# TSMC in Texas? Turning "Anchor Customer" Into an Actual Number Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/asianometry-2026-10-08-tsmc-is-coming-to-texas/ > Notes after listening to Asianometry's 2026-10-08 episode "TSMC is Coming to Texas?" — working back from Arizona's $265B to estimate how large a customer commitment a Texas fab would need, and where that chain of reasoning breaks. Educational, not investment advice; no price targets or stock recommendations. Published: 2026-10-09 Locale: en Tags: semiconductors, TSMC, supply chain, capex, Asianometry ![A half-built fab skeleton on a Texas plain at dusk, crane lights and transmission towers receding toward the horizon](/covers/asianometry-2026-10-08-tsmc-is-coming-to-texas-cover.png) > Mount Tai refuses no clod of earth, and so it became great; rivers and seas reject no small stream, and so they became deep. > > — Li Si, "Memorial Against Expelling Foreign Guests" (Qin, c. 237 BC, my translation) In the Asianometry episode "TSMC is Coming to Texas?" published 2026-10-08, host Jon Y walks through a rumor first reported by Tim Culpan in his Culpium newsletter: TSMC may build a Texas fab with Elon Musk's Terafab as the anchor customer. Jon Y says up front that he has no inside information and that the whole episode is speculation, written on October 6, 2026. He measures the idea against Arizona, where TSMC has committed $265 billion over eight to ten years across ten phases plus an R&D center and two packaging fabs — and if the company wants something like a 50% return on that capital, a Texas fab would need roughly $400 billion of revenue over its useful life, which puts an anchor customer's share at 20–30%, or $80–120 billion of purchase commitments. Two unproven premises hold that number up: the 50% return is his assumption (TSMC publishes nothing like it), and the claim that Texas will match or exceed Arizona comes from Taiwanese media, with TSMC's board yet to review the project. ## Where the rumor came from The starting point here is smoke, not a filing. Jon Y says he had been hearing rumblings about Texas for weeks, and this week the smoke got hotter: Tim Culpan wrote that Terafab might be the anchor client for a future TSMC fab in Texas, and when Musk was asked about the report on X he replied, "Just discussions, but something may come of it." Then he brings in Tim's reasoning. TSMC tends not to build new fabs unless customers are nearby, and Terafab — at least in its ambitions — qualifies. Musk has said Tesla and SpaceX will buy every chip TSMC, Samsung and Micron make for them, and even so he still needs to build Terafab, because what he wants is a terawatt of AI chips. One arrangement Tim sketches has TSMC owning the Texas fab while Terafab participates through equity or purchase commitments, along the lines of Tesla's $16.5 billion supply deal with Samsung. That all tracks. The question is the next step: how heavy does an anchor have to be? ## Turning the anchor into a number Jon Y doesn't skip that step, and this is the part of the episode I keep coming back to. Taiwan's UDN reported that the total mooted Texas investment will equal or exceed Arizona, where the running total is now $265 billion — TSMC added another $100 billion for four more phases as recently as July 2026. A company putting that much into one campus expects a visible return; he picks 50% as a working assumption, which implies around $400 billion of revenue over the fab's useful life. Then comes the jump that matters: how much of that must an anchor customer supply? He floats 20% to 30%. Run it through and Terafab is signing up for $80–120 billion of purchases. And then he brakes on his own conclusion. On top of everything else Tesla and SpaceX are doing, that is an enormous amount of cash. They spend heavily on semiconductors, but the AI9 accelerators for cars and robots cannot carry that volume. His read is that the anchor-customer story becomes plausible if Terafab also produces its own data center chips. What this section does is convert a respectable-sounding phrase — anchor customer — into a purchase order someone has to sign. In ordinary news flow, "X becomes anchor customer" reads as good news; this episode is a reminder that the phrase hides a number you can bound, and once you bound it you can see who could afford it. ## Where the chain breaks I like to ask what would falsify a chain like this, and there are three places worth flagging. The 50% return is the first. It is Jon Y's assumption; TSMC does not publish per-campus return targets. If the real hurdle is 30%, the end of the chain shrinks and the pressure on an anchor customer drops with it. If overseas fabs carry a higher hurdle — plausible, given US construction and labor costs — the number goes up. Error in this one cell propagates straight to the conclusion. Second: "Texas equals or exceeds Arizona" comes from Taiwanese media, not from TSMC. That sentence is the ruler for the whole calculation. Third: purchase commitments are only one form an anchor can take. Jon Y offers another: AVSMC, the Sony–TSMC joint venture just announced in Kumamoto, a $6 billion image-sensor project in which TSMC holds 37.8%. Sony brings customer demand and the core sensor technology; TSMC brings operational expertise. His analogy sticks with me — Sony buys the rental apartment and TSMC manages it. Japanese media reported that Sony was facing the Samsung juggernaut for the Apple iPhone account, which needs more capex, and that Sony was reluctant to commit that much cash to chips because it wants to stay fab-lite. Hence the structure. A TSMC–Terafab venture could take the same shape, with Terafab as majority owner and builder. Jon Y frowns at this version too: the current understanding is that Terafab starts with memory, and TSMC doesn't make discrete memory modules — embedded memories, yes — which limits what its expertise is worth here. Samsung or SK hynix would be the better partner for that. There is one more exit, and Tim built it himself: TSMC can move forward with a Texas plant without a committed anchor customer. Once that is on the table, the purchase-commitment arithmetic stops being a necessary condition and becomes one scenario among several. ## Why not the other explanations If the anchor math is this strained, why are the two sides talking? The episode offers reasons I find durable. One is culture. Jon Y reads Musk as wanting to rethink fabrication from first principles and take risks an older company would avoid. His example: the Free Electron Laser that appears in the Terafab rendering. He doesn't think TSMC would be first out the door with one, and his line about Musk is that the man is basically "FEL for the win." So a partnership would pair a set of inherited assumptions with a team that wants to throw inherited assumptions out. That tension weakens the technology-transfer reading of the deal: if Musk wants a clean sheet of paper, what TSMC brings is a distraction. From TSMC's side the explanation he gives is customer capture. There was a much-hyped Terafab–Intel partnership a few months ago, and for a foundry, customers are everything while a potential customer the size of Terafab comes along rarely. He reaches back for a precedent: Morris Chang built a fab for Apple, committing before Apple had signed, and he thinks the current CEO can do the same for Musk. I find that more convincing than complementary technology, because it doesn't require the two cultures to fit. It only requires TSMC to decide the customer is worth fronting. ## And Texas on its own terms One stretch of the episode sets Terafab aside and looks at Texas by itself, which I think is the most durable part, because it survives the rumor being wrong. TSMC needs a second large campus. Phoenix will keep expanding, but ten phases, two packaging fabs and an R&D fab sound like the ceiling for one site. Texas has an existing ecosystem: Austin hosts headquarters or offices for AMD, Texas Instruments, IBM, Silicon Labs, NXP and Microchip Technology, plus analog-inference startups like Neurophos and Mythic that have each raised hundreds of millions (though Jon Y doesn't think either is ready). Fabs run by Texas Instruments and Samsung have already built up a supplier base, and his verdict on Arizona is that its ecosystem was overfitted to Intel and Intel's peculiar circumstances, where Texas companies may be more flexible to work with. On talent, Texas has no state income tax, and UT Austin, Texas A&M and Rice graduate thousands of electrical engineers a year. Every TSMC person he knows brings up how much developing US talent matters. Then an off-the-record thread: he has heard that TSMC struggled with the Phoenix environment and regulatory ecosystem. Plenty has been written about finishing that fab — safety violations, allegations of anti-American discrimination, clashes with labor unions. He suspects some of it might have been avoided in business-friendly Texas, and immediately says he hasn't done the analysis to back that, with the aside that Phoenix isn't exactly Berkeley. The one concrete difference he can point to is that Texas has no state-level NEPA-style environmental review statute. Marking the strength of his own evidence like that is a large part of why I listen. He leaves two counterweights. Water and power: AI data centers are sprouting across Texas, grid interconnection queues are long enough that developers buy turbines and build their own plants, and can TSMC muscle into that queue? And local backlash: Governor Greg Abbott has a statewide moratorium on data centers in place, so the mood is tense. Fabs differ from data centers, but will the neighbors draw that distinction when a big box full of overseas arrivals making AI shows up in the backyard? ## A few other things worth remembering - Taiwan's share is being diluted. TSMC doesn't publish the share of advanced production in Taiwan; the third-party 85–90% figure dates to 2023, before Arizona and Japan came online, and Jon Y calls it old and unreliable. He substitutes a proxy from the ESG report — 85% of fab water and 90% of electricity usage in Taiwan — and treats the triangulation as a decent stand-in for wafer output. - A threshold may be close. Texas plus Arizona would mean $400–500 billion of capex into the US over eight to ten years, call it $50 billion a year. TSMC's 2026 capex guidance is $60–64 billion, up 50% from 2025's $40 billion. Unless next year soars to $90–100 billion, TSMC may soon see its first year of investing more abroad than at home. - The demand side is harder than the threshold: 75% of TSMC's first-half 2026 revenue came from the United States, and the US government wants more of that served by American fabs. - Singapore is on the list too. Bloomberg reported for the first time that TSMC has talked with Singaporean officials; Jon Y doubted the land, power and water were there until a friend showed him a large reserved plot in Tampines, near an IKEA. - Taiwan projects continue: Hsinchu Fab 20, Kaohsiung Fab 22, the Tainan Fab 18 expansion, Taichung Fab 25, and a second future Tainan fab. - A precondition hasn't been met: Bloomberg notes the Texas investment depends on lawmakers extending the advanced manufacturing tax credit from the 2022 CHIPS Act, which expires at year end. - The board has not reviewed the project. Jon Y closes by saying that under truth serum he thinks Texas moves forward — too much smoke, too compelling a client, an AI market still hot. ## Sources worth checking - Asianometry, "TSMC is Coming to Texas?", 2026-10-08 - Tim Culpan's Culpium newsletter (the original Texas and anchor-customer reporting) - TSMC ESG report (regional breakdown of fab water and electricity usage) - TSMC 2026 capex guidance and Arizona investment announcements - Announcements on AVSMC, the Sony–TSMC joint venture in Kumamoto ## One thing to take away An idea: with a large investment, read the structure of the commitment before the size of it. This episode holds up because it breaks "anchor customer" into three things you can question separately — who pays (purchases or equity), how much counts as an anchor (20% or 30%), and what single missing event stops it (the tax credit, the board, the grid queue). The dollar figure is the headline; the structure decides whether it happens. I used to get dazzled by the zeros, and then noticed the zeros carry the least information in the sentence. Something I've tried, and it works outside investing: next time someone describes a big plan — a friend starting a company, a colleague redesigning a process, family talking about moving — skip "how much will it cost" and ask three things instead. Who writes the check, what single thing failing to happen stops it, and when is the last day to decide. Put their answers in your phone and look again in a month to see which one broke first. Each time I've done it, the thing that broke wasn't the budget.