Why Do We Need Humanoid Robots Instead of Just AI?
Some version of this question turns up whenever a humanoid robot demo goes viral. A composite of the ones we keep seeing:
“I get why AI matters. It writes, codes, diagnoses, plans. But why does everyone want to bolt it onto a clunky metal person? Wouldn’t it be cheaper to keep the intelligence in the cloud and use ordinary machines for the physical stuff? What does a body add that a smarter model doesn’t?”
It’s a fair challenge, and the short answer has two parts. Most of what it costs to keep a human alive is physical work, and software alone never touches it. And the places where that work happens were built around one particular body: ours. Unscarcity treats AI and robots as two separate engines for exactly this reason, and it is blunt that the physical one is running behind.
The Brain Reaches One Line of the Bill
Chapter 1 of the book makes the case with an invoice. Take a $300,000 apartment, the kind Maria Delgado, the Detroit house cleaner at the center of the book, will one day paint sunsets in. About 40 percent of that price is labor: architects, contractors, electricians, plumbers. Another 35 percent is materials, roughly 10 percent is energy, and the last 15 percent goes to permits, profit and overhead.
Now let a perfect AI loose on that bill. It removes the design and coordination hours, and the book’s running tally after its AI section reads: “design and coordination, once the most expensive hours on the job, now cost about what an email costs. The rest of the bill still stands.”
A language model can draw the wiring diagram. It can’t pull the wire through the wall. The electrician, the plumber and the crew on the roof are the labor line, and only something with hands takes it away. The book goes further: materials are “energy and labor wearing a costume,” because concrete costs money only after someone quarried the limestone, fired the kiln and drove the truck. Remove physical labor and materials start to fall too. Leave it in place and abundance stays a software story that stops at the edge of the screen.
At the scale of the planet, the site’s research on humanoids counts 3.5 billion employed people, most of them in agriculture, construction and domestic work, “where AI chatbots are irrelevant.” Its own summary is hard to improve on: the brain can write your emails and compose your sonnets, but someone still has to take out the trash, build the houses and care for grandma.
Why the Body Is Shaped Like Ours
The second half of the question is sharper. Fine, robots, but why humanoids? Why not wheels, arms on rails, purpose-built machines?
Often the answer really is purpose-built machines, and the book says so. Outdoors, where “there is no human in it to work around,” drones and autonomous boats already do the work; the Ocean Cleanup project had removed over 450,000 kilograms of plastic from the Great Pacific Garbage Patch by 2025.
The humanoid case begins where the world was built for people. Amazon is the cleanest example. It runs hundreds of thousands of wheeled, Kiva-descended robots that carry shelving across a flat floor, and they can’t climb, reach or pick. In March 2026 it bought Fauna Robotics, maker of a 3.5-foot bipedal robot called Sprout, aimed at elevation changes and multi-level shelving. The wheeled robots solved the floor. The rest of the building still assumes a person.
Homes, hospitals, kitchens and job sites make the same assumption: stairs, door handles, counters at waist height, tools sized for a hand. The alternative is rebuilding the space around the machine, and Walmart shows what that costs. At one of its automated grocery warehouses, electricity can run to $800,000 a month, against about $250,000 when people did the work, and the engineers who keep the equipment running cost more than the workers they replaced. That is not proof a humanoid fleet would have done better. It is a receipt for the other route.
The Harder Half
Physical intelligence is also the more difficult problem. Language AI learned from the internet, trillions of words people had already written. There is no equivalent library for the finger pressure that cracks an egg or the wrist turn that folds a fitted sheet. Physical skill has always passed by apprenticeship, so it was never written down.
That is why the numbers look so different from chatbot benchmarks. When Google DeepMind released Gemini Robotics 2 in July 2026, it published success rates instead of a highlight reel: 76.3 percent picking from a shelf, 45.7 percent picking from the floor. In September, Figure tested its Helix 2.5 model in 30 Bay Area homes the robot had never seen. Trained from scratch, it succeeded on 9 percent of trials. Pretrained on footage from Index, Figure’s app that pays people to film their own chores (16 million videos by its August launch), it reached 56 percent. Six times better, and still failing almost every other time.
A robot is AI plus a body, and the body is where most of the remaining difficulty lives.
From the Podcast: The Mind as a Download, the Body Through Customs
The November 1, 2026 episode of Minds, Bodies & Terawatts, our daily podcast, put the whole question into one story. Its headline: Meta Has the Money, the Chips and the Agent. To Give Muse a Body It Flew Its Robot Chief to Shanghai for Two Months, and Says the Parts Aren’t Here.
Per The Information’s report, Meta built a 106-pound humanoid called Max with the Shanghai robot maker Fourier in about three months. Internally it was described as Muse, Meta’s personal assistant app, getting a real body, with household chores as the hope. Meta has told investors it will spend up to $145 billion this year, and Nvidia had just put $3 billion into its share sale for the Muse work. None of that bought motors. A Meta spokesperson conceded that sourcing parts hard to obtain domestically is a reality for the whole industry, and under the import rule the FCC adopted in July, a humanoid sold in the United States needs 65 percent US-made parts by value.
The episode’s framing answers this post’s question in one line: the brain crosses the Pacific as a download, and the body has to clear customs. Software ships as a file. Bodies have to be manufactured, shipped, certified and repaired, one at a time.
On unscarcity.ai: Where to Go Deeper
Humanoid Robots 2026 is the running record of the body engine: prices, shipments, the US import ban, and the data bottleneck in detail. Robots Run on Electrons follows the Walmart bill and argues that physical automation is an energy bet, decided at the substation as much as in the lab.
In the Book: Robots That Learned to Improvise
Chapter 1, “The Foundation (Solving for Survival),” names three engines, “the brain, the body, and the fuel,” and gives the body its own section, “The body: robots that learned to improvise.” It opens on the old problem: “A two-year-old could navigate a messy room better than a million-dollar industrial arm.” The fix it describes is training thousands of digital robot twins in simulation. In the book’s words: “We’re not programming robots anymore. We’re growing their intelligence in a time-accelerated bubble.”
The section prices the result: a 1X Neo at $499 a month, about 69 cents an hour, ten times under the federal minimum wage. Then it gets honest about the limit. Tesla taught its cars by quietly logging how its customers drove, but “there is no shadow GPU for folding laundry.” Each chore has to be shown to a robot by a human, one body at a time. The verdict: “The robots can do the warehouse. The fitted sheet remains, for now, beyond them.”
A later line in the same chapter marks the outer edge of what bodies can take: “Robots can pour the concrete. They cannot attend the planning commission.” It arrives at the end of the chapter’s accounting of Maria’s apartment, after robots and fusion have taken labor, energy and materials off the bill and only permits and land remain. The book spends Chapter 8 on who pays for that remainder.
More questions like this one are collected on the blog page. Follow the series on Substack.
