Buried in Musk's Terafab announcement was the most candid public statement from a major AI buyer on the state of the global foundry industry: "There's a maximum rate at which they're comfortable expanding. That rate is much less than we would like." The "they" was TSMC and Samsung — the two companies that, between them, fabricate essentially every leading-edge chip in commercial production today. The implication is the entire thesis behind Terafab: the world's foundry capacity is not growing fast enough for what AI buyers want to deploy, and the largest of those buyers are starting to vote with their balance sheets.
[HIGHLIGHT]Building leading-edge fab capacity is one of the slowest, most capital-intensive industrial activities humans do. The foundry industry's expansion rate is bounded by EUV equipment supply, talent, and yield-learning curves — none of which scale on AI investor timelines.[/HIGHLIGHT]
# How Foundry Capacity Actually Grows
The constraint Musk is describing is not a TSMC or Samsung policy choice. It is a series of physical and operational ceilings that bind every player in the leading-edge foundry industry.
## EUV Equipment Supply
The most advanced chips require Extreme Ultraviolet lithography tools manufactured exclusively by ASML in the Netherlands. ASML produces a small number of these systems per year — and the most advanced High-NA EUV variant, required for the most aggressive 2nm flows, is produced at fewer than ten units per year worldwide. Every fab on the planet competes for the same constrained equipment pipeline.
## Process Yield Learning
A new node — 3nm to 2nm, for example — does not come online at high yield. Foundries spend 18–36 months ramping yield through painstaking process optimization. There is no way to accelerate this curve through capital alone; it is bounded by experimental cycle time inside the fab.
## Capex Cycle Time
A greenfield leading-edge fab takes 4–6 years from groundbreaking to volume production. The shell construction is the easy part. Equipment install, qualification, and process tuning dominate the schedule.
## Talent
Process integration engineers capable of bringing a new node into volume are among the scarcest skill sets in the global semiconductor industry. The deep specialists exist almost entirely inside TSMC, Samsung, Intel, and a handful of equipment suppliers. Buying them in volume is a multi-year program in itself.
# Why Demand Is Outrunning Supply
The other side of the equation is the demand explosion. AI training compute requirements have grown by orders of magnitude in five years. Inference compute — the much larger long-term workload — is just beginning to scale. Robotics, autonomous vehicles, edge AI, and orbital compute compound the pull.
[QUOTE]"We either build the Terafab or we don't have the chips, and we need the chips, so we build the Terafab." — Elon Musk[/QUOTE]
Musk's specific claim — that current global AI compute output is roughly 20 gigawatts annually and that the rest of the world's output is "about 2% of what his companies need" — is in some sense an extreme statement. But the directional point is well-supported by the broader industry: every major hyperscaler is bidding aggressively for foundry capacity, lead times are stretching, and 2nm allocations for 2026–2028 are functionally spoken for.
# The Captive Fab Pattern
Terafab is not the first time a major chip buyer has built its own fab. The historical pattern is instructive.
- **Intel** has historically operated its own fabs but is now pivoting to also serve external customers
- **Samsung** runs both a captive fab business (for its own products) and a foundry business (for external customers)
- **Apple, Google, Meta, Microsoft, Amazon** have not built captive fabs — they continue to buy capacity from TSMC and Samsung at premium pricing
Terafab is closer in spirit to Apple's relationship with TSMC: a single customer with enough scale to functionally dictate roadmap. But unlike Apple, Musk's portfolio companies are vertically integrating into the fab itself rather than negotiating allocations.
# The 2nm Bet
The choice to target 2nm out of the gate is the most aggressive design decision in the announcement. There are two ways to read it.
## The Optimistic Read
Targeting 2nm avoids the strategic dead-end of building a fab that is already a generation behind by the time it comes online. It also signals that Terafab is a serious, multi-decade industrial commitment — not a stopgap.
## The Skeptical Read
2nm is the hardest node ever attempted, and the talent and equipment to build it from scratch are scarce even for incumbents. A first-time fab operator targeting the leading edge has a meaningfully harder execution path than one starting at a more mature node.
Morgan Stanley's note splits the difference: it calls Terafab "the right move" given Tesla's compute trajectory while emphasizing that the build-out is "a Herculean task" and that first chip output is unlikely before mid-2028.
# What Captive Fabs Mean for the Industry
If Terafab works — even partially — the pattern will be studied carefully by every other AI-scale operator without a captive supply path. The hyperscalers in particular will face a strategic question: continue to fund TSMC and Samsung's expansion through long-term commitments, or follow Musk into vertical integration. The capital and talent requirements are extreme, but the prize — guaranteed leading-edge supply on internal timelines — is exactly what the AI-and-robotics buildout is constrained by.
For the foundry industry, the implication is more nuanced. A handful of vertically integrated captive fabs would not displace TSMC or Samsung. But they would absorb a meaningful share of the next generation's leading-edge demand and reshape the negotiating dynamic for everyone else.
[FAQ]
Q: Why can't TSMC and Samsung expand faster?
A: Leading-edge fab expansion is bounded by EUV equipment supply (ASML produces a small number of advanced tools per year), process yield learning curves (18–36 months), capex cycle time (4–6 years from groundbreaking to volume), and the global scarcity of process integration talent.
Q: What is 2nm?
A: 2-nanometer refers to the most advanced commercial chip process geometry available today. TSMC and Samsung are only beginning to bring 2nm into volume production. It is the leading edge of the EUV lithography roadmap.
Q: Has anyone else built a captive fab at AI scale?
A: Intel has historically run its own fabs. Samsung does both captive and foundry. The major hyperscalers (Apple, Google, Meta, Microsoft, Amazon) have not built captive fabs and continue to buy from TSMC and Samsung. Terafab is the first announced AI-scale captive program from a non-incumbent.
Q: When will Terafab actually produce chips?
A: Morgan Stanley projects mid-2028 at the earliest, even under aggressive build-out assumptions.
Q: Will Terafab serve external customers?
A: No. It is explicitly a captive fab for Tesla, SpaceX, and xAI internal demand.
[/FAQ]