Elon Musk is accelerating construction of "Terafab," a facility to produce artificial intelligence chips in-house, complicating the calculus for Samsung Electronics (005930.KS) and SK hynix (000660.KS). If the plan to manufacture both logic chips for computation and memory takes shape, a major customer of the Korean companies could turn into a competitor. Analysts say, however, that the early phase of Terafab's operation could instead create room to expand memory supply and chip equipment exports.
Musk said at SpaceX's earnings presentation on the 20th that he was grateful to ASML and all partner suppliers for moving up equipment delivery schedules, and that the company aims for full-scale mass production in the third or fourth quarter of 2028, according to industry sources on the 22nd. Test chips are to be produced at a research facility in Austin, Texas, in May or June next year. Musk acknowledged that the timeline is highly aggressive and that there is uncertainty over meeting the target. He did not specify whether orders had been placed for ASML's extreme ultraviolet (EUV) lithography systems, which are essential to advanced chip production.
Terafab is designed to produce chips for Tesla's self-driving vehicles and humanoid robots as well as for SpaceX's AI data centers and space computing. The goal is to handle everything in one place, from chip design to logic and memory wafer fabrication, advanced packaging and testing. SpaceX plans to invest more than $16.8 billion in the first phase in Grimes County, Texas. Over the long term, it envisions scaling the site into a mega-plant processing 1 million wafers a month. The industry expects an initial production facility with capacity of 100,000 wafers a month to be in place before 2030.
Samsung Electronics' foundry business stands to be affected first. Samsung signed a $16.5 billion chip contract manufacturing deal with Tesla last year, agreeing to supply next-generation AI6 chips through 2033. Production will be handled at its plant in Taylor, Texas. If Tesla succeeds in producing chips in-house at Terafab, contract manufacturing volumes for products after the AI6 could shrink. That does not mean the existing contract will be affected immediately. If Tesla's chip demand grows faster than its own production capacity, there is a strong chance it will continue using outside contract manufacturing in parallel.
Success in in-house production hinges on securing process technology. Intel is participating in Terafab, and Musk said he would use Intel's 1.4-nanometer-class "14A" technology. Talks were also held with TSMC, the world's largest foundry, but the two sides reportedly differed over who would operate the plant. Musk intends to run the facility himself, while TSMC holds to a principle that plants using its technology must be operated by TSMC. Whether the fine process technology and stable yields needed for mass production can be secured in a short period remains unclear.
The impact on the memory businesses of Samsung Electronics and SK hynix is expected to emerge somewhat later. If Terafab also produces memory in-house, outside purchases could decline over the long term and suppliers' pricing power could weaken. High-bandwidth memory (HBM), however, requires not only DRAM manufacturing but also sophisticated packaging technology to stack and connect multiple chips. Terafab has not disclosed when or how much HBM it will produce. The target for next year's test production is also reported to be static random-access memory (SRAM) circuits for process validation rather than DRAM. If logic chip production expands first, the memory needed is likely to depend on outside supply for a considerable period.
New business opportunities could open up for Korean chip materials, components and equipment companies. Tesla and SpaceX reportedly made contact with a number of Korean firms at a semiconductor exhibition held in the United States this month. Most of those contacts, however, are at the evaluation or early discussion stage, and no orders tied to mass production facilities have been confirmed.
The key questions are when Terafab actually begins mass production and what yields it achieves. Building a plant and mass-producing advanced logic and memory chips at competitive prices are separate matters. Some observers say that, given so-called "Elon Time," full-scale mass production at Terafab could take far longer than the two years Musk has indicated. Elon Time is a term that pokes fun at Musk's habit of presenting overly optimistic schedules. It emerged as actual timelines repeatedly slipped by months or years from initial announcements whenever Musk set targets for new Tesla and SpaceX products, the completion of self-driving technology or migration to Mars.
The second-generation Tesla Roadster, unveiled in physical form only recently, is a prime example. Initially targeted for release in 2020, it was delayed several times before the actual vehicle was shown this month, six years later. The target for formal mass production and customer deliveries has been set at 2028. Some worry that even this schedule could be pushed back again under Elon Time.
"Gap World" digs into the gaps in the flood of news from the era of technology supremacy competition, seen through the eyes of reporter Seo Jong-gap. Find the core and outlook of cutting-edge technology and semiconductor issues in Gap World. Please subscribe to the Gap World column and the reporter's page.
Inquiries of any kind are always welcome. Contact the reporter by email and the topic will be followed up and reflected in a future installment.
