Key Takeaways
1. The new Tesla AI5 chip boasts ten times the performance per dollar and three times the performance per watt compared to Nvidia’s AI chips.
2. The AI5 chip outperforms the existing AI4 computer by 40 times in certain benchmarks and is set to be produced at Samsung and TSMC facilities starting in 2026.
3. Tesla’s AI5 chip will have a performance range of 2000–2500 TOPS, focusing on simplicity by removing unnecessary components for their specific needs.
4. The AI5 chip is designed as a GPU for AI inference, enabling more efficient integration and eliminating the need for certain subsystems found in previous models.
5. Tesla is refining the AI5 chip for its exclusive use in self-driving cars and Optimus robots, with an emphasis on reasoning capabilities for improved decision-making in navigation.
The new Tesla FSD computer is said to have ten times the performance per dollar compared to Nvidia’s AI chips, and it can achieve three times the performance per watt, as stated by Elon Musk.
Performance Boost
In certain benchmarks, the Tesla AI5 chip outperforms the existing AI4 computer found in the 2026 Model Y by a staggering factor of 40. Overall, it is ten times quicker than the current hardware powering FSD 14.
The AI5 chip will be produced by Tesla at Samsung and TSMC facilities located in Texas and Arizona, with manufacturing set to begin in 2026.
Specifications Revealed
Previously leaked specifications for the AI5 chip indicated a performance range of 2000–2500 TOPS (trillion operations per second), which is about five times more than the existing HW4 computer. However, Musk mentioned that he has been spending most weekends collaborating with the AI team to enhance the AI5 design.
Instead of crafting complex AI chips like Nvidia’s offerings, such as the DGX Spark AI desktop available on Amazon, Tesla has opted to simplify the design by removing unnecessary components. Since they only need to meet their own requirements, Musk explained:
“If you’re unsure about how much data will travel between each logic block on the chip, it leads to huge highways everywhere. This makes the design problem very complex. NVIDIA has handled these tough requirements well, but we’re aiming for radical simplicity.”
Integrated Design
Unlike the HW4 computer in the 2026 Model Y, which has a dedicated graphics subsystem, the AI5 chip is designed as a GPU for AI inference. This allows Tesla to eliminate elements like the image signal processor (ISP). Musk noted, “Due to these removals, we can actually fit the AI5 chip in a half reticle while maintaining ample margin for memory traces to the Tesla Trip accelerators, the ARM CPU cores, and the PCI-X blocks.”
In essence, what Tesla is doing with the AI5 chip reflects a broader trend in AI application. Instead of investing in costly general language models, they are focused on creating specific solutions that don’t require the vast billions of parameters found in all-encompassing models like ChatGPT.
Focused Development
Since Tesla’s sole customer is itself, and it knows the AI5 chip will be used just for its self-driving cars and the Optimus robots, the company can refine the entire hardware stack for critical software points. Even with AI4, Tesla assured that cars equipped with FSD 14.2 will experience a seamless self-driving experience, as this update will introduce reasoning capabilities.
With reasoning, the system will evaluate which parking spot to choose. It will recognize that while there may not be a spot right at the entrance of a full parking lot, it can drop passengers off at the entrance and then search for parking elsewhere. It aims to identify open spots more effectively than humans, thanks to its 360-degree vision. As Musk stated, it will employ reasoning to navigate these decisions.
Future Compatibility
It remains uncertain whether the FSD 14.2 version will be compatible with older vehicles. Tesla announced that it would offer an FSD 14 Lite for cars equipped with the HW3 computer in Q2, but they haven’t disclosed which features will be omitted. By that time, they may have already initiated production scaling for the AI5 chip ahead of the Cybercab launch planned for the April-June period. Musk hinted that any surplus AI5 chips produced will be utilized in Tesla’s FSD data centers, complementing the existing Nvidia hardware.
