The candid basis
At the Q4'25 call (Jan 2026) Musk conceded Optimus was 'not in usage in our factories in a material way', with units mainly collecting data; at Q1'26 he said the Fremont line would start late July-August, slowly, with yearly rate 'impossible to predict' (10,000 unique parts on a new line). Circulating figures like '50,000 cumulative units' carry no audited company basis. We rank Tesla last among the four on public operating evidence only - not on technical reserves.
Training route: mocap distillation plus Dojo
Optimus' brain inherits FSD directly: an end-to-end vision network processes camera streams and understands motion and spatial context; human workers in mocap suits provide demonstrations distilled into neural policies via imitation learning. The cloud side rides Dojo (D1 at 7 nm, 362 TFLOPS BF16, 400 W TDP, planned 100 EFlops); the device side inherits the automotive AI chip - precisely the automotive-grade version of the 'cloud model issues instructions, on-device model executes' two-brain architecture.
| Item | Basis |
|---|---|
| Factory use | Musk: not material yet (Jan 2026) |
| Line | Fremont under build, summer 2026 start, slow ramp |
| Training | Human mocap demos to distillation; Dojo |
| Architecture | FSD-derived on-device brain plus cloud inference |