On July 23, Tesla held their Q2 earnings conference. The conferece included two interesting semiconductor related tidbits:
- The AI5 hardware chip (leveraging TSMC N3) has slipped a full year, originally "very end of 2025/early Q1 2026" and now "end of next year [2026]'
- Tesla's AI6 chip is planned to be a scalable design that can be scaled down for use in humanoid robots, self-driving cars, but also scaled up to AI data center usage (i.e. potentially replacing Nvidia..)
They also mentioned they think the can improve AI inference performance on the existing silicon (HW4/AI4) by "roughly 10X" from where it is performing today. There have been rumors that Tesla has already hit performance limits on AI4 hardware, and infact Jim Keller has stated* that he thought Tesla needed another 5X-10X performance over AI4 to achieve (unsupervised) full self driving.
Context:
- No reason for the AI5 slip was given.
- Tesla used to call their self driving hardware "HW", i.e. "HW2", "HW3", etc, but now refer to it as "AIx".
- Tesla's first "2.5" generations of hardware were Nvidia-based chips, and Gen 3 was the first in-house design, led by Jim Keller
- The AI / self driving hardware is a separate system from the infotainment system (which is currently AMD Ryzen based.)
(Source webcast here if curious - https://ir.tesla.com/#quarterly-disclosure )
*Jim Keller statement is from a Lex Fridman podcast
- The AI5 hardware chip (leveraging TSMC N3) has slipped a full year, originally "very end of 2025/early Q1 2026" and now "end of next year [2026]'
- Tesla's AI6 chip is planned to be a scalable design that can be scaled down for use in humanoid robots, self-driving cars, but also scaled up to AI data center usage (i.e. potentially replacing Nvidia..)
They also mentioned they think the can improve AI inference performance on the existing silicon (HW4/AI4) by "roughly 10X" from where it is performing today. There have been rumors that Tesla has already hit performance limits on AI4 hardware, and infact Jim Keller has stated* that he thought Tesla needed another 5X-10X performance over AI4 to achieve (unsupervised) full self driving.
Context:
- No reason for the AI5 slip was given.
- Tesla used to call their self driving hardware "HW", i.e. "HW2", "HW3", etc, but now refer to it as "AIx".
- Tesla's first "2.5" generations of hardware were Nvidia-based chips, and Gen 3 was the first in-house design, led by Jim Keller
- The AI / self driving hardware is a separate system from the infotainment system (which is currently AMD Ryzen based.)
(Source webcast here if curious - https://ir.tesla.com/#quarterly-disclosure )
*Jim Keller statement is from a Lex Fridman podcast