Goldman Sachs has torn up its own forecast for humanoid robots and replaced it with a much larger one. In a new 80-page report on what it calls "physical AI," the bank now expects roughly 6.5 million humanoids to ship annually by 2035, up from about 1.4 million previously, a nearly fivefold jump. It raised its 2030 estimate from 256,000 units to 890,000, and its near-term 2026 figure from 51,000 to 75,000. The market those machines represent, in Goldman's telling, grows from 38 billion dollars to about 138 billion. The revision is not driven by a single breakthrough but by a change of mood: the bank increasingly treats reliability, cost and the shortage of training data as engineering problems to be solved rather than reasons to stay out.
Big round numbers are easy to distrust, and Goldman is careful to hedge. What makes this forecast worth a second look is a much smaller figure that arrived alongside it. JPMorgan estimates that a humanoid could soon cost roughly 10 dollars an hour to run inside a warehouse, against about 30 dollars for the human doing similar work. Today it takes about two robots to match one worker's output, so the real comparison is nearer 20 to 24 dollars for the same job. By 2030 the bank expects that gap to close to around 1.2 robots per worker, pulling the effective cost toward 12 to 16 dollars an hour. That is the number that changes the arithmetic on a factory floor, and it lands in the middle of a real shortage: JPMorgan counts about 462,000 unfilled US manufacturing jobs today, potentially 1.6 million by 2030.
The early deployments are concrete enough to test the theory. Over a ten-month run at BMW's plant in Spartanburg, Figure's robot moved more than 90,000 components and logged some 1,250 hours across the production of over 30,000 vehicles; a newer model is now doing harder, less scripted picking work. Amazon already operates more than a million robots across 300-plus facilities, and Goldman reckons its automation push could yield 72 billion dollars in cumulative savings by 2030. Hyundai plans to build as many as 30,000 Boston Dynamics Atlas units a year by 2028. The obstacles are equally concrete: dexterous hands remain a bottleneck, integration is expensive, and the physical world offers none of the free training data that the internet handed to chatbots, which is why Figure says it has paid contributors to upload 16 million videos of ordinary human tasks.
Goldman's sharper point is about where the profit sits, and it is not with the robot brands. The bank estimates each humanoid carries 3,000 to more than 6,000 dollars of semiconductor content: edge computing, analog and mixed-signal chips, memory, sensors. At 6.5 million units that is 19.5 to 39 billion dollars of chip demand a year, a new leg of the same AI-infrastructure build-out already reshaping the industry. The advice to investors is old as gold rushes: forget guessing which humanoid wins the popularity contest, and own the picks and shovels, the processors, actuators and machine vision that every robot needs regardless of the badge on its chest.
Two cautions keep the story honest. The forecasts diverge wildly the further out they run, from Goldman's 6.5 million by 2035 to Morgan Stanley's 1 billion by 2050 and RBC's 9 trillion dollar market, which is a polite way of saying nobody knows. And the whole build cycle rests on AI capabilities continuing to improve at pace, the very thing that the industry's own leaders spent this week arguing they should slow down. A quintupled forecast is a statement of confidence. Whether it becomes a shipment schedule depends on machines proving, job by unglamorous job, that they can earn their ten dollars an hour.