There is a theory of why software AI conquered the world in three years, and it is not the one in the keynotes. The keynote version says the models got good. True, but incomplete. The quieter half is that the models got good at a price of roughly twenty dollars a month — or free — delivered through a device billions of people already owned. No installation. No shipping. No customs officer in Apapa asking what is in the crate.
I could sit in Lagos, on the same day as an engineer in San Francisco, and use the same frontier model for the price of two shawarmas. That has essentially never been true of any transformative technology at launch. Electricity, cars, computers, the internet — all of them rolled out rich-first, spreading over decades. Software AI skipped the queue because its marginal cost rounds to zero and its distribution channel was an app store. Accessibility was not a nice property of the boom. It was the engine.
So the question I keep asking about the next wave — AI in bodies, robots that fold laundry and stack pallets — is not when the capability arrives. It is whether the economics can ever rhyme.
What the price list says right now
For the first time, the price list is real. A Unitree G1 — a genuinely capable humanoid — ships today for about sixteen thousand dollars, and you can literally order one on Amazon. Unitree's smaller R1 runs about five thousand. 1X's NEO, built for homes and holding over ten thousand pre-orders, is twenty thousand dollars up front or four hundred ninety-nine a month. Tesla's Optimus targets twenty to thirty thousand at scale — a target worth respecting and discounting, since Tesla has missed every Optimus date since 2021 and outside customers likely wait until 2027. Industrial machines run far higher: Agility's Digit deploys around a quarter million or thirty dollars an hour; Figure bills BMW roughly twenty-five dollars per robot-hour. And maintenance adds ten to twenty percent of the purchase price, every year, forever.
Elon Musk told Forbes in June there might be a hundred million humanoid robots within five years, maybe a billion. Set the forecast aside and look at the unit economics instead: hardware AI launches at the price of a car, not the price of lunch. The twenty-dollar trick is not available, because atoms refuse to be cached.
Software AI was the first revolution in history that a junior developer in Lagos and a VP in Palo Alto adopted in the same week. The robot revolution will not work like that. It will work like cars: industrial first, financed always, rich-first by a decade.
How it will actually spread
My honest prediction has three phases, and none of them looks like ChatGPT's launch curve.
First, robots-as-labor-arbitrage. The early buyers are not households; they are operations where a robot offsets eighty to a hundred sixty thousand dollars a year of loaded wages — US warehouses, German assembly lines. At those wages the payback is months, which is why BMW and Amazon are already customers. Note the corollary: the robot is most attractive exactly where human labor is most expensive. In Lagos, where the wage the robot competes with might be three hundred dollars a month, the same machine is economically absurd for a decade or more. Hardware AI, unlike software AI, arrives first where workers cost the most — which inverts who feels it, and when.
Second, robots-as-subscription. Nobody outside the top decile buys a twenty-thousand-dollar appliance outright; NEO's four-ninety-nine a month is the actual consumer product. Expect the car-market apparatus — leases, financing, insurance, certified service — to graft itself onto humanoids almost immediately. Which also means the credit systems of each country will gate adoption as much as the technology does.
Third — and this is the wildcard — the China curve. Over eighty percent of humanoid installations last year were Chinese; Unitree alone out-shipped everyone combined, and shipments are forecast to grow several hundred percent this year. Chinese manufacturers are doing to robot prices what they did to phones and solar panels, with a five-thousand-dollar humanoid already in pre-sale and sub-appliance prices plausibly a few years out. Meanwhile Unitree open-sourced a vision-language-action model anyone can run on its hardware. The likely near-term shape of the industry: Chinese bodies, American brains — the smartphone pattern again, with the value fight happening at the software layer.
The Lagos test
I judge every technology by one question: when does it reach the person who needs it most and can pay the least? Software AI passed that test in year one, which is why adoption looked vertical. Hardware AI will fail it for years — not from malice, just physics and freight and customs and cost of capital.
But cost curves are the one thing this industry reliably delivers. Sixteen thousand today was ninety thousand two years ago; five thousand exists already; the appliance price is coming. My guess: industrial ubiquity in the rich world by the late twenties, a real consumer market there by around 2030, and the Lagos moment — the robot equivalent of the twenty-dollar subscription — sometime in the mid-2030s, probably wearing a Shenzhen badge and a payment plan.
The pattern to hold onto: software AI democratized intelligence before it concentrated wealth. Hardware AI will concentrate first and democratize later, because atoms always do. The people planning for both waves as if they behave the same are going to be exactly one decade wrong.