Why Tesla Cut AI5 and AI6 Memory—and Then Raised AI5 to 96GB

Tesla’s memory changes for its upcoming AI5 and AI6 chips are a supply-and-cost trade-off, according to Elon Musk: use less RAM per chip to secure enough memory for Optimus production. The plan changed quickly, though: AI5 was first cut from 144GB to 72GB, then raised to 96GB, while AI6 stayed at 144GB. Those are planned specifications, not demonstrated production results. 11719

The planned memory changed twice for AI5

The initial reported plans were 144GB of LPDDR5 memory for AI5 and 216GB of LPDDR6 for AI6. Musk first said Tesla had cut AI5 to 72GB and AI6 to 144GB. He later raised AI5 to 96GB, saying Tesla would otherwise be the only company using the minimum-memory version of LPDDR5. That leaves AI5 below its original plan, but above the briefly announced 72GB configuration. 71743

Musk described the cuts as necessary to get enough volume for Optimus production and said they would greatly reduce cost. His later explanation for AI5’s 96GB figure points to a component configuration choice; the available reporting does not give a separate explanation for keeping AI6 at 144GB. 11943

Musk expects little effect on Optimus performance—but there are no cited benchmarks

Musk said the lower capacities should have a “negligible effect” on Optimus performance because memory bandwidth is a bigger constraint than total capacity, and he said bandwidth would remain unchanged. That is Tesla’s expectation, not an independently verified performance result. The cited coverage provides no benchmark showing how the revised configurations perform in a robot. 1821

The distinction matters: saying bandwidth is the limiting factor is an explanation for why Tesla expects the change to be manageable, not proof that reducing capacity will have no effect across every workload.

How the decision fits into the memory-supply picture

Tesla’s stated rationale is specific: Musk said the changes were needed to obtain enough memory supply for Optimus production. Separately, Micron has warned that memory and storage supply-demand conditions could be tighter in 2027 and 2028 than in 2026. Together, those reports put Tesla’s decision in the context of broader supply pressure, but they do not quantify how much of the constraint comes from any one market or buyer. 12

Micron has also estimated that a future humanoid robot could use more than 200GB of DRAM and several terabytes of storage. That is a forecast for possible robot designs—not a confirmed specification for Optimus. Tesla’s revised AI5 plan is 96GB, and the available evidence does not establish how much memory future robots will actually need or how quickly they will be adopted. 23

AI5’s proposed uses and performance claims remain unproven

Available reference material describes AI5 as intended for AI inference in Optimus and vehicles, and also lists use in computing clusters for AI training. These descriptions do not establish deployment volumes or a firm production schedule. 2224

Musk has also been reported as claiming that AI5 could deliver roughly two to three times the performance per watt of Nvidia products at about one-tenth the cost. That comparison should be treated as a claim, not a verified, like-for-like test result. 20

For now, the clearest takeaway from the memory revisions is practical rather than performance-related: Tesla says reducing the memory configuration helps it secure supply and lower costs as it plans for Optimus production. Whether the revised chips meet Tesla’s performance expectations will require results from actual deployments or independent testing.