The more profound truth behind the spectacle is, as it has been pointed out, that there are still too many demos that use scripted behaviors or require direct oversight. The line proposed by Lei Yang, in a message during an email exchange on the subject of humanoids, is too close to what Tesla intends to do today, which is invest heavily, promise heavily and hope that people will mistake a flashy demonstration with a workable employee.
Tesla has projected an annual capital-spending which is less like a factory upgrade of an automaker than an off-the-computer-then-off-the-manufacturing-land grab. The firm has steered towards over 20 billion of capex in 2026, and automotive sales in Q4 2025 were down by 11 percent. The most noticeable redeployment is Fremont: the plant will cease producing Model S and X and instead re-equiped to produce Optimus, to which Elon Musk promises investors that he is certain of achieving one million units a year of Optimus 3 there, and he says that he will make the plant produce Optimus at a cost of $20,000 per robot.
The question of engineering is not whether it is possible to construct a line and create many humanoid-shaped machines. Whether or not such machines are capable of performing work of economic significance, without a hidden supportive cast, is the question. Musk has confirmed that Optimus is not operational in a material sense in Tesla factories in the present day, but states that it is still in R&D and predicts large volumes of production only towards the end of the year. That discontinuity between a prototype that handles those simple tasks and a fleet that handles paid labor is precisely what McKinsey had drawn attention in the gap between what is technically demonstrated in pilots and what is commercially viable at scale.
The prolonged gap has led robotics researchers and operators to be more open in explaining why it does not disappear. The issue of human in the loop demonstrations is one of the recurring themes: spectacular manipulation, multi-step activity, or cloth manipulations are regularly shown to be teleoperation or pre-programmed behavior. This breakdown of communication is discussed in detail, including examples, like Stanford Mobile ALOHA and an Optimus folding video uploaded by Musk, in a discussion of why robot videos are deceptive out of context. Teleoperation can be applied to training and data gathering, yet this is not a scalable model of labor and may mislead the expectations of the whole field.
Next comes safety, the component of humanoid robotics which is not impressed by quarterly earnings calls. Humanoids are frequently dynamically stable, i.e. power is needed to keep them up; conventional emergency-stop logic, kill power, stop motion, may become kill power, fall over. Aaron Prather, a standards worker in ASTM and IEEE, simply stated: You cannot do that with a humanoid. Standards standards involving industrial mobile robots are being developed with actively controlled stability, which attempts to define the safety outcomes but does not pre-determine the mechanical implementation, which is an uncomfortable imposition on anybody who is offering general-purpose machines in environments where untrained individuals coexist.
The humanoid form factor is even less nostalgic to industry veterans. Roboworx VP Dale Walsh claims that the idea of Humanoid is a form factor, not a capability, and since simple physics: a five-foot, 150200 lb machine in the state of nature it is unstable, and should the power go dead, it will fall. Computer scientist Rodney Brooks of MIT robotics has been more pessimistic, terming proximity-compatible humanoids well short of the next decade as mere more extreme thinking and opining that dexterity and close-proximity safety will not be high-performance in the next decades.
In the meantime, actual robots are already good-earning–by being unromantic. Hotels and other large facilities install machines who serve towels or vacuum corridors, or otherwise provide a limited concierge interface, frequently on the controlled routes and typically limited in interactions. A robot usually works well, even when it is modeled after a person, not by mimicking a workplace partner but by performing small, repetitive activities. The difference in that contrast is important to Tesla: it is not the same to scale production as usefulness.
As Tesla reverses its expenditure, the corporation is in fact betting on the fact that muscle and AI compute will enable it to squeeze a list of recalcitrant issues that encompass autonomy, dexterity, safety engineering and social acceptance into a product cycle. The funds are able to construct clusters and factories. The final challenge is how Optimus can stand on his feet, work, and create value without having an operator somewhere that is just out of view.
