Tesla Cuts Memory Specs for AI5/AI6 to Secure Supply for Optimus Mass Production
nashnova research
Musk announced cuts of 50% to one-third in AI5/AI6 chip memory, redirecting supply to Optimus robot production — his bet: memory bandwidth is the real bottleneck, not total capacity.
How deep are the memory cuts?
AI5 drops from a planned 144 GB LPDDR5 to 72 GB — a 50% reduction.
AI6 drops from roughly 216 GB to 144 GB LPDDR6 — about one-third cut.
This means → both chip generations are being significantly de-specced at once. This is not a trim; it is a cross-generational reset.
Won't cutting memory hurt performance?
Musk's argument: memory bandwidth — how fast data moves — is the real bottleneck, not total capacity.
In plain terms = think of a factory line. Throughput depends on conveyor speed (bandwidth), not warehouse size (capacity).
He claims the performance impact is negligible, but that claim has not been validated in a mass-production environment.
Where does the freed-up memory go?
The target is clear: mass production of the Optimus humanoid robot.
Musk wrote on X that this is the only path to secure Optimus scale-up while driving costs down under constrained memory supply.
This reflects Tesla's internal resource ranking: Optimus production now carries enough weight to force chip specs to yield.
How powerful is AI5 on its own?
AI5 will be co-manufactured by TSMC and Samsung, with Samsung using a 2 nm process node.
Expected compute: 2,000–2,500 TOPS (trillion operations per second). Musk positions a single AI5 as roughly Hopper-class and a dual-chip setup as Blackwell-class.
AI5 has entered pilot production at Samsung's fab in Taylor, Texas, with mass production planned for 2027.
What about the AI6 roadmap?
AI6 and AI6.5 are expected to be manufactured by Samsung and TSMC, respectively.
Both will integrate large amounts of SRAM — on-chip cache that is extremely fast but costly.
This means → Tesla's chip roadmap spans both major foundries, and the on-chip cache build-up is deliberate — faster, pricier SRAM compensates for the external memory reductions.
What is the risk in this decision?
AI5/AI6 will serve two product lines simultaneously: Full Self-Driving (FSD) and Optimus.
The memory cut is fundamentally a strategic trade-off: tilting limited resources toward Optimus under supply constraints.
Whether Optimus can hit its mass-production timeline is the key test of this bet — if Optimus slips, the memory saved becomes a performance sacrifice with nothing to show for it.
市场有风险，内容仅供研究参考，不构成投资建议。