Nobody Knows Just How Close Humanoid Robots Are

Everyone is stuck on doomsday AI headlines. Lab leaders warn we are all going to die. Governments talk about bans. China versus the United States eats the feed. The current that matters more is quieter: physical artificial intelligence, humanoid robots, and the supply chains that already know how to build the hard parts.

I walked that argument on tape. Watch the upload if you want the clips. This is the same case in writing.

The clues sitting in plain sight

Jason Calacanis, host of the All-In podcast, keeps teasing Tesla Optimus. On a clip I used in the video, he says he saw a recent Optimus build that is not public yet, asked Elon whether it was CGI, and was told it was real. His line on that tape: Optimus will be the bestselling product of all time. Treat that as Calacanis describing what he says he saw, not a Tesla product sheet.

Separately, Chinese automaker XPeng has been showing humanoids that walk off the line after they are built. And Tesla has already put Cybercab into the world in Austin: a purpose-built autonomous car with no steering wheel and no pedals, driven by the same class of real-world AI that humanoids will need to stand, walk, grasp, and work.

Those three signals are not the same product. Together they say physical AI is no longer a lab poster. It is shipping in pieces.

World models already work

The reason this moves faster than most people price is world models: AI systems that understand how the physical world behaves. The cleanest live proof is Cybercab in Austin. As of that recording, Tesla had harvested on the order of 14 billion miles of real-world camera driving from customer cars, trained a large driving brain on that data, and reached a point where a rider can fly into Austin, open the robotaxi app, and take a Cybercab. Rider-reported fares have come in a bit under a typical Uber on some trips and above it on others, and the waits get long when there are not enough cars. That price path is my framing from the tape.

That stack is the same class of perception and control a humanoid needs. Instead of left, right, stop for a pedestrian, you ask it to pick something up, set it down, fold laundry, wash dishes, mow a lawn, or run an errand. The body changes. The hard problem, understanding the world through cameras and acting safely in it, is the same problem Tesla already forced through cars.

Why car factories matter

Humanoids need batteries, inference chips, actuators, motors, cameras, plastics, and a supply chain that can feed a factory every day. Electric vehicle makers already buy or build most of that stack. You can reuse a huge share of it for a bipedal robot.

The loudest signal on my tape is Tesla ripping out Model S and Model X capacity at Fremont and turning that floor toward Optimus production. XPeng is on a similar path from the car side in China. Once an automaker decides robots are the next line, the missing piece was never "can we invent metal." It was "can the robot see and act." That is the part now catching up.

Cost math that forces adoption

Most financial analysts I track land humanoids under about $30,000 a unit within a few years once the actuator and motor supply chain ramps. That is less than the average new car in the United States. Divide build cost plus electricity, maintenance, repairs, and insurance across the hours the machine works, and I put worst-case labor near about $4 an hour, with a more typical path closer to about $2 an hour and falling as the hardware and software improve.

Those dollar figures are my framing from the video, not a Tesla price list or a government labor series. The point is the comparison. If a human costs $15, $40, or $100 an hour for the same physical work, and the robot can do the job, the robot wins every time. Capability is the remaining variable. Capability improves with software updates and fleet learning. Elon has described an Optimus Academy with on the order of 10,000 robots, and maybe a few times that, practicing tasks in the real world. The warehouse picture, and the idea that one robot learns a task and the skill copies across the fleet, is my framing from the tape. One robot learns to fold laundry. Millions can fold laundry.

Disruption, then a new layer

In the short to medium term, physical robots and digital agents displace work the same way: cheaper, always on, improving every model drop. New college grads already feel agent pressure. Humanoids extend that into the physical world.

On the other side of every industrial jump, new jobs appear that did not exist before. This channel is one of them. The old economy does not get a soft landing. The question is what you are building on top of agents and robots once the forcing function is obvious.

I wrote two books on that stack, Master Plan and Abundance or Collapse. Links sit in the video description. The upload is the full walkthrough with the Calacanis clip and the factory visuals. The takeaway I want you to keep: almost nobody is pricing how close this already is.