The iPhone Moment for Robots Just Happened: Why Machine Hours Rewrite Labor

Three robots named Bob, Frank, and Gary worked non-stop for days in a California warehouse for Figure AI. They sorted packages at human-parity speed, swapped when batteries ran low, and put the runtime counter on a giant screen so the internet could watch. Past 60 hours and still climbing when that clip was recorded. That is not a CES stunt. That is the kind of break that looks obvious only after it happens.

June 29, 2007 was the day the original iPhone went on sale. Before that day the phone was a gadget. After that day it reorganized wallets, maps, cameras, dating, and work. The hardware was mostly ready in 2005. The software was not. When those two converged, the economy rearranged around one object. Twenty years later that market is about $4 trillion a year. The same kind of inflection just hit the human body. The market it touches is not $4 trillion. It is about $40 trillion a year of human physical labor: warehouses, harvests, packages, hospital lifts, cars bolted together. Roughly a third of global GDP paid for hands and feet moving through space.

The doomer story (robots end us) is wrong. The cheerleader story (instant utopia) is also wrong. The real story is weirder and arriving faster than most models price.

Three S-curves hit together

Humanoids stopped being a punchline when three curves bent at once.

First, the brain: vision-language-action models. Nvidia researcher Jim Fan’s framing of Groot is clean. One network takes camera pixels, a language instruction like “pick up that red tool,” and the robot’s body state, then outputs continuous motor commands hundreds of times per second. No hand-written “if red then close gripper at 30% force.” Same family of architecture as large language models and Tesla FSD, pointed at a body. That was not practical until very recently because you need massive paired data of senses, actions, and language.

Second, the data flywheel. Tesla has been collecting factory manipulation data with Optimus and years of fleet FSD data across millions of vehicles. Nvidia’s Cosmos generates synthetic robot trajectories on GPUs that look real enough to train on. The brain finally has food.

Third, the body, meaning actuators. A full humanoid needs at least 28 actuators, often more. In 2020 a high-torque hip or knee actuator could cost $2,000 to $5,000 at low volume. The joint pack alone could hit $50,000 to $100,000. In 2026 the same class of actuator is often $500 to $2,000 at low volume. At Tesla scale, Morgan Stanley sketches $100 to $300 each once volumes hit a million units a year. China’s EV supply chain is why. The motor and gear makers behind BYD and peers make humanoid joints too. Unitree sells a full humanoid for about $13,500. Bank of America sees material cost falling from roughly $35,000 toward $13,000 to $17,000 by 2035. IDTechEx sees average selling prices falling from about $115,000 to $37,000 over four years. Tesla’s Optimus target is around $20,000 by the time they hit a million robots, aimed near 2030.

Brain plus data plus body in the same 18-month window means parts are available and volume is about to show up everywhere at once.

They are already in plants

Figure’s bots are headed into BMW’s Spartanburg body shop, sheet metal, and weld inspection. Apptronik’s Apollo is in Mercedes plants, Jabil electronics sites, and GXO warehouses. Agility’s Digit crossed 100,000 totes at a GXO site in Flowery Branch, Georgia. Tesla Optimus is in internal pilots at Fremont and Giga Texas with mass production planned to start this summer. Figure’s Brett Adcock said the BotQ factory went from one robot per day to one per hour in about four months, a 24x ramp in 120 days, with a calendar-year plan north of 12,000 and a long-term aim of a million per year. Live streams since mid-May show end-to-end neural control from camera pixels, with teleop only when something breaks.

Wave one fills vacancies, not pink slips

The first deployments go where humans already refuse or cannot be hired: overnight warehouse shifts, hot and cold extremes, high turnover roles. ManpowerGroup’s 2026 survey put Japan employer hiring difficulty at 84% and Germany at 83%. Japan’s elder-care sector shows roughly 3.9 jobs per applicant and a projected shortfall of about 570,000 care workers by 2040 with nearly 30% of the population over 65. Germany crossed a point where more workers retire than enter. Sites like BMW South Carolina, GXO Georgia, Amazon warehouses, and Tokyo care facilities are not first because executives wanted mass layoffs. They are first because the seats were empty. The short-term story is robots filling vacancies nobody wanted, not stealing jobs from a line of eager applicants.

Wave two is the hard politics

Around 2030 the middle wave hits roles people actually want to keep: fast food prep, lower-turnover warehouse work, light manufacturing, hotel housekeeping, retail stocking, framing, large-scale harvest. The U.S. alone has on the order of 7 million material-mover jobs near a $30,000 median wage, about 2 million assemblers near $44,000, and millions more in nearby categories. Combine those and you are past 10 million American workers in directly automatable physical roles before counting Europe and Asia. That is where the story stops being demographic gap-fill and becomes a real transition.

If cost curves and software generalization collapse into the same five-year window, the middle of the workforce that moves things for a living can get crushed inside a decade. That is the scenario Farzad worries about in Abundance or Collapse: robots scale faster than government competence, millions of handlers lose work in a few years, and politics answers too late or so badly it turns violent. People who learn to build with the stack will fare better than people who refuse to look. The window for ignoring it is closing.

One flywheel, not a niche industry

Humanoids are not a standalone sector. They are the manufacturing face of one stack that also includes AI models, batteries, vehicle production, and data centers. Tesla builds Optimus inside factories that already make millions of cars a year, on the same AI training path as FSD and Grok. Figure builds with actuators from the same Chinese supply chain that feeds EVs and home batteries. Nvidia’s Cosmos trains humanoids on the same GPU class that trains language models. Pause the flywheel and you pause AI, batteries, EVs, and competition with China. Those are politically untouchable. So every warehouse, farm, hospital, kitchen, and job site eventually sits in the path of something that does not pause cleanly.

The iPhone analogy fails in one critical way. Phone-era displacement was mostly invisible over a decade. Humanoid displacement is visible. Bob, Frank, and Gary have name tags on a livestream. Headlines will pair a robot at a station with a person who used to stand there, even when the underlying job count is still filling vacancies. Politics will respond to the photo, not just the spreadsheet.

Short term, the economic case is still vacancy fill, demographic repair, and new categories of work that do not exist yet. Longer term, the iPhone moment for physical labor is already on camera. The question is whether institutions adapt as fast as the actuator cost curve.

Check the video here.