According to Dynamic Beating monitoring, Musk announced today on Twitter that Tesla's AI5 chip has completed tapeout (i.e., final chip design and delivery to the chip fab for manufacturing) and also revealed that AI6 and Dojo3 are in development. He shared a photo of the physical AI5 chip in the tweet, showing the Tesla logo on the chip package.
AI5 is Tesla's next-generation chip designed for autonomous driving and AI inference. Compared to the current in-car HW4 chip, the computing power in specific scenarios has increased by up to 40 times, and the memory has been expanded by a factor of 9. The increased memory allows the on-board system to run larger vision models, process longer spans of video data to understand changes in road conditions, which is a key bottleneck in achieving Full Self-Driving (FSD) without supervision.
According to Musk's previous statements, AI5 will be co-produced by TSMC and Samsung, with both fabs manufacturing the same design but with some differences in physical implementation. This dual-sourcing approach both mitigates supply chain risks and aims to secure production capacity. AI5 is expected to have a small number of samples by the end of 2026 and go into mass production in 2027. The initial batch will be used for the Cybercab autonomous taxi, Optimus humanoid robot, and Tesla's data center.
The emergence of Dojo3 is noteworthy. Tesla had previously put its in-house training chip Dojo project on hold and instead heavily relied on purchasing NVIDIA GPUs for AI training. Musk's mention of Dojo3 now indicates that Tesla has not abandoned the path of self-developed training chips but is simultaneously advancing on the inference (AI5/AI6) and training (Dojo3) fronts, aiming to gradually reduce dependence on NVIDIA.
