Tesla has cut the memory on its next-generation AI chips to secure enough supply for Optimus humanoid robot production, according to Elon Musk. In a post on X on October 1, Musk said the company halved the RAM on its AI5 chip to 72GB of LPDDR5 and cut the AI6 by a third to 144GB of LPDDR6, describing the move as the only way to get sufficient volume for Optimus while sharply reducing cost.
The disclosure came just a day after memory maker Micron told investors that humanoid robots may each require more than 200 gigabytes of memory, highlighting how the global DRAM crunch is becoming a defining constraint for the robotics industry.
What Tesla Changed
Based on Musk's post and earlier statements, the reductions are substantial:
- AI5: memory cut from 144GB to 72GB of LP5.
- AI6: memory cut from 216GB to 144GB of LP6.
- Bandwidth: held constant, according to Musk.
Musk argued the change should have a negligible effect on Optimus performance, because memory bandwidth, not total capacity, is the bigger bottleneck for the robot's AI workloads. That claim is Musk's own and has not yet been demonstrated in deployed robots.
Musk previously described AI5 as delivering roughly eight times the raw compute of Tesla's current HW4 computer, with significantly more memory. Reports indicate trial production of AI5 is underway at Samsung's Taylor, Texas facility, with volume production expected in 2027.
Micron's 200GB Warning
On Micron's fiscal fourth-quarter earnings call on September 30, CEO Sanjay Mehrotra said humanoid robots are expected to need more than 200GB of memory and multiple terabytes of storage per unit, a profile similar to autonomous vehicles.
Musk's post was a reply to an X user who used that figure to illustrate the scale problem. At 200GB per robot, a fleet of billions of humanoids would require orders of magnitude more DRAM than the world currently produces in a year. Whatever one thinks of such long-range projections, the arithmetic underscores a real tension: AI data centers, autonomous vehicles and now humanoid robots are all competing for the same constrained pool of high-performance memory.
Engineering Around the Memory Crunch
Tesla's decision illustrates a strategy that other robotics makers may be forced to follow. Rather than waiting for memory supply to catch up, engineers can:
- Prioritise bandwidth over capacity, keeping data moving quickly while storing less on-chip.
- Shrink or compress models so that on-robot inference fits into smaller memory footprints.
- Offload heavier reasoning to the cloud or to fleet-level systems where latency allows.
Each approach carries trade-offs. Smaller on-device memory can limit how large a model a robot can run locally, how much sensory context it can retain, and how gracefully it can operate when connectivity is poor. Tesla is betting that careful system design can absorb those limits without visible performance loss.
Why It Matters
The humanoid robot race is increasingly a supply chain race. Industry trackers report that humanoid shipments grew sharply in the first half of 2026, with Chinese vendors supplying the overwhelming majority of units globally. For Western manufacturers trying to scale, access to chips and memory may matter as much as advances in AI models or actuators.
Tesla's move is also a signal to the memory market. If a company with Tesla's purchasing power is cutting memory specifications to secure volume, smaller robotics startups could face even tougher allocation and pricing. That may push the industry toward more efficient on-device AI and closer partnerships with memory makers such as Micron, Samsung and SK Hynix.
For investors and buyers, the key question is whether reduced-memory hardware can still deliver the dexterity and autonomy that Optimus has been promised to achieve. As several commentators noted, the real test will come when these chips are running inside working robots on factory floors.
The Bigger Picture
Tesla has positioned Optimus as a central pillar of its long-term strategy, and it has invested heavily in vertical integration, including the planned Terafab chip facility announced with SpaceX in August. Cutting memory now suggests the company is prioritising near-term production volume over maximum on-board specifications, a pragmatic choice in a market where every major technology company is chasing the same components.
Whether that pragmatism pays off will depend on how quickly Tesla can move Optimus from demonstrations to meaningful deployments, and whether its bet that bandwidth beats capacity holds up in the messy, unpredictable environments where humanoid robots must ultimately work.
