Tesla's AI5 autonomous driving inference chip has completed tapeout, according to Elon Musk, who posted a photo of the chip package on X on April 15. The package carries Tesla branding. Musk also said AI6 and Dojo3 are in development, putting three in-house AI hardware efforts back into view at the same time.
AI5 uses a dual-foundry plan with TSMC and Samsung
Tesla's AI5 is set to be manufactured through two process paths: TSMC at 3nm and Samsung at 2nm. The design is the same, but the production base is split between two foundries. That gives Tesla a way to reduce supply-chain concentration and secure more capacity for future rollout. It is also a practical response to tight leading-edge foundry supply.
On performance, the report says AI5 can deliver up to 40x more compute than the current in-vehicle HW4 in specific scenarios, with 9x more memory and 5x higher memory bandwidth. Total inference performance is estimated at 2,000 to 2,500 TOPS at int8 precision, roughly 5x HW4. The larger memory footprint matters. It allows the car to run bigger vision models and process longer spans of video context while interpreting changing road conditions.
Sample chips are planned for late 2026, mass production for 2027
Tesla's timeline points to limited samples by the end of 2026 and mass production in 2027. The first deployment targets are listed as the Cybercab robotaxi, the Optimus humanoid robot, and Tesla data centers.
There is still a gap between announcement and deployment. Earlier reporting on Cybercab had suggested that the first vehicles may use AI4 hardware as a bridge, with AI5 only arriving once mass production starts in 2027. Musk also framed AI5 against Nvidia's lineup: a single AI5 SoC is described as roughly comparable to Hopper, while a dual-chip configuration is said to be around Blackwell class, at a cost more than an order of magnitude below Nvidia's $30,000 data-center GPUs.
Dojo3 returns with a new “space-based AI compute” label
Dojo3 is notable because the project had looked uncertain. Bloomberg reported in August 2025 that Tesla had dissolved its in-house Dojo training-chip team and shifted toward large purchases of Nvidia, AMD, and Samsung GPUs for AI training. That report fed speculation that Tesla was stepping back from its own training hardware plans.
Then in January 2026, Musk brought Dojo3 back into the discussion and described it as “space-based AI compute”. The label suggests a different role from earlier Dojo expectations. In that setup, AI5 and AI6 would focus on vehicle-side and ground-based inference, while Dojo3 would sit on the training side with a more remote and larger-scale compute mission.