In Tesla's latest presentation, the chips so far called AI4, AI5 and AI6 are labeled SI4, SI5 and SI6, for "Super Intelligence." According to Tom's Hardware, Elon Musk recently confirmed SpaceXAI will be renamed SpaceXSI for the same reason. Technically, SI6 is expected to bring 80 times the compute of the SI4 chip in use today.
The most notable change is the return of Dojo, Tesla's in-house training chip. Musk cancelled the Dojo project in August 2025, and in August 2026 said SpaceX and xAI would use only Nvidia hardware. Dojo 3 (D3) differs sharply from its predecessors: instead of an SRAM-centric, wafer-scale architecture like D1, which linked 25 processors with 11GB of SRAM, it will be a chiplet design using 2.5D and 3D packaging with stacked HBM.
Tesla describes D3 as optimized for "space, training, and scaling," meaning orbital data centers as well. The company has not disclosed D3's performance, memory capacity, manufacturing process or schedule, and it is unclear whether it has dropped wafer-scale integration entirely.