Elon Musk said Tesla’s AI5 chip will be manufactured by both Samsung (Texas) and TSMC (Arizona), with the goal of intentional oversupply for use in cars, robots, or data centers. He also confirmed that the AI6 chip will be made by Samsung.
Here are the key takeaways from Tesla’s Q3 2025 earnings call:
AI5 Chip Strategy
- Manufacturing Partners: AI5 will be produced by Samsung in Texas and TSMC in Arizona, both using U.S.-based fabs.
- Oversupply by Design: Musk stated, “Our explicit goal is to have an oversupply of AI5 chips.” If not used in vehicles or robots, they’ll be deployed in Tesla’s data centers.
- Form Factor: AI5 is a half-reticle chip, meaning it’s smaller than full-reticle designs like NVIDIA’s H100 or AMD’s MI300X, potentially improving yield and cost efficiency.
- Use Cases: AI5 is designed for Tesla’s Autopilot hardware, Optimus robot, and robotaxi fleet, but is flexible enough for inference in Tesla’s internal compute clusters.
AI6 Chip – The Successor
- Foundry: Musk confirmed that Samsung will manufacture the AI6 chip, continuing the partnership from the AI5 program.
- Timing: While no specific launch date was given, AI6 is expected to follow AI5’s deployment cycle, likely aligning with next-gen robotaxi and Optimus rollouts.
Tesla’s Compute Ambitions
- Tesla now has compute capacity equivalent to 81,000 NVIDIA H100s, underscoring its hybrid strategy: using in-house chips for inference and NVIDIA GPUs for training.
- Musk clarified: “We’re not about to replace NVIDIA, to be clear, but we do use both in combination.”
This dual-foundry, oversupply-first strategy mirrors how Apple diversified M-series chip production across TSMC nodes. It also hints at Tesla’s ambitions to become a vertically integrated AI inference platform, not just a car company.
Want to dig into how AI5’s half-reticle design might affect yield economics or how this compares to Groq’s SRAM-based LPU? I’d love to riff on that :Copilot
PS
Shrink n Stack tailwinds blow strong.
Leading-Edge maniac club assures it.
ASML
Village |
