Terafab is a planned, vertically integrated semiconductor fabrication plant (fab) jointly developed by Tesla, SpaceX (and involving xAI/Intel), not a traditional commercial company offering retail products or services to the public.
Its core purpose is large-scale manufacturing of advanced chips to meet the massive internal demand from Tesla and SpaceX that external foundries cannot supply. It consolidates the full semiconductor production chain under one roof: chip design, lithography/fabrication, memory production, advanced packaging, and testing.
Key products it will manufacture
- Logic/AI inference chips optimized for edge computing (e.g., Tesla’s AI5 and successor chips).
-Memory modules and multi-chip modules. - High-power chips designed for harsh environments.
These will primarily support:
- Tesla Optimus humanoid robots and Cybercab (robotaxi/self-driving vehicles) for on-device AI inference.
- SpaceX’s planned space-based/orbital AI data centers and AI spacecraft.
The facility targets production of over **1 terawatt of AI compute capacity per year** (far exceeding current global output for similar needs), with full-scale capacity potentially reaching hundreds of billions of custom AI and memory chips annually once ramped.
A smaller prototype/research fab is already underway at Tesla’s Giga Texas (North Campus) in Austin for rapid iteration, while the main multi-phase facility (initially ~$16.8 billion investment, potentially scaling much higher) is planned for Grimes County, Texas, and is designed to become one of the world’s largest buildings (>100 million square feet).
In short, Terafab’s “products and services” center on producing specialized AI semiconductors at unprecedented volume for Tesla and SpaceX’s robotics, autonomy, and orbital compute ambitions rather than selling finished goods or external foundry services on the open market. The project remains in the planning and early construction phases as of August 2026.
The scale is driven by the extreme production targets and the physical realities of advanced semiconductor manufacturing, not just raw automation. Even with full 24/7 automation (which modern fabs already use extensively via robots, automated material handling, and continuous operation), the footprint remains enormous because of equipment size, process complexity, and the need for massive parallel capacity.
Core production goals that force the size
Terafab targets over 1 terawatt of AI compute capacity per year at full scale. That translates to long-term goals of up to 1 million wafer starts per month and 100–200 billion custom AI/memory chips annually.
Most of that output is intended for SpaceX’s orbital AI data centers/spacecraft (~75–80%), with the rest for Tesla’s Optimus robots, Cybercabs, and related edge inference. Current global advanced chip production is only a small fraction of what’s projected to be needed, so the facility is designed for captive, ultra-high-volume output that no existing foundry network can match.
Why automation alone doesn’t shrink it dramatically
- Tool size and throughput limits: Extreme ultraviolet (EUV) lithography scanners and other process tools are room-sized (or larger), extremely expensive, and have limited wafers-per-hour throughput. Achieving high volume requires hundreds of parallel tools. Cleanrooms must accommodate them with precise spacing for vibration isolation, laminar airflow, and contamination control.
- Hundreds of process steps: Each wafer goes through hundreds of sequential steps (deposition, etch, implant, etc.). High volume demands wide process lines and buffer space.
- Full vertical integration: Unlike a pure logic foundry, Terafab combines chip design/iteration, logic fabrication, memory production, advanced packaging, and testing under one modular complex. Packaging and test alone add substantial floor space. The co-location enables rapid “make–test–revise mask” loops without shipping wafers between sites—an intentional design for recursive improvement that conventional split supply chains lack.
- Support infrastructure: Ultra-pure water systems, chemical delivery, multi-gigawatt power distribution, HVAC for massive cleanrooms, waste treatment, and material storage take huge additional area. Cleanrooms themselves are typically only a portion of total facility space.
- Modular layout, not one continuous floor: The ~100–110 million square feet is the total manufacturing space across phased modules. This allows independent ramp-up, better vibration control for lithography, and linear material flow. Stacking many production floors is avoided for structural and cleanliness reasons.
Real-world comparisons
Leading-edge fabs already operate at high automation levels yet remain vast:
- A single high-volume TSMC-style advanced fab (tens of thousands of wafers per month) often involves several million square feet of total floor space, with cleanrooms measured in tens of thousands of square meters.
- Scaling to ~1 million wafers/month (roughly 10–50× a large existing advanced fab, depending on the node and mix) requires proportional expansion of tools and cleanroom capacity.
- Existing mega-campuses (TSMC, Samsung, Intel) for far lower volumes already span millions of square feet and hundreds of acres. Terafab’s target is an order of magnitude beyond those for both wafer volume and the added memory/packaging scope.
Musk and project analyses have indicated that ~100 million square feet is the right order of magnitude for the compute/chip volume goals under current and near-term process technology (targeting advanced nodes such as Intel 14A). Efficiency gains and denser packaging help, but they do not eliminate the need for parallel physical capacity at this extreme scale.
In short, the size is a direct consequence of aiming for terawatt-scale AI silicon output with full vertical integration and rapid iteration. A substantially smaller facility could support more modest volumes, but it would fall far short of the stated needs for Optimus, robotaxis, and especially orbital compute. The project is phased (research fab first at Giga Texas, then the large Grimes County complex), so the full footprint will be built out over years as capacity ramps.
At full scale, Terafab targets 100–200 billion custom AI and memory chips per year.
This is the long-term figure repeatedly cited by Elon Musk and project descriptions. It combines:
- AI/logic chips (edge-inference processors such as AI5/AI6 successors for Tesla Optimus robots, Cybercabs/FSD vehicles, and high-power D3-style chips optimized for SpaceX orbital AI data centers/spacecraft).
- Memory chips/modules (including DRAM and related high-bandwidth memory needed to pair with the AI silicon).
No official public breakdown separates the exact counts of pure AI/logic chips versus memory chips. The total is presented as a combined output of “custom AI and memory chips.” Some independent analyses estimate that high-performance logic/compute chips form a smaller portion of the unit count (potentially on the order of billions for the 1 TW compute target), while memory chips make up a larger share of the volume because systems require many more memory devices per AI processor.
Supporting capacity numbers
- Wafer starts: Up to 1 million per month at full scale (earlier phases start much lower, e.g., research/pilot fab aiming for thousands to ~100,000 wafers per month).
- Compute output: Over 1 terawatt of AI compute capacity per year (roughly 20–25% for ground-based Tesla uses such as Optimus and ~75–80% for space-based systems).
These volumes are far beyond current global advanced-node production and are the reason the facility is designed at such extreme physical scale with logic, memory production, packaging, and testing integrated together. The numbers remain long-term targets; actual ramp-up is phased over years (research fab volume targeted around 2027, larger phases into the late 2020s and early 2030s).
Terafab is not a standalone company with a single parent. It is a joint semiconductor fabrication project primarily between Tesla and SpaceX (with xAI integrated under the SpaceX umbrella following its acquisition). Intel participates as a key technology and manufacturing partner (providing process technology such as its 14A node and expertise in design, fabrication, and packaging), but it is not described as a co-owner in the same way.
- Tesla is leading the smaller research/prototype “Advanced Technology Fab” at Giga Texas in Austin.
- SpaceX is taking the lead on the initial phases of the large-scale facility in Grimes County, Texas.
- Official descriptions and filings refer to it as a collaborative effort or “general framework” between the companies. Some details (exact financial splits, intellectual property arrangements, and long-term commitments) remain subject to further negotiation and are not fully binding in every respect according to SpaceX’s earlier SEC disclosures.
It is not publicly traded on its own.
There is no independent Terafab stock or entity that investors can buy shares in. Indirect exposure comes mainly through:
- Tesla (TSLA), which is publicly traded.
- SpaceX, which remains private (it has filed an S-1 in connection with a potential IPO, but as of the latest information it is not yet publicly listed).
- Intel (INTC), which is publicly traded and involved as a partner.
Cost estimates have evolved and are phased:
- The most recent confirmed figure (August 2026 announcement) is an initial investment of $16.8 billion by Tesla and SpaceX for the Grimes County site and first phase.
- Earlier projections included roughly $20–25 billion at announcement, then SpaceX filings citing an initial ~$55 billion and a multi-phase total of up to $119 billion.
- A smaller research fab at Giga Texas has been described in the low-single-digit billions (around $3 billion in some comments).
- Analyst estimates for reaching the full long-term ambition (1 TW of annual AI compute) have ranged much higher in some cases, but the concrete project numbers currently discussed center on the tens to low hundreds of billions across phases.
In short, Terafab is a multi-company collaboration rather than a subsidiary of one parent, it has no independent public listing, and its cost is structured as large, multi-phase capital investments starting at $16.8 billion for the current first phase.