before commitment.

This assessment uses public information available on November 1, 2026. It is a technical and commercial analysis, not a recommendation to buy or sell securities.

Author’s Note

I believe Elon has great ambitions. Like many people with big imaginations, we can sometimes be too enthusiastic about how quickly our ideas will become reality. My intention is not to say this is impossible, but to caution against convincing ourselves that the process will be quick. There is no substitute for time, patience, and hard work. Willing it to happen is not enough.

If TSMC and Elon Musk’s SpaceX turn these discussions into a successful collaboration, it could become one of the semiconductor industry’s most consequential partnerships.

— William Wright