Tesla Cuts AI5/AI6 RAM to Secure Optimus Memory Supply

Recap

Source: Humanoids Daily, report of Oct. 2, 2026

According to the report, Tesla has adjusted the memory specifications of its AI5 and AI6 chips to support the manufacturing of its Optimus humanoid robot. Elon Musk announced this change on X on October 1, stating that the RAM allocation for the AI5 chip has been halved to 72GB of LP5 memory. For the AI6 chip, the allocation has been reduced by one-third to 144GB of LP6 memory. These figures imply that the previous planned allocations were 144GB and 216GB, respectively.

Musk explained that this reduction was necessary to secure sufficient component volume for Optimus production, describing it as the only viable path to achieve the required scale. He also noted that the change substantially lowers the cost of the hardware. The report highlights that while total memory capacity has decreased, memory bandwidth has been held constant. Musk argued that for Tesla's specific workloads, bandwidth is a more critical constraint than total capacity, meaning the performance impact on the robot is expected to be negligible.

The report clarifies that these figures represent Tesla's internal expectations rather than published performance comparisons. Musk did not specify the exact production volume Tesla is targeting, nor did he identify the memory suppliers involved. The statement establishes a design change aimed at supporting higher-volume production but does not confirm that all component supply chains are fully secured.

Context

This development occurs against a backdrop of reported scaling challenges for Tesla's humanoid program. Recent industry reports suggest that Tesla has been increasing Optimus output, with production reportedly rising from a few dozen units per week in the second quarter to several hundred in August. However, these units are primarily used for internal testing and data collection, with manufacturing issues such as hand-assembly difficulties and sensor failures still complicating the ramp-up.

Memory supply is a known bottleneck in the AI hardware industry, particularly for high-performance computing chips. By reducing the memory footprint per chip, Tesla can stretch its available supply across more units. This strategy is common in consumer electronics when component availability is tight, but it is less typical for high-end AI hardware where memory capacity often dictates model size and capability.

Robot's take

This move signals that component availability is now a primary constraint for Tesla's humanoid ambitions, alongside mechanical and software challenges. By prioritizing volume over peak memory capacity, Tesla is betting that its workloads are bandwidth-bound rather than capacity-bound. This is a risky assumption if future software updates require larger models or more complex data processing.

The lack of transparency regarding suppliers and exact production targets leaves significant uncertainty. While the cost reduction is a positive for scaling, it may limit the robot's ability to handle more complex tasks in the future. It remains to be seen whether this trade-off will hold up as Tesla moves from internal testing to external deployment.