Tesla is cutting the memory capacity of its next-generation AI chips as it looks to secure enough supply for large-scale Optimus production while lowering costs. According to Wccftech, Elon Musk said Tesla’s AI5 will now feature 72GB of LPDDR5, half the previously planned capacity, while AI6 will use 144GB of LPDDR6, a reduction of roughly one-third.
Musk said on X that the change was the only way to secure sufficient volume for Optimus production while significantly reducing costs. He expects the impact on Optimus performance to be negligible, arguing that memory bandwidth is a greater constraint than total memory capacity. Tesla has kept memory bandwidth unchanged despite the capacity reduction.
Tesla plans to use AI5 and AI6 to power its Full Self-Driving (FSD) compute stack and Optimus humanoid robot, Wccftech adds.
Meanwhile, Tesla’s AI5 chip is already progressing toward production at Samsung’s Taylor facility in Texas. According to Seoul Economic Daily, citing industry sources, the plant has begun producing wafers using Samsung’s 2nm process, bringing AI5 into the prototype stage. Some lines are already making AI5 chips for Tesla while Samsung conducts final mass-production verification, which is expected to be completed by the end of 2026 ahead of full-scale supply in 2027.
Tesla Emphasizes SRAM as Memory Bandwidth Takes Priority
Looking further ahead, Tesla is already developing AI6 and AI6.5 with a greater emphasis on memory bandwidth. Musk said in April that AI6 will pair LPDDR6 memory with Samsung’s 2nm process in Texas and is designed to deliver roughly twice the performance of AI5 within the same half-reticle size. AI6.5 is expected to further improve performance using TSMC’s 2nm process in Arizona.
Both chips will also make extensive use of on-chip SRAM. According to Musk, roughly half of their TRIP AI compute accelerators will be dedicated to SRAM. For calculations performed within the SRAM cache, he said effective memory bandwidth would be roughly 10 times that of DRAM.
The memory adjustment also comes as humanoid robots emerge as a potentially significant source of memory demand. As noted by Wccftech, Micron recently estimated that each humanoid robot could require around 200GB of DRAM and several terabytes of NAND, with physical AI potentially becoming a significant memory and storage demand driver by the end of the decade.
