Tesla cut its next robot chip's memory in half, then walked a third of that back within hours

Tesla's next-generation self-driving and robotics chip was supposed to carry 144 gigabytes of memory. On Oct. 1, Elon Musk said Tesla had cut that in half, to 72GB of LPDDR5, to keep Optimus robot production moving through what he called a historic memory shortage -- and cut the follow-on AI6 chip's memory by a third, from a planned 216GB to 144GB, in the same breath. The AI5 number didn't last a day: by Oct. 2, Musk had pushed it back up to 96GB, saying Tesla would otherwise be shipping the lowest-spec memory configuration in the industry. AI6's cut held.

AI5 is the chip meant to replace Tesla's current AI4/HW4 computer across cars, Optimus, and -- Musk has said separately -- some of xAI's own inference servers. Musk has called early AI5 silicon a "monster," claiming up to 40 times AI4's practical throughput on the workloads that matter most and roughly five times the memory bandwidth, with the chip's processor die surrounded by a dozen memory packages on the version Musk demoed earlier this year. The planned 144GB was already nine times AI4's 16GB -- the number Musk said Tesla was cutting, not the chip's compute itself.

The chip matters because Optimus's own production ramp is still barely starting. Tesla's stated target for 2026 is 50,000 to 100,000 units, with the Fremont line designed for an eventual 1-million-a-year run rate and a second line at Giga Texas aimed at 10 million units annually once it comes online in 2027. On Tesla's Q1 2026 earnings call, Musk said the actual pace would be "literally impossible to predict," warning output would be "extremely slow at first" across roughly 10,000 unique parts on an entirely new line -- "this is not like making a car." A chip that's cheaper to build in volume matters most exactly when the production line underneath it is this far from settled.

“We cut our RAM in half for the Tesla AI5 chip (now 72GB of LP5) and 1/3 for AI6 (now 144GB of LP6). This was the only way to get enough volume for Optimus production and greatly reduces cost.” — Elon Musk, Oct. 1, 2026, on X

Musk's framing treats the cut as a pure production-math tradeoff, not a design compromise: memory *bandwidth* -- how fast the chip moves data -- is, in his account, the bigger limit on Optimus's performance than total capacity, and bandwidth stayed where it was even as capacity came down. That is Tesla's own characterization, posted by its CEO on his own platform the same day as the cut; no independent benchmark of a reduced-memory AI5 exists yet to confirm it, and Tesla has not published a technical rationale beyond Musk's social posts.

The reversal came almost as fast as the cut, and for a different reason than performance. Replying directly to a Tesla-focused account on X, Musk wrote that Tesla had "decided to nudge AI5 up a little to 96GB, as Tesla would otherwise be the only company using the min RAM version of LP5" -- a competitive-optics concern about where Tesla's spec would sit next to rivals, not a retraction of the bandwidth argument. AI6, a third off its planned 216GB, was never revised.

The chip didn't move in a vacuum. One day before Musk's first cut, on Micron's fiscal fourth-quarter earnings call, chief executive Sanjay Mehrotra told investors that a single humanoid robot would need more than 200GB of DRAM plus multiple terabytes of storage -- roughly ten times the memory content of today's average driver-assist car -- and that Micron saw no line of sight to supply catching up with that demand before 2027. Micron is positioning physical AI, not data centers, as the next multi-decade driver of memory demand; Tesla, trying to ship Optimus at volume through the exact shortage Micron described, is the first company visibly trading chip spec for chip supply in public.

Micron isn't alone in describing a market this tight. J.P. Morgan's own research estimates DRAM prices will have risen more than 400% from the start of 2024 to the end of 2026, with spot prices up closer to 700% over the past year alone as data centers are forecast to absorb roughly 70% of all memory supply. Samsung, SK Hynix, and Micron have all shifted production capacity toward higher-margin server and HBM chips -- HBM is reportedly sold out through 2026 -- leaving less conventional DRAM for everything else; analysts covering the squeeze expect it to shrink the PC market by up to 9% and smartphones by 5% this year as vendors pass the cost through. Tesla designing its own silicon, rather than buying a merchant chip off a vendor's roadmap, is what let Musk cut and partially restore a memory spec within a single day instead of waiting on someone else's allocation.

What Musk's numbers don't settle is whether 96GB is where AI5 actually ships or just where the spec is sitting until the next supply update. The number changed twice in roughly 24 hours once Tesla started discussing memory publicly at all -- which is itself more convincing evidence that the DRAM market Micron described is tight than any single statement from either company. A chip spec that moves twice in a day over a memory shortage is a supply chain talking, not a finished design.