CEO Elon Musk stated that both Samsung and TSMC will manufacture the chip at their U.S. facilities, with Tesla aiming for “excess supply” to deploy in vehicles, robots, and data centers, CNBC reports.
Revolutionary performance gains through custom design
The AI5 chip, first introduced at Tesla’s 2024 shareholder meeting, was described by Musk as “a stunning development” during the earnings call, emphasizing that the performance improvements stem from Tesla’s deep understanding of both hardware and software requirements. “By some measures, the AI5 chip will be 40 times better than the AI4 chip,” Musk told investors, explaining that such a significant improvement comes from the chip’s custom optimization for Tesla-specific use cases.
Architectural advancements include 8 times greater compute power, 9 times more memory, and 5 times improved memory bandwidth compared to AI4. Tesla achieved the 40-fold performance boost by removing bottlenecks present in the AI4 chip, particularly in SoftMax operations, which currently require 40 emulation steps but will execute natively in just a few steps on AI5. The chip also features built-in support for mixed-precision models and sparse tensor operations optimized for real-world AI workloads.