Tesla Cuts Memory on AI5 and AI6 Chips to Lower Costs
Elon Musk details reduced RAM on next chips while holding bandwidth steady
TL;DR:
- AI5 drops to 72GB LP5 memory and AI6 to 144GB LP6 after RAM cuts
- Bandwidth stays constant so Optimus performance impact stays small
- Tesla chooses volume-specific inference hardware over general flexibility
Headline
Tesla is cutting memory on its AI5 and AI6 chips to reduce costs and speed up Optimus robot production.
Summary
Elon Musk stated that the AI5 chip will have 72GB of LP5 memory and AI6 will have 144GB of LP6, after cuts to RAM. Tesla sees little effect on Optimus performance since memory bandwidth stays the same and is the main limit rather than total capacity. The choice favors high-volume edge inference tailored to specific tasks over flexible hardware.
Analysis
Capacity sets how much model state and context fit on the chip, while bandwidth controls processing speed for workloads that fit. The change improves Tesla's costs per unit and helps plan supply for large Optimus output. The downside is less room for bigger models, longer contexts, multiple models, or future autonomy tasks, though current Optimus work is not affected much. Tesla is choosing hardware tuned for inference over general capacity, which helps if it controls the models on its robots.
Impact Assessment
- Significance: Medium
- Categories: Technical Insight; Industry Trend