Tesla's AI5 now carries 96 GB of LP5, a nudge above the 72 GB minimum Musk had set during the memory-shortage cut. His stated reason, in a single X post: at 72 GB, Tesla "would otherwise be the only company using the min RAM version of LP5." The AI5 taped out in April. The AI6 sits at 144 GB of LP6, cut by a third from its original spec.

The 24 GB delta is not a performance figure. It is a spec-sheet position. In a shortage, moving from the minimum LP5 bin to the next tier changes the allocation conversation with the memory supplier and the BOM cost per Optimus unit. What the document does not prove is that the AI5 architecture actually needs the extra 24 GB; the stated justification is competitive positioning, not a measured workload requirement. For the fab, the constraint is unchanged: the die is already taped out, so the memory bump is a packaging and procurement decision, not a silicon one. The buyer of an Optimus unit gets more headroom, but the fab gets a harder allocation in a tight market.

No document in the dossier states the LP5 data rate, the die size, or the process node. The three-times performance-per-watt claim carries no benchmark. The 96 GB figure is a procurement number, not a measured one. For the buyer, the change is price per unit; for the fab, it is memory allocation.

The 24 GB bump is a procurement and positioning decision, not a measured architectural requirement. The die is already taped out; the memory change is a packaging and allocation problem, not a silicon one.

The LP5 data rate, die size, process node, and any benchmark at 96 GB versus 72 GB are absent from every document in the dossier.

The 96 GB figure is a BOM-level decision in a shortage market. It does not reflect a measured workload need. The fab constraint is memory allocation, not die redesign.