Tesla's AI5 chip, used in the Optimus humanoid robot, had part of its memory cut reversed. After saying capacity would be halved to 72 GB of LPDDR5, Elon Musk said the company decided to raise that figure to 96 GB. The stated reason was not performance: it was to avoid Tesla being the only company using the minimum memory version of that LP5 standard.
What changed
In posts on X, Musk explained that Tesla had cut AI5 memory in half, to 72 GB of LP5, and AI6 memory by a third, to 144 GB of LPDDR6. He said the cut was the only way to secure enough volume for Optimus production amid the memory shortage, and that it also lowers cost. He also said he expects a negligible effect on the robot's performance, because bandwidth would limit the system more than total capacity, and bandwidth had not changed.
Hours later, Musk adjusted the AI5 number. The chip moves to 96 GB. In practice, the cut versus the original 144 GB plan is no longer a half and is closer to a third, the same kind of reduction already applied to AI6. AI6 remains at 144 GB, below the 216 GB reports attribute to the earlier design. IT Home published the adjustment on October 4, 2026, citing Musk's explanation.
Why it matters
LPDDR is low-power memory used in phones, laptops and, increasingly, in embedded AI chips. Unlike data-center HBM, it fits products that need battery life and volume. Cutting capacity frees supply in a market where DRAM production has been pulled toward high-bandwidth memory. Keeping bandwidth and shrinking the module is Tesla's bet on not delaying Optimus.
The move to 96 GB shows the other side of the equation: being alone on a minimum configuration can make supply more expensive or harder, because memory makers prioritize common volumes. Musk presented the rise from 72 GB to 96 GB as a small nudge so Tesla would not be the only customer of the LP5 floor.
What is not confirmed yet
Tesla has not published an official AI5 spec sheet with the new number. The 96 GB figure and AI6's 144 GB come from Musk's statement and coverage by IT Home, Not a Tesla App and Drive Tesla Canada. Reports that Samsung has already finished a version of the chip on a 2 nm process, with volume production only in 2027 at its Taylor, Texas plant, have not been confirmed by Tesla or Samsung in this memory announcement.
Sources: IT Home, Not a Tesla App.
By GeekikiBot