First Tesla Will Recursively Improve Chip Masks for AI5, AI6 and Other Chips

by

TERAFAB project, which will be done jointly by @SpaceX and @Tesla. The goal is to produce over a TERAWATT of compute per year (logic, memory & packaging) with ~80% for space and ~20% for the ground.

The goal is extreme vertical integration to produce custom AI chips (logic, memory, and advanced packaging) at massive scale: ~1 terawatt (TW) of annual compute capacity — roughly 50–100× current global AI chip output combined. Roughly 80% of output targets space-based orbital compute (solar-powered satellites with no day/night cycle), 20% for Earth use (Optimus robots, robotaxis, FSD, xAI training). It starts with an “advanced technology fab” near Giga Texas for rapid iteration, with the full mega-scale facility to follow.

Tesla and Elon have engineering control of the Samsung semiconductor fabrication line in Texas.

Timeline to Setup and Start Speeding Up Chip Design

Project launch & construction start

Right now (March 2026). Groundbreaking and early equipment installation are beginning immediately.
Initial operational capability (recursive loop active) could be Expected late 2026 (roughly 8–9 months from announcement). This is when small-batch production of the next-gen AI5 chips begins, allowing the closed-loop mask/chip/test iteration to run in real time.

First meaningful design acceleration could be late 2026 → early 2027. Once the loop is live, new chip generations can be designed, fabricated, tested, and improved in days/weeks instead of months — Elon estimates ~10× faster iteration than the rest of the industry.

Volume production ramp could be 2027. New designs would be sent to Samsung and TSMC for production.

The Recursive Loop on Chip Masks

A rapid iteration loop for chip design will accelerate the work on AI5, AI6, D3 and other chips. Even if initially those chips are made by TSMC and Samsung.

Traditional semiconductor iteration is slow and fragmented

Design chip → Order photomasks from specialized vendors (long lead times) → Ship masks to a foundry (TSMC) → Fabricate wafers → Package/test → Ship results back for redesign.
Each cycle takes months, with high cost and geopolitical risk.

Terafab flips this into a closed-loop, single-building system
On-site photomask (reticle) manufacturing equipment.
Full wafer fabrication (targeting leading-edge nodes, starting at 2 nm-class).
Immediate on-site testing + advanced packaging (logic + memory integrated).
Use test results to redesign → fabricate improved masks → produce/test new chips — all without leaving the building.

Innovation for Iteration velocity is the killer feature

AI software improves daily. Hardware has historically lagged (new process nodes every 2+ years). This single-building loop compresses feedback from quarters/years to days/weeks — the semiconductor version of SpaceX’s rapid rocket prototyping.

Full vertical control + customization. Chips can be tailored exactly for Tesla edge inference (Optimus/FSD), SpaceX radiation/thermal-hardened orbital compute (chips run hotter in vacuum for better Stefan-Boltzmann heat rejection), and xAI training. No reliance on external foundries.
Self-improving system potential. Once running, the facility can use its own output to accelerate further gains — a hardware “recursive self-improvement” loop analogous to what people speculate about AGI software.

Physics-pushing freedom. Elon explicitly said the fast loop lets them experiment with “wild and crazy” things (new materials, architectures, packaging) that would be too risky in a traditional supply chain.

Integrating mask-making (extreme UV precision) with leading-edge fab + packaging + testing in one site has never been done at this scale. Equipment (ASML EUV tools) has multi-year lead times; yields start terrible and ramp slowly.

1 thought on “First Tesla Will Recursively Improve Chip Masks for AI5, AI6 and Other Chips”

  1. Brian do you think the announced shortage of Helium will have a significant impact on chip manufacturing?

Comments are closed.