The AI5 chip taped out in April and will carry 96GB of LP5 DDR5, a step up from the 72GB minimum that Tesla set during the global memory shortage. The chip powers the Optimus humanoid robot. 72GB was the floor for the LP5 tier; 96GB is the next allocation step.

The shift from 72 to 96 is a procurement decision, not an architectural one. The filing does not state that the Optimus workload requires 96GB, only that Tesla would otherwise be the sole buyer of the minimum LP5 tier. For the fab, the 72GB configuration is no longer a single-customer job, and the tooling amortizes across more orders. For the buyer, the extra 24GB sits in the same LP5 package, but no page measured whether the memory controller's bandwidth or power draw changes with the larger allocation.

No page measured the chip's bandwidth, power draw, or performance against the Hopper or Blackwell comparisons Musk cited. The process node is not stated. The exact day within April is not given. The memory shortage volume is not quantified, and the 144GB figure for AI6's LP6 is noted but not tied to a measured workload.

The 72-to-96 shift is a procurement decision, not an architectural one. The filing establishes a supply constraint, not a design constraint, and the extra 24GB does not change the LP5 package.

The fab amortizes 72GB tooling across multiple customers rather than a single one. The buyer gains 24GB of headroom in the same LP5 package, but bandwidth and power draw remain unmeasured.