Humanoid robot narratives usually begin with market size, not with the physics of the work, and that leads to distorted expectations. The common framing treats all human labor as addressable, which implies a market measured in tens of trillions of dollars if one aggregates global wages across sectors. Zach Shahan asked the question recently, Who Is Tesla Selling 1 Million Humanoid Robots A Year To?, because the claims are distinctly out of hand. A better starting point is a two-axis reality check. One axis is dexterity burden, the difficulty of physically manipulating objects and environments. The other is human-proximity safety burden, the difficulty of operating near people without creating unacceptable risk, liability, or supervision overhead. I published a lengthy piece dealing with these aspects a few months ago, citing robotics industry giant Rodney Brooks’ deep knowledge, as well as my personal explorations. When those two axes are combined with a simple market lens, with a large outer halo representing Total Addressable Market and a smaller inner circle representing the near-term serviceable market, a clear pattern emerges. The largest theoretical markets cluster where both burdens are highest, and the near-term serviceable markets cluster where both burdens are lower. Infographic of obtainable markets by humanoid robots, by author with ChatGPT. The market segmentation underlying the chart is based on a simple but structured scoring model rather than intuition. Each segment is evaluated across two primary axes, dexterity burden and human-proximity safety burden, using weighted sub-factors such as object variability, deformability, precision requirements, need for bimanual coordination, and tactile dependence for dexterity, and human exposure, injury potential, environmental unpredictability, intervention frequency, and regulatory sensitivity for safety. Each sub-factor is scored on a 1 to 5 scale, then combined into a normalized 1 to 10 score for each axis. Market size is represented as a relative index rather than an absolute dollar figure, with the outer halo reflecting Total Addressable Market based on global task prevalence and labor value, and the inner circle reflecting a near-term serviceable market based on capability fit, deployment evidence, and economic viability. The goal is not precision but consistency, so that each segment is assessed against the same criteria and the relative position of segments is defensible and repeatable rather than anecdotal. Walking is no longer the gating problem. Modern bipedal systems can traverse flat floors, handle modest obstacles, and maintain balance under perturbation. Those capabilities are necessary but not sufficient. The harder problem is manipulation. A robot picking up a rigid box with known geometry and mass is a low-variance task. A robot handling a pile of mixed objects, a tray of dishes, a length of cable, or a person’s arm during assisted movement is a high-variance task. Each added degree of freedom in a robotic hand, each added sensor, and each added control loop increases the number of failure modes. In a simplified view, if a system has 10 independent subsystems each with 99% reliability per hour, the combined reliability is 0.99 to the power of 10, or about 90.4% per hour. Increase the number of interacting subsystems to 30 and the same assumption yields 0.99 to the power of 30, or about 74%. That is not how real systems are modeled, but it illustrates the direction. Dexterity does not scale linearly. Failure opportunities multiply. That is why structured logistics shows up as an early market. The objects are standardized. The workflows are repetitive. The environment can be constrained. A warehouse moving 10,000 totes per day across 3 shifts has a clear unit of work, and the robot can be engineered to that unit. If a human worker moves 200 units per hour and costs $25 per hour fully burdened, the cost per unit is $0.125. A robot that costs $50,000, operates 6,000 hours per year, and delivers 150 units per hour with 90% uptime has an effective output of 810,000 units per year. Amortized over 5 years, ignoring maintenance and supervision for the moment, that is $10,000 per year in capital cost, or about $0.012 per unit. Add $0.03 per unit for maintenance, supervision, and energy and the total is still below the human benchmark. That is the kind of arithmetic that makes a market real. Structured manufacturing support is similar but slightly more complex. Tasks such as kitting, line-side material movement, and machine tending involve higher precision and tighter integration with existing processes. A plant producing 300,000 vehicles per year may have thousands of repetitive handling steps per vehicle. If a humanoid system can replace or augment a subset of those steps with stable cycle times and low intervention rates, the value compounds quickly. But manufacturing also has strong alternatives. Fixed robotic arms, conveyors, gantries, and custom automation often deliver higher throughput and reliability for specific tasks. The humanoid value proposition is not raw performance, it is flexibility across tasks. That flexibility has to be priced against lower reliability and higher integration cost. Move up the dexterity axis and the problem changes character. Homes, care environments, and general labor tasks are dominated by deformable objects, clutter, and exceptions. A towel does not have a fixed shape. A pile of dishes varies by material, size, and fragility. A child or an elderly person introduces motion, unpredictability, and safety constraints that are orders of magnitude tighter than those in a warehouse. If a robot drops a box in a warehouse, the loss is a damaged unit. If a robot drops a glass near a child, the risk profile changes immediately. If a robot assisting an elderly person misjudges force by 10%, the outcome may be injury. These are not marginal differences. They are structural constraints on deployment. Safety is not only about catastrophic failure. It is also about frequency of minor events and the cost of managing them. A system that requires human intervention once every 10 minutes is not autonomous in any economic sense. If each intervention takes 30 seconds of human time, that is 3 minutes of human labor per hour of robot operation, or 5%. At $30 per hour for a supervising worker, that adds $1.50 per hour to operating cost before maintenance. If the robot itself is saving $10 per hour of labor, that 15% overhead matters. Increase the intervention frequency to once every 5 minutes and the overhead doubles. In public or home environments, intervention is not just a cost. It is a safety and trust issue. This is where the distinction between Total Addressable Market and a serviceable market becomes critical. A common claim is that robots can address a $20 trillion global labor market. That number is not wrong in the abstract, but it is not actionable. If only 5% of that work is technically and economically accessible in the next decade, the real market is $1 trillion. If only 1% is accessible, it is $200 billion. And that accessible portion will not be evenly distributed. It will be concentrated in domains where dexterity and safety burdens are manageable. The bubble chart makes this visible. Large outer halos sit in the upper-right corner, representing homes, eldercare, and general labor. Small inner circles sit inside them, representing what can actually be done in the near term. In the lower-left and center-left, the halos are smaller but the inner circles occupy a larger share of the area. Remote and hazardous inspection sits somewhere in the middle. The human-proximity safety burden is lower because people are not present, but the environment can be unstructured. A refinery inspection task may involve stairs, valves, and uneven surfaces, but also heat, corrosion, and limited visibility. Drones, tracked robots, and fixed sensors already cover parts of this space. A humanoid form factor may be useful where human-designed infrastructure dominates, but the market is narrower than the general labor narrative suggests. Public service environments, retail, and hospitality are often used in demonstrations, but they sit higher on the safety axis. A robot navigating a crowded store or airport must handle unpredictable human motion, varying lighting, and social expectations. The liability environment is different from that of a warehouse. A near miss in a warehouse is an internal issue. A near miss in a public space can become a reputational event. That shifts the economics. Insurance, certification, and risk management costs increase. The serviceable portion of the market shrinks. Construction and agriculture are large markets in labor terms, but they are high on both axes. Outdoor environments introduce weather, terrain variability, and material inconsistency. A construction site changes daily. A field varies by soil, moisture, and crop. The dexterity required to handle tools, materials, and living plants is high. The safety environment includes human workers, heavy equipment, and dynamic conditions. There may be niche opportunities, but broad adoption will require substantial advances in both manipulation and autonomy. For companies building humanoid robots, including Tesla with Optimus and a range of Chinese entrants, this framework reframes the competition. Manufacturing scale matters, but it is not the primary constraint. If a company can produce 100,000 robots per year but can only deploy them economically in a few constrained workflows, production capacity will not translate into revenue. Conversely, if a company proves a workflow with strong economics, suppliers will emerge to support scaling. The competitive advantage lies in proving useful work, not in announcing capacity. China’s manufacturing ecosystem can likely compress hardware costs once demand is proven. Tesla’s vertical integration and software stack may offer advantages in system design and deployment. But both face the same underlying constraints. A robot that can move boxes in a warehouse does not automatically become a robot that can do laundry or provide eldercare. Each step up the dexterity and safety axes requires new capabilities, new validation, and new regulatory acceptance. The evidence that would change the current assessment is straightforward. It is not more demonstrations of mobility or manipulation in controlled settings. It is data from real deployments. Robot-hours in production environments, measured in the tens of thousands. Task success rates above 99% for defined workflows. Intervention rates below one per hour, ideally much lower. Throughput comparisons against human workers and existing automation. Maintenance costs expressed in $ per operating hour. Safety incident and near-miss data. Repeat orders from customers after initial pilots. Those are the metrics that convert a theoretical market into a serviceable one. None of this argues that humanoid robots will not be important. It argues that the path to importance is narrower and more incremental than the broad market narratives suggest. The early value will come from constrained niches where the physics, economics, and safety can be aligned. From there, capabilities may expand and markets may grow. But the sequence matters. Starting from a $20 trillion labor market and working backward to technology is a recipe for overestimation. Starting from a specific task, with defined objects, environments, and safety requirements, and building outward is how real markets form. The more human-like the task, the more it bundles dexterity, perception, judgment, and safety into a single requirement, and the less attainable it is in the near term. That is the central constraint on humanoid robot markets today. The answer to Zach’s question about who will buy Tesla’s Optimus is likely Tesla itself for its factories, but it won’t need a million of them.
+75 others
