You Can Buy the Tools. You Can’t Buy the Yield.

The announcement

Elon Musk has ruled out a role for TSMC in Terafab, the Texas chip complex SpaceX is building for itself, Tesla and xAI. His companies will build and operate it on their own. The plan is a vertically integrated site covering logic, memory and advanced packaging, with early cost estimates around $55 billion for the first phase and as much as $119 billion for the full buildout. Intel has been the reported technology partner since April, with its 14A process named as the likely node.

Musk has been blunt about the motive. Terafab, he’s said, wouldn’t exist if TSMC could keep up with his demand. TSMC’s chairman, C.C. Wei, answered with a schedule instead of an argument: there are no shortcuts in this business. A new fab takes two to three years to build and another one to two years to ramp.

Both of them are right, and the gap between their two statements is where this story actually lives.

What money buys

It’s worth being precise about how much of a leading-edge fab is purchasable, because the answer is: most of it.

Tools are for sale. Applied, Lam, KLA, Tokyo Electron and ASML will ship to anyone with an order, a site and a place in the queue. Lead times on the scarcest tools, EUV above all, are long, but that’s a scheduling problem, not a barrier.

A process can be licensed or co-developed. That’s what the Intel partnership is for. It brings a design kit, integration recipes and years of 14A development that Terafab didn’t have to fund from scratch.

Not all borrowed learning is equal, though. A process partner hands over its recipes and its ramp habits along with them. Samsung has generally been an efficient ramper, quick to move a node from development to volume, even if it came late to HBM and has had its own yield struggles at 3nm foundry. Intel is a different case. It has spent the better part of a decade slipping node schedules, from the long 10nm delay onward. Terafab isn’t just borrowing a process. It’s borrowing a ramp culture, and Intel’s is the one with the longer record of running late.

People can be hired, at least at the senior level. Experienced fab managers, integration engineers and yield leads move between companies all the time.

Demand is the piece most new entrants can’t buy, and Musk already owns it. SpaceX, Tesla and xAI are a captive customer base large enough to fill a line. That isn’t a small advantage. It’s the thing that killed GlobalFoundries’ leading-edge ambitions, which I’ll come back to.

Put all of that together and the money looks like enough. It isn’t, because one input doesn’t come with a price tag.

The learning clock

Yield is learned, and the learning is paced by how long a wafer takes to come out the other end.

At the leading edge a fab takes roughly 1 to 1.5 days to process each mask layer. The best fabs get closer to 0.8. Industry estimates put 7nm-class logic at 80 to 85 days of cycle time and 5nm-class near 100. EUV removes some multi-patterning layers, but each new node adds complexity back. A wafer started on Monday at an advanced node tells you how it did roughly three months later.

Engineers don’t sit and wait, of course. They run short-loop test vehicles that exercise a handful of steps, keep many lots in flight at once, and stagger experiments so data arrives continuously. That’s real, and it’s why mature fabs learn faster than new ones. A large captive customer helps here too, since more lots in flight means more data per cycle.

What short loops can’t give you is the integrated answer. Yield is multiplicative across every step. A die that clears 999 steps and fails the thousandth is still scrap, and the interactions that kill yield often appear only when the full flow runs together. The full-flow verdict on “did that fix work?” arrives about once a quarter. That’s four complete learning turns a year, no matter how big the budget is.

The incumbents learn on the same clock, and that’s the trap. The leader isn’t standing still while the newcomer catches up. It’s running the same quarterly cycle on a mature line, with a decade of learning already behind it, and it’s spending those quarters on the next node.

The GlobalFoundries precedent

I remember watching GlobalFoundries try to do this in upstate New York.

Construction on Fab 8 in Malta started in 2009, and production began in 2012. It was a greenfield site built specifically for the leading edge, backed by serious money and stocked with the same tools everyone else bought. On paper it had what it needed.

The ramp ran slow. Its own path to FinFETs didn’t keep pace, so GF licensed Samsung’s 14nm process and built from that, which meant leaning on someone else’s learning to stay in range. In August 2018 it stopped 7nm development altogether, citing costs it couldn’t justify. AMD, its biggest leading-edge customer, moved 7nm to TSMC. That left three companies at the frontier: Intel, Samsung and TSMC.

GF didn’t fail as a business. It pivoted to specialty processes, went public in 2021, and has been one of the largest foundries in the world by revenue. What it lost was the race the fab was originally built to win, and it lost it permanently. Leading-edge learning compounds. The leader’s yield funds its next node while the follower is still paying for the last one. Miss a node and the gap doesn’t close, it widens.

Notice who GF borrowed from. Samsung was, and generally still is, one of the faster rampers in the industry. GF leaned on that efficiency and still couldn’t close the gap on its own line. Terafab is leaning on Intel, whose recent ramp record is slower. If borrowed learning from a fast ramper wasn’t enough for GF, borrowed learning from a slow one is a harder bet.

The specific failure is worth naming, because it’s the one Terafab is designed to avoid. GF didn’t walk away from 7nm because its engineers couldn’t do it. It walked away because it couldn’t find enough customers to pay for it. Musk has solved that problem in advance. He has the customers.

What he hasn’t solved is the clock.

Logic, and some ego

Musk’s reasoning holds up better than the skeptics give it credit for. Supply security is real when your demand outruns your supplier’s allocation. A domestic supply chain is real when export controls and geopolitics keep reshaping who can build what, and where. Vertical integration is real when a single owner can align chip design, packaging and system architecture. Captive demand, as covered above, is the advantage GF never had.

The ego shows up in a single assumption: that the institutional learning curve is a bureaucracy problem and not a physics-and-statistics problem.

That assumption has paid off for Musk before. SpaceX compressed launch-vehicle development by testing hardware aggressively and accepting public failures. The method works because a rocket test gives an answer in minutes. Fly it, watch it break, fix it, fly the next one.

A fab doesn’t give you that loop. You can’t fail fast when each failure takes a quarter to show up. First-principles thinking can design a better fab layout, a smarter automation stack and a leaner organization. It can’t shorten the time a wafer spends moving through a thousand process steps.

That isn’t a criticism of the ambition. It’s a description of which part of the problem responds to Musk’s usual tools and which part doesn’t.

Where the constraint relocates

Terafab doesn’t remove Musk’s chip bottleneck. It relocates it.

Today the constraint is TSMC’s allocation queue, set by someone else’s capacity and someone else’s customer priorities. Once Terafab is running, the constraint becomes his own yield curve: how many good die per wafer his line produces, and how fast that number climbs. That trade may well be worth making. It’s still a trade, not an escape.

The timing matters for the rest of the SpaceX story as well. By Wei’s math, meaningful Terafab output is a 2029-plus event. SpaceX is targeting close to 10 gigawatts of compute in 2027 and is lining up $40 billion of Apollo-led debt to buy Nvidia chips to get there. That isn’t a contradiction, since the GPUs bridge until the fab can deliver. A lender holding those GPUs as collateral should still ask what they’re worth in a world where the borrower’s own fab succeeds and the borrower stops needing them.

Upstream sits the tool queue. EUV systems remain the scarcest equipment in the industry, and Terafab joins a line that Intel, Samsung, TSMC and every new national fab program are already standing in.

What to watch

The cost estimates will get the headlines. The signals that tell you whether Terafab is beating the clock are quieter:

• Tool-in dates, not groundbreaking dates. A building is easy. The first EUV system installed and qualified is the real start of the learning clock.
• First full-flow silicon. The date the first wafer comes out the far end of the line marks the start of real yield learning.
• How deep Intel’s role goes. Process license, integration support, or Intel engineers running the line? The deeper Intel goes, the more borrowed learning Terafab gets. Given Intel’s ramp record, that cuts both ways, so watch Intel’s own 14A milestones as closely as Terafab’s.
• Any yield or defect-density disclosure. New fabs rarely publish these. When one does, the trend matters more than the level.
• Who the first chip is for. A low-volume internal part is a learning vehicle. A high-volume AI accelerator is a bet that the curve has already been climbed.

The question

Musk can fund Terafab. That was never really in doubt. The money is the fast part of building a fab.

The slow part runs one wafer cycle at a time, roughly four complete turns a year, while the leader runs the same clock on the next node. GlobalFoundries had the money and the tools, and it still lost the race by a node.

So the question isn’t whether Terafab gets built. It’s how many learning cycles you can afford to lose before the node moves on without you.