Every Optimus reveal arrives wrapped in trillion-dollar language. The filing is quieter and more useful. Tesla's FY2025 10-K, retrieved via EdgarBeast and filed at sec.gov, states plainly that the company is applying the artificial-intelligence learnings from its self-driving technology to Bots, such as Optimus, and that it uses custom-designed inference chips.
Read that as an engineer, not a fan. The claim is not that Tesla built a bespoke humanoid-robotics organization from scratch. The claim is that the same neural-network perception, the same training pipeline, and the same in-house inference silicon that interpret a road scene are being pointed at a manipulation-and-locomotion problem instead. Optimus, in the company's own words, is downstream of FSD.
That framing is the story. It explains why Tesla can plausibly stand up a humanoid effort faster than a pure-play robotics startup: it is amortizing an AI stack it already pays for. It also sets the ceiling. If the self-driving perception stack has a weakness — a long-tail edge case, an inference-budget limit — the robot inherits it, because the robot is running a variant of the same brain.
The custom-inference-chip detail matters for the same reason. A humanoid that must see, plan, and act in real time on battery power is an inference-budget problem before it is a dexterity problem. Tesla telling the SEC it designs its own inference chips is telling you it intends to control that budget rather than rent it — the one durable advantage a car company brings into a field crowded with software-first humanoid entrants.
So the honest read of the sec.gov document is this: Optimus is not a separate bet. It is the autonomy bet, redeployed onto two legs. Audit the humanoid demos against that, and the right question is not how many degrees of freedom the hands have — it is whether the FSD-derived perception stack generalizes from driving to grasping. The filing makes the dependency explicit; the keynote tends to hide it. Source surfaced via EdgarBeast.