Tesla is adjusting the memory specifications of its next-generation AI chip, used for both autonomous driving and Optimus, to meet the demands of mass-producing humanoid robots.
According to recent reports, Tesla CEO Musk announced a significant reduction in the memory specifications of the AI5 and AI6 chips to ensure sufficient storage for Optimus mass production. Musk also stated that this adjustment will not affect the performance of the humanoid robot.
Specifically, Tesla has halved the memory capacity of the AI5 chip to 72GB LP5 and reduced the memory capacity of the AI6 chip by one-third to 144GB LP6. Musk stated that this adjustment is the “only way” to achieve sufficient production volumes for Optimus while significantly lowering costs.
Elon Musk also stated that the impact of this adjustment on Optimus's performance should be "negligible," as memory bandwidth, rather than total memory capacity, is the primary limiting factor.
More notably, this comes just over two months after Tesla thanked Micron during its July earnings call for providing an “extremely substantial” memory allocation. Even with this allocation, adjustments to chip specifications were still required, underscoring the tightness of memory supply and signaling that humanoid robots are emerging as a new source of high-end memory demand.
AI5/AI6 storage downgrade; performance impact remains unclear
Tesla has not yet disclosed the adjusted specific memory specifications for the AI5 and AI6 chips. However, given the prolonged shortage of memory supply, Tesla’s decision to reduce these specifications itself indicates a reallocation of limited memory resources, prioritizing the memory needs required for Optimus mass production.
During the July earnings call, Musk stated that the AI5 was progressing well and would initially be used for Optimus, and that AI6 would become "the best edge computing chip in the world." Now, the memory specifications for both chips have been significantly reduced, creating a contrast with their high-end positioning. Adjustments to memory specifications typically involve metrics such as bandwidth, capacity, or latency, which could impact the chip’s data processing efficiency in real-time perception and decision-making scenarios. Tesla has not quantified the actual impact of this adjustment on FSD performance.
Musk stated that the storage adjustment will not affect Optimus's performance, but this claim is currently unsupported by technical specifications or test data.
Optimus relies on the coordination of visual perception, motion control, and decision-making models; storage specifications affect model inference speed and the efficiency of multi-sensor data fusion. If the adjustment indeed has no impact on performance, possible explanations are that Optimus’s current software stack has not yet reached the upper limit of storage bandwidth after the adjustment, or that Tesla has compensated for the hardware changes through algorithmic optimizations. However, Tesla has not specified which scenario applies, and external parties cannot independently verify either possibility.
Storage shortages are spreading from data centers to cars and robots.
The significance of Tesla's recent adjustment lies in providing new evidence that storage shortages are spreading to the automotive and robotics sectors.
In July this year, Musk acknowledged on the earnings call that the expansion of AI systems had caused a shortage and high prices for memory chips, and he thanked Micron for providing Tesla with a “substantial” allocation of memory under “very reasonable” terms, despite “extremely wild” memory prices. Vice President of Supply Chain, Karn Budhiraj, assisted in securing this allocation. Even now that Tesla has obtained the allocation, it still needs to reduce the memory specifications for AI5/AI6, indicating that memory resources remain tight and reflecting growing competition for memory resources across Tesla’s different business units.
Optimus’s weekly production has increased from dozens of units per week in the second quarter to hundreds, with management targeting over a thousand units per week by the end of the year and a final plan of approximately 20,000 units per week. As mass production scales up, demand for storage will rise accordingly, making humanoid robots a new major driver of demand for high-end storage capacity, following AI servers.
Optimus takes precedence over the FSD chip.
Tesla's trade-off between AI chips and the mass production of robots points to two possibilities: either Optimus's production ramp-up is exceeding market expectations, requiring early securing of critical components; or storage supply has become a new bottleneck in Optimus's production scale-up, forcing Tesla to reallocate resources from other product lines.
Elon Musk has previously emphasized that Optimus is the most challenging product Tesla has ever attempted to scale, with production ramp-up following an S-curve. Tesla’s recent adjustment to the storage specifications for AI5 and AI6 suggests that, under current storage supply constraints, the production demands for Optimus may now receive higher resource priority.
If storage constraints persist, Tesla may need to make more similar trade-offs between Optimus and other business areas in the future.