By TeslaLearning Objectives
- Understand what Terafab is, who the partners are, and the production targets
- Identify what chips Terafab will produce and how that fits the Tesla/SpaceX/xAI stack
- Evaluate the execution risks and why Musk's ecosystem is betting on vertical integration
What Is Terafab?
Terafab is a $20-25 billion chip foundry joint venture announced by Elon Musk on March 21, 2026 at the former Seaholm Power Plant in Austin, Texas. The founding partners are Tesla, SpaceX, and xAI, with Intel joining on April 7, 2026 as the process-technology partner contributing its 18A node. Terafab's stated goal: produce 1 terawatt of AI compute output per year — roughly 50× current global AI chip production.
The name reflects the target, not just a brand — Terafab is designed to scale AI silicon supply to a level no existing foundry (TSMC, Samsung, or Intel alone) has ever attempted for a single customer group.
💡Key Concept
Why "tera-" scale matters: A modern leading-edge fab typically runs 100,000–150,000 wafer starts per month (wsm). Terafab's long-term target is 1 million wsm — roughly 5-10× the largest TSMC fabs — producing 100-200 billion chips per year on 2 nm and Intel 18A nodes.
The Announcement Timeline
| Date | Event |
|---|---|
| March 21, 2026 | Musk announces Terafab at Seaholm Power Plant, Austin — Tesla, SpaceX, xAI as founding partners |
| April 7, 2026 | Intel joins as process-technology partner, contributing its 18A node (gate-all-around, backside power delivery, EMIB/Foveros 3D stacking) |
| April 15, 2026 | Tesla's AI5 chip tapes out — powers FSD v15 and next-gen Optimus (dual-sourced at Samsung Taylor + TSMC Arizona for initial ramp) |
| April 17, 2026 | Terafab reported to be in contact with Applied Materials, Tokyo Electron, Lam Research, Samsung, and ASML for equipment supply |
| 2029 target | Pilot line operational at 3,000 wafers per month |
| Long-term | Scale to 1 million wafer starts per month; 100-200 billion chips per year |
What Terafab Will Produce
Terafab is positioned as the vertical-integration play that ties the entire Musk-ecosystem silicon roadmap together:
| Product | For | Status |
|---|---|---|
| Tesla AI5 | FSD v15 + next-gen Optimus | Taped out April 15, 2026 (dual-sourced during ramp; Terafab long-term) |
| Tesla AI6 | Optimus + Robotaxi edge compute | Tape-out targeted December 2026; Samsung Taylor exclusive |
| Dojo3 | Space-based AI compute (Musk, January 2026) | Restarted January 2026 after August 2025 Dojo shutdown; D2 declared evolutionary dead end |
| xAI training accelerators | Colossus 2 / Colossus 3 (Memphis) | Long-term non-NVIDIA supply; Colossus 2 still on GB200 near-term |
| SpaceX orbital silicon | Starlink V3 on-satellite compute + radiation-tolerant inference | Required for the 1 million orbital data-center satellites in SpaceX's January 2026 FCC filing |
Why Vertical Integration Now
Musk's bet: foundation-model-era compute demand is so large that even TSMC, Samsung, and Intel combined cannot supply one customer group. By owning the fab, Musk's ecosystem controls the full stack — chip design, fabrication, data center, ground deployment, and orbital compute.
Musk's public framing: "We either build the Terafab, or we don't have the chips, and we need the chips, so we build the Terafab."
How Terafab compares to other custom-silicon plays
| Program | Approach | Foundry | Scale Target |
|---|---|---|---|
| Terafab (Tesla/SpaceX/xAI/Intel) | Own the fab + chip design | Terafab + Samsung Taylor + TSMC Arizona (transition) | 1 million wsm long-term |
| OpenAI Stargate | Rents capacity + partners with Broadcom/Marvell/TSMC | TSMC | 400+ billion dollar infrastructure target |
| Amazon Trainium3 | Designs chips + partners with TSMC | TSMC | Internal AWS workloads |
| Google TPU v7+ | Designs chips + Broadcom partnership | TSMC | Internal Google workloads |
| Apple M-series / C-series | Designs chips only | TSMC (no fab) | Apple products only |
Terafab is unique among the above — no other frontier AI buyer is building its own foundry.
Strengths
- True vertical integration — Tesla controls silicon from design through manufacturing; eliminates dependency on TSMC or Samsung allocation
- Intel 18A partnership — Intel's process technology is leading-edge (gate-all-around transistors + backside power delivery); Terafab gets the node without the Intel-fab overhead
- Multi-customer JV structure — Tesla, SpaceX, and xAI each have different silicon needs; sharing fab capacity improves utilization
- Austin geographic concentration — adjacent to GigaTexas + existing Tesla manufacturing ecosystem
- Equipment supplier engagement — contacted Applied Materials, Tokyo Electron, Lam Research, Samsung, and ASML at "light speed" (per Musk)
- Aligned with Dojo3 + orbital compute roadmap — Terafab output feeds space-based AI data centers, the Musk-ecosystem long-term bet
Limitations and Execution Risks
- Tesla has no foundry manufacturing experience — Electrek has called the project "reeks of desperation" given Tesla's total lack of semi-fab operational history
- AI5 is roughly two years behind schedule — a warning sign that custom-silicon timelines remain brutal even for the best-resourced players
- Dojo pivoted twice in eight months — August 2025 shutdown, January 2026 restart, D2 declared a dead end; project execution is unstable
- Pilot line not operational until 2029 — Terafab output does not help Tesla's current chip supply constraints
- Capital-intensive beyond $25 billion — analysts project full build-out could exceed $5 trillion; funding plan unclear
- Competes with Intel's own foundry ambitions — Intel contributes 18A but also sells it to external customers; partnership dynamics will be tested
- Regulatory and export controls — 2 nm / 18A silicon is subject to US export controls; international deployment will be complex
- Bear case (Electrek, Jim Chanos): first-time fab operators typically underestimate yield, ramp time, and equipment conditioning by 3-5×
Tool Details
| Detail | Info |
|---|---|
| Founding partners | Tesla, SpaceX, xAI, Intel |
| Announced | March 21, 2026 (Intel joined April 7, 2026) |
| Location | Seaholm Power Plant + GigaTexas, Austin, Texas |
| Investment | $20-25 billion initial; projected $5 trillion full build |
| Process nodes | Intel 18A, 2 nm |
| Pilot target | 3,000 wafers per month by 2029 |
| Long-term target | 1 million wafer starts per month; 100-200 billion chips per year |
| Compute output goal | 1 terawatt of AI compute per year |
| Status | Under construction (pre-operational) |
Key Takeaways
- Terafab is a $20-25 billion Tesla/SpaceX/xAI/Intel chip foundry joint venture announced March 21, 2026, targeting 1 terawatt of AI compute output per year (about 50× current global AI chip production)
- Will produce Tesla AI5 and AI6, restarted Dojo3 for space-based AI compute, xAI training accelerators, and SpaceX orbital inference silicon — the silicon backbone for the entire Musk-ecosystem AI strategy
- Intel's April 2026 partnership contributes the 18A process node; equipment suppliers Applied Materials, Tokyo Electron, Lam Research, Samsung, and ASML have been engaged
- Pilot line targets 3,000 wafers per month by 2029; long-term goal is 1 million wafer starts per month and 100-200 billion chips per year
- Execution risks are significant — Tesla has no foundry experience, AI5 is two years late, and Dojo has pivoted twice in eight months; the project is under construction and pre-operational
- If Terafab succeeds, it would create the first credible non-NVIDIA frontier AI training platform at scale and reshape the chip-supply landscape for the entire AI industry
Related Tools and Topics
- Tesla FSD and Cybercab — the immediate AI5 customer
- Tesla Optimus — also AI5/AI6-powered
- Grok 4.1 — xAI's flagship model, trained on Colossus
- xAI Company Overview — context on the Colossus training cluster
- Tesla Company Overview — context on vertical-integration strategy
- SpaceX Company Overview — orbital AI data-center strategy
