Terafab will make two chip families: edge-inference silicon for Tesla vehicles and Optimus robots, and radiation-hardened processors for SpaceX's orbital data centers. Both are planned on Intel's 14A process. And both depend entirely on lithography tools that only one company on Earth makes — ASML.
👉 The one thing to understand: Terafab is not competing with TSMC for customers. It is competing with TSMC — and everyone else — for ASML's machines.
- Chip 1 — terrestrial: edge-inference processors in the AI5 lineage, for Full Self-Driving, Cybercab and Optimus.
- Chip 2 — orbital: high-power, radiation-hardened silicon for space-based AI data centers.
- The node: Intel's 14A, in what looks like a technology-licensing arrangement rather than outsourced production.
- The bottleneck: ASML is the sole supplier of EUV lithography. Its CFO confirmed in July 2026 that Terafab is factored into 2027–2028 capacity plans.
- The scale problem: one analysis calculated that equipping Terafab's logic fabs alone would take ASML roughly 70 years at its current output rate.
What Is a Terafab Chip? The Two Families
Musk was specific at the March 2026 announcement: Terafab would technically be two fabs, each producing a single chip design. That is a manufacturing decision — one design per line means no changeovers and maximum yield.
| Terrestrial chip | Orbital chip | |
|---|---|---|
| Purpose | Edge inference — running AI models on-device | AI training and inference in orbit |
| Target hardware | Tesla vehicles (FSD), Cybercab robotaxis, Optimus robots | SpaceX space-based data centers and AI satellites |
| Design lineage | AI5 and successors | Purpose-built space-hardened design |
| Special requirements | Automotive-grade reliability, power efficiency | Radiation hardening, thermal management with no air or water cooling |
| Stated volume split | Roughly 100–200 GW/year of compute | On the order of 1 TW/year of compute |
Musk framed the split in gigawatts rather than units: on the order of one hundred to two hundred gigawatts a year of terrestrial chips, and roughly a terawatt of chips destined for space. That ratio is the single most important number in the whole project.
💡 Insight: "Terafab chip" is not one product. The terrestrial chip is an evolution of silicon Tesla already ships. The orbital chip is a category that essentially does not exist commercially — which is why it needs a dedicated fab.
The Technology Stack: What Makes Terafab Different
Intel 14A, licensed rather than outsourced
Musk announced that Terafab facilities would use Intel's 14A process technology, with SpaceX responsible for high-volume manufacturing — an arrangement that reads as a technology licence rather than Intel producing the wafers. Tesla has since hired a long-serving Intel fab veteran, a signal consistent with running 14A in-house.
Everything under one roof
Terafab is designed to consolidate design, lithography, fabrication, memory production, advanced packaging and testing in a single facility — make a chip, test it, revise the photomask, repeat, without shipping wafers between sites. No existing fab operates this way.
The research fab is a physics laboratory
Musk described the Austin research fab as roughly a $3 billion initiative supporting a few thousand wafers per month, explicitly intended to test ideas about how chips are made — including new physics — rather than to fill orders.
Packaging as a first-class problem
Modern AI chips are multi-die assemblies, and packaging is frequently the real constraint. Intel's EMIB bridge technology, manufactured on US soil, was cited by analysts as a second source to TSMC's Taiwan-concentrated advanced packaging capacity.
Terafab and ASML: The Real Bottleneck
Every leading-edge chip on Earth is printed with ASML equipment. The Dutch company is the sole global supplier of EUV lithography systems, and it already has orders from every major foundry and memory maker — TSMC, Samsung, SK Hynix, Micron and Intel. Any new entrant to advanced chipmaking must buy billions of dollars of ASML tools, and join a queue.
What ASML has actually said
- May 2026 — CEO Christophe Fouquet confirmed direct conversations with Musk and called the project "very serious," while warning that projects of this type will strain equipment makers' capacity for years.
- June 2026 — Musk joined ASML's annual technology conference remotely for a fireside chat with Fouquet, with ASML signalling that Musk's team is joining the broader semiconductor ecosystem.
- July 15, 2026 — on the Q2 earnings call, CFO Roger Dassen said the 2027–2028 capacity expansion plans account for customer building plans and that "Terafab is also part of those plans".
That third item is the most consequential. A supplier building capacity around your project is a stronger signal than any press release — ASML has no incentive to expand for a customer it does not expect to order.
The 70-year problem
Here is the arithmetic that makes the terawatt target hard to take literally. Tom's Hardware modelled a Terafab logic fab as needing roughly 100 DUV and EUV lithography systems; at 126 such fabs, that is about 12,600 lithography tools for logic alone.
Against that, ASML shipped 48 EUV and 131 immersion DUV scanners in 2025 — 179 machines in total, up from 173 the year before. At that rate, equipping Terafab's logic production alone would take roughly 70 years, and only if ASML sold to no one else.
| Metric | Figure | What it implies |
|---|---|---|
| ASML 2025 shipments | 48 EUV + 131 immersion DUV = 179 systems | Global annual supply of leading-edge litho tools |
| Tools per Terafab logic fab | ~100 DUV + EUV systems | One fab module consumes over half a year of global EUV output |
| Tools for full-scale Terafab | ~12,600 for logic alone | Excludes DRAM lithography entirely |
| Time at current rate | ~70 years | Assumes ASML supplies no other customer |
| High-NA EUV shipped globally | Low single digits | Next-gen tools barely exist yet |
⚠ How to read this: the 70-year figure is not a prediction that Terafab fails. It is a demonstration that the one-terawatt goal cannot be reached with today's tool supply. Either ASML expands enormously, or Terafab's real capacity lands far below the headline.
High-NA EUV: the technology that could change the math
High numerical aperture EUV uses a 0.55 NA lens, enabling roughly 2.9 times the transistor density of current EUV systems in a single exposure. Intel is the earliest adopter, having completed acceptance testing of a Twinscan EXE:5200B at its D1X fab in Oregon in late 2025.
Fewer exposures per wafer means fewer tools per fab and higher throughput. It is the only realistic lever that bends the tool-count arithmetic — and Intel's head start on it is another reason the Intel partnership matters.
✔ For ASML watchers: Terafab is now a named factor in ASML's own capacity planning. The stock has responded to that narrative, which also means it carries timing risk if the project is delayed or scaled back.
Terafab vs TSMC: Not the Comparison You Think
Terafab is often framed as a TSMC challenger. Structurally it is not — it is a captive fab, building chips for its own owners rather than competing for external customers.
| Terafab (planned) | TSMC (operating) | |
|---|---|---|
| Business model | Captive — produces for Tesla, SpaceX and xAI | Merchant foundry — produces for external customers |
| Customers | Its own owners | Apple, Nvidia, AMD, Qualcomm and hundreds more |
| Process technology | Licensed Intel 14A | Developed in-house over decades |
| Status | Research fab under construction; production site confirmed Aug 2026 | Volume production at multiple nodes worldwide |
| Advanced packaging | Planned on-site, using Intel technology | CoWoS and SoIC, capacity-constrained and Taiwan-concentrated |
| Stated capacity goal | 1 million wafer starts/month long-term | Established global leader in leading-edge output |
The competitive pressure runs a different way. TSMC's own CEO has said publicly it will be a long time before the company can meet customer demand. Terafab does not need to beat TSMC — it needs to remove Musk's companies from TSMC's queue.
Terafab and Nvidia: Customer, Competitor, or Both?
Tesla already runs tens of thousands of Nvidia GPUs, and with Dojo cancelled, its own AI5 processors are slated for data centre duty alongside them. So Nvidia is a supplier today and a partial substitute tomorrow.
But the shared problem is bigger than the rivalry. Nvidia has itself locked up enormous supply commitments — reported at around $145 billion inclusive of inventory, purchase commitments and prepaids — precisely because leading-edge capacity is scarce.
Nvidia's CEO was also an early sceptic, warning that building advanced chip manufacturing is extremely hard before Terafab was formally announced. That scepticism and Nvidia's own supply hoarding describe the same market condition from opposite sides.
👉 The honest framing: Terafab is not an anti-Nvidia project. It is an anti-scarcity project. If it works, it reduces Musk's dependence on Nvidia, TSMC and Samsung simultaneously.
What SpaceX's Own Filing Admits
The most credible source of scepticism about Terafab is SpaceX's IPO paperwork, where risk factors are stated in plain language.
- On the shortage: orbital AI at scale requires access to AI chips "significantly more than are currently available to us", with capacity constraints named at TSMC and Samsung Foundry.
- On Terafab's odds: the filing states Terafab may not be successful, in which case sufficient AI chips may not be available from other sources.
- On the partnership: neither Tesla nor Intel is obligated to remain in the project, and definitive agreements may never be signed.
- On continued outsourcing: SpaceX expects to keep sourcing a significant share of compute hardware from third parties regardless.
Coverage of the August site confirmation made the same point: behind the price tag sits a supply problem the company has already documented for regulators.
⚠ Read risk factors carefully, not literally. IPO filings are written to disclose worst cases. But these are unusually specific — a named process dependency, a named partner exit risk, and an admission that the core goal may not be achievable.
Checklist: Signals That Terafab Silicon Is Becoming Real
- ASML tool orders. Lithography systems have multi-year lead times. Orders are the hardest evidence available.
- High-NA EUV allocation. With only low-single-digit units shipped globally, any allocation to Terafab is meaningful.
- A definitive 14A licensing agreement replacing the current framework arrangement.
- Fab process engineer hiring in Texas — staffing precedes tooling.
- Tape-out announcements for the orbital chip, which currently has no public design details.
- Packaging capacity commitments, since packaging is as likely to be the constraint as wafers.
FAQ
What chips will Terafab make?
Two families: edge-inference processors for Tesla vehicles, Cybercab and Optimus, and radiation-hardened high-power chips for SpaceX's orbital data centers.
Will Terafab use ASML machines?
It will have to — ASML is the sole supplier of EUV lithography. ASML's CFO confirmed in July 2026 that Terafab demand is included in the company's 2027–2028 capacity expansion planning.
Is Terafab a TSMC competitor?
Not in the usual sense. Terafab is a captive fab producing for its own owners, while TSMC is a merchant foundry serving external customers. They compete for equipment and talent, not customers.
What process node will Terafab chips use?
Intel's 14A for the full-scale facility, apparently under a licensing arrangement with SpaceX handling high-volume manufacturing. The original March 2026 announcement had targeted 2nm.
Does Terafab threaten Nvidia?
Only indirectly. It reduces Musk's companies' dependence on external AI chips, but SpaceX has stated it expects to keep buying substantial compute hardware from third parties.
Summary
Terafab's chip plan is coherent: two purpose-built families, one process node, everything under one roof, with packaging treated as a first-class problem rather than an afterthought. On paper the technology story holds together better than the financing story does.
The constraint is not ambition or capital — it is lithography. ASML is the only supplier, its output is measured in low hundreds of machines a year, and the arithmetic between that and a terawatt of compute does not close without either a High-NA breakthrough or a dramatically smaller Terafab.
The most telling development so far is not the site or the funding. It is that ASML has begun planning capacity around a fab that has not broken ground.
Tracking the silicon behind Optimus and orbital compute? Lithography orders will reveal Terafab's true scale long before any chip does — bookmark this page for updates as tool allocations surface.
NEVER MISS AN OPTIMUS UPDATE
We track every Tesla Optimus development — specs, deployment milestones, pricing and competitive moves — updated as news breaks.