Tesla Cut Its AI Chips' Memory in Half — to Feed Optimus
October 2, 2026 · 5 min read
Elon Musk announced on X on October 1 that Tesla's next-gen AI chips are getting a memory diet: AI5 drops from 144GB to 72GB of LPDDR5, and AI6 goes from 216GB down to 144GB of LPDDR6. The reason is blunt — "only this way can we pool enough memory quota for Optimus mass production" — with a big cost cut as a bonus. His defense: the real performance bottleneck is bandwidth, not capacity, and bandwidth stays the same, so the impact on Optimus performance is "negligible."
Cutting memory in half and calling it negligible takes confidence. Let's check the math.
Bandwidth vs capacity: he's actually right (mostly)
For AI inference — especially the real-time kind a humanoid robot needs — what matters most is how fast you can feed the model weights into the compute units, not how many weights you can park in memory. If the model fits and the bandwidth is unchanged, halving capacity genuinely doesn't slow much down. It's the same reason you can game fine on 16GB of VRAM even though 32GB exists: throughput beats the closet.
The catch is the word "if." Models keep growing. A chip specced tight for today's model sizes gets uncomfortable fast when next year's model is 30% bigger. Tesla is betting its model roadmap stays inside the new memory envelope. That's a bet on its own software discipline as much as on silicon.
The real story: memory is the currency of the robot ramp
Read the announcement again and the technical justification is the side dish. The main course is supply: every gigabyte of memory on an AI chip is memory that has to be manufactured, bought, and allocated. Cutting per-chip memory roughly in half means roughly twice as many robots per memory shipment — or the same number of robots at a much lower bill of materials. When you're trying to mass-produce humanoids, memory stops being a spec and becomes logistics.
It's also a quiet admission about where the industry is: memory — not compute — is now the binding constraint on scaling AI hardware. The entire sector is rationing it. Tesla just said the quiet part out loud.
The timeline to watch
AI5 is being built by Samsung and TSMC on 2nm, with mass production targeted for 2027. That's the chip that has to carry both the car fleet's next-gen autonomy and the first real Optimus production wave. If the memory math is wrong, 2027 is when the bill comes due — in the form of models that don't fit and robots that can't think fast enough. If it's right, Tesla just shaved a fortune off every robot it will ever build.
Either way, it's the clearest signal yet of Tesla's priority stack: the car is the present, the robot is the future, and the silicon serves the robot first.