Elon Musk ‘shows off’ the Tesla AI5 AI chip

Elon Musk said the Tesla AI5 chip model is 40 times faster than its predecessor AI4, used for Tesla cars, Optimus robots and xAI data centers.

“Congratulations to the Tesla AI team on completing the AI5 chip design,” Musk wrote on X on April 15. “AI 6, Dojo 3 and other exciting chip models are also in development.”

He also thanked TSMC and Samsung for taking care of the production of new chips. “This will be one of the most produced AI chips ever,” the American billionaire emphasized.

According to the image shared by Musk, AI5 has a fairly small ASIC integrated circuit surrounded by 12 memory chip packages from SK Hynix (most likely using GDDR6 or GDDR67 technology). According to Tom’s Hardwarethis parameter shows that the chip has a fairly wide memory input/output interface (I/O Interfacing).

In microprocessors and microcontrollers, I/O Interfacing is considered important, the purpose of ensuring effective data and information transmission between internal and external components of a computer system. Therefore, if AI5 indeed has 12 GDDR6/7 memory packages, the chip will have a 384-bit memory interface. Depending on the type of memory used, it can provide memory bandwidth from 768 GB/s to 1,536 GB/s.

Musk said AI5 had just “completed the design”, meaning the final chip design had just been sent to the manufacturing plant, but the image showed that the processor was built with the symbol “KR 2613”, meaning it was packaged in the 13th week of 2026.

 

Tesla AI5 chip. Image: X/Elon Musk

Tesla has not announced the performance of AI5. However, in the third quarter 2025 financial report, Musk emphasized that the chip is significantly improved compared to AI4. “I think the Tesla engineering team is really designing an amazing AI chip model,” he said at the time. “AI5 will be up to 40 times better than AI4. By eliminating outdated hardware, Tesla can actually integrate AI5 into half the area, while speeding up memory connectivity, Arm CPU cores and PCIe blocks.”

Musk’s announcement on April 15 also shows that Tesla has not given up its ambition to build the Dojo 3 supercomputer project. In August 2025, some sources said that the team in charge of the project “has been abandoned”. Peter Bannon, head of the Dojo project at Tesla, also retired at that time.

But earlier this year, the billionaire announced that Tesla was promoting this project and using entirely self-developed chips to reduce dependence on Nvidia. “The AI5 design is in very good shape, allowing us to put resources back into Dojo 3,” Musk wrote on X on January 21.

Dojo 3 will be the first supercomputer built by Tesla using only internal hardware. Previously, its predecessors Dojo and Dojo 2 still used a mixed design between self-developed chips and Nvidia graphics processors. The highlight of the project lies in the roadmap to launch hardware at a “breakneck” speed. According to Musk, Dojo 3 uses AI5, AI6 and AI7 chips with a release cycle every 9 months, although experts worry that building a completely new architecture in such a short time is very difficult, if not impossible.

Dojo 3 supercomputer is not only used for data processing on FSD (Full Self-Driving) self-driving cars. Musk revealed that supercomputers will be applied to AI in space. However, experts are still cautious about Dojo 3’s ability to succeed.

Dojo 1 was considered the most powerful supercomputer when it launched but was later surpassed by Nvidia, and Dojo 2 was canceled midway. If favorable, Dojo 3 will help Tesla have complete autonomy in AI infrastructure, reduce operating costs by billions of dollars each year and consolidate its leading position in the smart car era. Musk even announced that Tesla will produce AI chips with a larger output than all other chip companies combined.


By Editor