WTF Is Happening at Tesla: Why Optimus Hands Bet on Software Over Demo Hardware

CES-era demos from Boston Dynamics’ Atlas, Figure’s factory bots, and Unitree-style showpieces look more athletic than Optimus today, yet Tesla’s bet is the opposite of flashy hardware: overbuild the hand’s degrees of freedom and let neural nets catch up, the same pattern that made early FSD look clumsy until end-to-end networks improved across versions 10 through 14. A human hand has about 27 degrees of freedom; Optimus Gen 2 offered 11 per hand and Gen 3 about 22, versus simpler three-finger grippers or Figure-class hands near 16—harder to control now, but already capable once models, data, and reinforcement learning improve without a full mechanical redesign. Tesla’s fleet of millions of cars collecting real-world video, plus factories scaling from roughly two million toward five million vehicles a year, supplies the training flywheel competitors still lack. Overall, this aims to make software-updatable dexterity—not backflip demos—the scarce advantage when humanoids enter factories and homes later this decade.

Check the video here.