Musk Confirms Terafab Talks With TSMC as Intel Remains Sole Named Partner

QbitAI reported that TSMC has not commented. Intel is the only confirmed partner and plans to supply Terafab with its 14A process, roughly a 1.4-nanometer-class technology. News of the TSMC discussions quickly drew attention because Intel and TSMC are longtime rivals, and because Terafab is not meant to be an ordinary foundry.

Terafab is a super-large vertically integrated chip manufacturing project being advanced by Musk’s Tesla, SpaceX and xAI. Its stated goal is to secure a continuous and sufficient supply of custom chips, not to take outside foundry orders. Musk has previously stressed the need to reduce reliance on TSMC-made chips. In April, he said: “TSMC just can’t make the staggeringly large number of chips needed! If they could, we would not need to do this.” The new talks therefore appear less like a simple reversal than an effort to add options as the project grows.

The scale explains why. According to QbitAI, Terafab’s first phase involves $16.8 billion in investment, with long-term potential total investment of up to $119 billion. It is planned to cover more than 100 million square feet and ultimately target about 1TW of AI compute per year. QbitAI estimated that this is on the order of 50 times current global annual AI compute output, based on Musk’s recent statement that global compute is around 20GW. No single chip factory now is designed at that scale.

Demand pressure is the main reason Musk is talking to TSMC, but risk hedging is likely another. The confirmed manufacturing route relies on Intel’s 14A process, which is still moving toward production. Risk production is expected around 2027, with volume production targeted around 2028. Terafab selected Grimes County, Texas, in August; site clearing and early earthworks have begun. Large-scale construction is planned for December 2026, with the first phase targeted for completion at the end of 2028. The two schedules are tight. Relying only on 14A could be risky for a project in which a delay in one link could push back a chain of plans, including rockets, robots and autonomous driving that await chips.

TSMC may also have reasons to look at Texas. Around the time the discussions were reported, reports said TSMC was evaluating a new production base in Texas. If TSMC wants to build another large U.S. base, it needs long-term, stable customers capable of absorbing large capacity, and Terafab’s demand would be large enough. QbitAI said the talks look less like a change in Terafab’s route than a move by Musk to keep several paths open as the project gets bigger.

Media reports described three possible forms of cooperation. Under one model, TSMC would own and operate the new plant, while SpaceX or Terafab would provide funding and long-term orders, and possibly take equity. That resembles TSMC’s JASM project in Japan and ESMC project in Germany. Under another model, SpaceX would hold a controlling stake, TSMC would take a small stake, operate the plant and bring its process technology. Under the lightest model, cooperation would be limited to technology, yield experience, talent training and operational know-how. These remain outside speculation, and Musk could still produce other plans.

Intel is already under pressure. Culpium reported that if TSMC also participates, Intel’s 14A process could be dropped because TSMC would use its own process technology to make chips. Culpium also described a mixed possibility: Intel continues to supply advanced front-end process, with 14A used at the front end, while TSMC handles back-end factory operations, yield and packaging. Terafab has plenty of demand but lacks certainty, so whoever can fill a gap may be used, and whoever can make the project more stable may get more room. For Intel, as long as 14A can win orders and expand its foundry business through Terafab, TSMC’s presence may not be the most important issue.

Musk’s underlying goal is to secure computing resources. Tesla needs chips for FSD, Optimus and Cybercab. xAI needs chips for training and inference. SpaceX has planned orbital data centers and space compute. Terafab is intended as internal chip infrastructure for Tesla, SpaceX and xAI rather than a foundry competing with Intel, Samsung and TSMC. With the Texas site chosen and construction preparations under way, the practical questions are who will run the fab, who will ramp yield, who will supply advanced process technology, how packaging will be handled and who will bear the risk if something goes wrong. Bringing Intel and TSMC onto the same field fits Musk’s approach: keep core capabilities as close as possible, and bring in the strongest partner for each remaining link.