Tesla tapes out AI5 chip

Tesla says its AI5 chip has reached tape-out, the point where a design is finished and sent to a factory for first silicon. (x.com) (ibtimes.com.au) Chief executive Elon Musk posted on April 15 that Tesla’s AI chip team had taped out AI5 and that “AI6, Dojo3 & other exciting chips” are also in work. Yahoo Finance, citing Tesla’s comments, said AI5 is expected to enter high-volume output in 2027 and could replace the current AI4 chip in Tesla vehicles. (x.com) (finance.yahoo.com) In chipmaking, tape-out is the handoff from design to manufacturing, not the start of mass production. The next steps are test chips, debugging, and yield work to see how many usable chips a factory can make from each wafer. (techpowerup.com) (finance.yahoo.com) Tesla’s own artificial intelligence page splits its chip work into two jobs: inference chips that run models in cars and robots, and Dojo chips that train those models in data centers. Tesla says it is pursuing autonomy for vehicles and its Optimus humanoid robot, both of which need fast on-device computing after the training is done. (tesla.com 1) (tesla.com 2) That makes AI5 the “run it in the machine” chip, while Dojo3 points to the “teach the model” side of the stack. Reuters reported in March that AI5 is mainly optimized for edge compute in Robotaxi and Optimus, even as Musk said Tesla and SpaceX AI would keep ordering Nvidia chips at scale for other workloads. (videocardz.com) (finance.yahoo.com) The manufacturing roadmap is still thin. Musk thanked Samsung Electronics and Taiwan Semiconductor Manufacturing, and recent reports said Samsung plans high-volume AI5 production in Texas in the second half of 2027. (x.com) (notateslaapp.com) The AI6 timeline is thinner still. Reports circulating in March said Musk had previously pointed to a December 2026 tape-out target for AI6, with Samsung aiming at second-half 2027 production on a 2-nanometer process, though Samsung-related delays have also been reported. (videocardz.com) (electrek.co) Early photos of the AI5 package suggest Tesla is packing a large compute die with multiple memory chips around it, a layout meant to keep data close to the processor. Several reports identified the memory as SK hynix and described 12 modules on the package, but Tesla has not published a full specification sheet. (videocardz.com) (techpowerup.com) Tesla has been building toward this for years. Its support pages already refer owners to “AI computer” upgrades for Full Self-Driving hardware, and its public recruiting pages describe Dojo as an in-house training chip effort that spans from design through mass production. (tesla.com 1) (tesla.com 2) The immediate change is simple: Tesla has finished one chip design and is now trying to turn it into working silicon at scale. The harder part starts after tape-out, when AI5 has to survive factory bring-up and prove it can ship in volume in 2027. (x.com) (finance.yahoo.com)