Tesla reduced memory on its AI5 and AI6 chips to secure sufficient supply for Optimus production, marking a first where memory constraints dictate chip design.
Hardware engineers know this moment well: when the product definition meeting stops debating performance and starts asking how much inventory is available, chip specifications are no longer designed but reverse-engineered from supply limits.
On October 1, Musk confirmed on X that Tesla has halved the memory on its AI5 chip to 72GB LPDDR5 and cut the AI6 chip by one-third to 144GB LPDDR6. He stated this is the only way to secure enough production volume for Optimus while significantly reducing costs, adding that the impact on performance is negligible since bandwidth, the primary bottleneck, remains unchanged.
The first interpretation views this as a rational engineering trade-off. As humanoid robot production scales from dozens to thousands, and eventually tens of thousands, per week, lowering per-unit specifications to boost total output is a sound calculation, especially since bandwidth, which drives real-time perception and motion control, remains intact.
The second interpretation is more cautious: if capacity were truly unimportant, why was it originally designed at 144GB and 216GB? The AI5’s half-reticle structure, with a central compute die and 12 surrounding DRAM modules, was initially planned to accommodate local large models, a vision now corrected by supply realities.
I lean toward the second interpretation, with one correction: this is not a design failure but a prioritization. Musk has noted that Optimus is Tesla’s hardest product to mass-produce, following an S-curve ramp-up where securing critical parts for every unit takes precedence over optimizing single-unit parameters.
Memory Shifts from Procurement Item to Design Input
Placed on a longer timeline, this minor adjustment carries significance beyond a simple spec change. During the earnings call in July of this year, Tesla thanked Micron for providing a "substantial" memory allocation on "very reasonable" terms despite "extremely crazy" memory prices, with Supply Chain Vice President Karn Budhiraj participating in the negotiations. Less than two months after securing the allocation, the company is still forced to lower chip specifications, indicating that obtaining the quota has not alleviated the pressure of tight supply.
144→72
AI5 Memory GB
216→144
AI6 Memory GB
200GB
Estimated DRAM per Robot
The first two groups represent the memory capacity before and after this adjustment, corresponding to a 50% reduction and a one-third decrease, respectively. The third group comes from Micron's assessment in its latest earnings call, which states that each humanoid robot requires over 200GB of DRAM and several TB of NAND flash. Viewing the third group alongside the first two explains why the supply side views robots as a new engine for memory demand.
Optimus production plans have been steadily rising. In the second quarter, weekly output was in the tens of units; by August, it had climbed to the hundreds. Management targets exceeding 1,000 units per week by the end of the year, with long-term plans for approximately 20,000 units per week. Each step up in production volume corresponds to a step up in memory procurement, and the current tightness in DRAM and NAND supply is widely expected to persist until 2028.
Reducing memory capacity allows the same allocation to support more units. For a production line in a ramp-up phase, this choice yields immediate results: the same procurement volume supports a higher output.
Another structural shift is worth noting. Tesla's existing AI4+ platform for autonomous driving is configured with 64GB of memory, while the reduced AI5 configuration is 72GB; the two are now very close. Musk previously stated that AI4 is already sufficient to achieve safety better than human drivers. The implication is that the additional compute capability in AI5 primarily serves robots, not in-car autonomous driving.
This also explains the rationale behind the memory adjustment. Optimus relies on the coordinated operation of visual perception, motion control, and decision-making models, which must reside locally. If the entire software stack still fits after the capacity is halved, performance remains unaffected. If it does not fit, frequent swapping is required to compensate, and paging itself consumes bandwidth—this is the boundary condition where bandwidth becomes more critical than capacity.
My assessment: Tesla claims the performance impact is negligible, but that may not hold for workloads where capacity and bandwidth are not equivalent, particularly in stages like multi-sensor data fusion that require large buffers to reside simultaneously. If upcoming public end-to-end inference tests for Optimus show significant degradation in these stages, it indicates that cutting capacity did affect performance. If no degradation appears over time, it suggests the software stack’s memory footprint is well below 72GB, and this adjustment merely trimmed redundancy.
From AI data centers to robots and then to smartphones, the transmission path for memory tightness has formed a single line. Smartphone manufacturers have generally raised prices in this cycle, and robot manufacturers are beginning to adjust chip specifications; both are drawing from the same pool of wafer capacity. New memory production lines typically require two to three years from groundbreaking to mass production, making significant supply-side improvements unlikely in the short term.
Tesla’s choice provides a reference point for the industry: when memory allocation becomes a hard constraint, adjusting product specifications is more effective than adjusting procurement plans. This reference is particularly valuable for companies building edge AI hardware—the room to 'max out parameters first and then figure out sourcing' is shrinking rapidly.
If memory tightness persists for another two years, which category of products will be next to see specification downgrades—robots, in-vehicle systems, or edge AI devices? The first item on the spec sheet to be compromised often most honestly reveals a company’s priorities.