The humanoid scaling curve does not run through cleverer software. It runs through a building that does not exist yet.
Between 1903 and 1906, the Ford Motor Company produced approximately 1,700 Model A automobiles from a rented wagon shop at 696 Mack Avenue in Detroit, a single-story brick building roughly 50 by 250 feet, with workers carrying chassis components by hand from station to station. Three years. Seventeen hundred units. The output of a modest county fair.
In 1910, Ford opened Highland Park. Within four years that same volume cleared the line in roughly two weeks. By 1916, Highland Park was producing more than half a million Model Ts annually. The Rouge complex, built between 1917 and 1928, then absorbed iron ore at one end of the site and shipped finished automobiles out the other, on a 1.5-square-mile footprint with its own deep-water port, coke ovens, steel mill, glass plant, and 100,000-person workforce.
The historical lesson is not that Ford got better at engineering cars. The cars at Highland Park in 1913 were nearly identical to the cars at Mack Avenue in 1908. The lesson is that each order-of-magnitude jump in output was a real estate event. Ford did not scale by writing better assembly instructions. He scaled by commissioning Albert Kahn to design a building that turned the assembly itself into a property of the architecture.
Humanoid robotics in 2026 sits exactly on Mack Avenue. Figure, 1X, Apptronik, Agility, Sanctuary, and Tesla Optimus are all hand-building units at single-digit thousands per year out of repurposed industrial space, with bills of material that look closer to an aerospace prototyping shop than a consumer-electronics line. The technology is real. The demand curve is bending. The thing that does not yet exist is the Highland Park.
What Mack Avenue Actually Was, and What Changed at Highland Park
The Mack Avenue plant cost Ford $75 a month to rent. It had no purpose-built power, no purpose-built lighting, no integrated material handling. Workers fetched parts from piles. Engines, transmissions, and bodies were sourced from outside suppliers across Detroit and trucked in by horse cart. A finished Model A took roughly 12.5 labor-hours to assemble and sold for $850, against per-capita US income near $300. The cars were good. The factory was a workshop pretending to be a factory.
Highland Park, designed by Kahn and opened on January 1, 1910, did three things that the wagon shop could not. It collapsed the assembly process onto a single, column-light reinforced-concrete frame engineered for 75-foot clear spans, so the line could move continuously without obstruction. It generated its own power on site, freeing the plant from the grid limitations of early-century Detroit Edison. And it absorbed an industrial workforce that grew from roughly 12,000 in 1913 to more than 68,000 by 1916, made possible by Ford's January 1914 $5-day announcement, which doubled the prevailing manufacturing wage and converted the labor pool from a constraint into a magnet.
By 1925, a Model T came off the Highland Park line every 24 seconds. The price had fallen from $850 to roughly $260. Throughput rose by orders of magnitude while labor-hours per car fell from 12.5 to under 1.5. None of that was a software improvement. It was a building, a wage policy, and a piece of land.
Where Humanoid Robotics Sits On That Curve Right Now
Figure AI publicly described its BotQ facility as targeting up to 12,000 humanoid units per year at full ramp, in a converted production space adapted from existing square footage. Agility Robotics opened RoboFab in Salem, Oregon, with a stated nameplate capacity of 10,000 Digit units per year, in a retrofit industrial building under 70,000 square feet. 1X stood up a Hayward, California factory in roughly 58,000 square feet and sold out its first-year NEO allocation within five days of opening preorders. Apptronik builds Apollo units out of an Austin, Texas facility that resembles a research lab more than a production plant. Sanctuary AI assembles Phoenix in Vancouver out of a multi-tenant industrial bay. Tesla is repurposing legacy Model S and Model X assembly capacity at Fremont for Optimus pilot production while signaling future Gigafactory Texas integration.
The headline ambitions from these operators publicly exceed one million annualized units per company by the late 2020s. The current ground truth is single-digit thousands of units per year, hand-assembled, in buildings that were leased because they were vacant, not designed because they were correct.
That gap is the Mack Avenue gap. It is the gap between a working prototype shop and an industrial system. It does not close by hiring more ML researchers.
Why The Next Leap Is A Real Estate Event, Not A Software Event
The hard constraints on producing 100,000 humanoid units a year are not algorithmic. The control stack is converging. Locomotion, manipulation, and end-to-end neural policies are improving on a curve set by foundation-model scaling and synthetic-data generation, and that curve is steeper than the industrial one.
The hard constraints are physical.
A humanoid robot at production specification carries 40 to 60 high-torque actuators, a multi-kilowatt-hour battery, several hundred sensors, an onboard compute module, and a precision-machined structural skeleton. Manufacturing one demands die casting at 2 to 4 MW per press, CNC actuator lines at 500 kW to 1.5 MW each, injection molding, battery-module assembly under controlled humidity, vibration-isolated calibration rooms, and an end-of-line burn-in farm where every shipped unit walks, lifts, falls, and recovers under instrumented load for hours before it leaves the building.
Aggregate that into a 100,000-unit annual line running three shifts and the campus footprint lands near 500,000 to 750,000 square feet of production, plus 150,000 to 200,000 square feet of test and qualification volume, with total electrical demand in the 30 to 80 MW band, structural floor loads of 600 to 1,000 pounds per square foot in localized zones, and a substation interconnect that takes 18 to 24 months from initial engineering to energization.
None of that exists inside a converted logistics warehouse. Floor slabs at 250 to 350 psf will not host die-casting presses. Roof structures rated for HVAC and sprinklers will not host 50-ton overhead cranes. Warehouse power services of 2 to 5 MW will not feed a battery-module clean line, a die-casting cell, and an in-factory robot charging fleet on the same bus. Class-100,000 cleanroom conditions, sub-micron-stable calibration rooms, and 20 to 25 degrees Celsius humidity-controlled battery integration cannot be retrofitted into a building that was built around HVAC for human comfort.
The bottleneck on humanoid scaling is the building. It always was. It was the bottleneck in 1908 and it is the bottleneck in 2026.
The Site Profile A Robotics Highland Park Requires
The campus that breaks this constraint has a recognizable shape.
It needs 80 to 200 acres, in a single contiguous parcel, with rail or heavy-truck access for inbound aluminum billet, rare-earth magnets, battery cells, polymer feedstock, and structural steel. It needs at least 60 MW of firm electrical service, with a clear path to 150 MW expansion, sited near transmission with substation room on parcel rather than across a public road.
It needs reinforced-concrete production floors at 600 to 1,000 psf in machining and casting zones, with isolation footings under high-vibration cells to prevent harmonic interference between adjacent stations. It needs 50-to-75-foot clear-span structural bays so the assembly line can be reconfigured without restraint as the product platform evolves, because every robotics company is going to redesign its actuator topology twice before it stabilizes.
It needs indoor test arenas with reinforced catch floors, instrumented terrain surfaces, drop-test cages, and a 50-to-100-meter walking course held at controlled lighting and temperature. It needs an outdoor obstacle range. It needs an end-of-line burn-in hall sized for several hundred humanoids charging and operating simultaneously, with the electrical density that implies.
It needs a workforce shed. Highland Park did not work because of the building alone. It worked because Ford bought himself a labor pool by paying $5 a day. The robotics Highland Park needs to be sited where it can recruit and retain 5,000 to 15,000 skilled industrial workers per shift cycle, which means proximity to a metropolitan labor market, attainable housing within commuting distance, and a wage policy that pulls technicians, machinists, and assembly-line workers away from automotive, aerospace, and logistics competitors.
It needs fiber, security, and regulatory clarity. The site profile overlaps heavily with AI training campuses on power, fiber, and security. It diverges sharply on floor loading, indoor volume, and workforce density.
What Compounds For Whoever Sites It First
The 24-month window in front of the industry is not symmetric. The first humanoid company to operate from a purpose-built single-site campus at 50,000-plus annualized units captures four compounding advantages that latecomers will struggle to close.
Cycle time compounds. Ford at Mack Avenue ran a 12.5-hour build. Ford at Highland Park, four years later, ran a 93-minute build. Cycle-time compression on a purpose-built line is not 2x. It is roughly an order of magnitude, and it compounds against unit cost.
Supplier proximity compounds. Once a campus is established, actuator suppliers, battery integrators, sensor manufacturers, and machining houses cluster around the gate. The robotics equivalent of the Detroit supplier ecosystem builds itself within five years of the first anchor campus, and every subsequent operator either pays a freight-and-time penalty or relocates onto the same logistics network.
Talent magnetism compounds. The first plant that runs at scale becomes the training ground for the industry's first 10,000 humanoid technicians, line leads, and quality engineers. The second plant hires from the first. The third plant hires from the first two. Wage policy, schedule design, and human-factors engineering compound on the campus that scales earliest.
Regulatory imprinting compounds. The first operator at production volume sets the inspection regime, the safety standards, the workforce-health protocols, and the inbound-supply rules that subsequent campuses will be measured against. That regulatory pattern is set by the first physical site, not by the trade association.
By 2028, three to five humanoid companies will be producing at five-figure annual rates, and each of them will be doing so from a purpose-built single-site campus, not a distributed network of prototype shops. By 2030, the robotics-industrial campus is its own asset class, with site selection that overlaps heavily with AI training campuses on power, fiber, and security but diverges on heavy-machinery floor loading, indoor test volume, and end-effector calibration infrastructure. By 2032, fully-amortized cost per humanoid unit out of a purpose-built campus falls below $20,000, against current hand-build cost estimates ranging from $40,000 to $150,000 per unit, and the production model converts from aerospace prototyping to automotive mass manufacturing.
The companies that locate their Highland Park inside this window will dictate the cost curve. The ones that wait will buy their components from the supplier base that grew up around the first campus, hire their line leads out of the workforce that the first campus trained, and ship at unit economics set by the first campus's regulatory regime.
Robotics will not be won by the team with the best policy network. It will be won by the team that, sometime between now and the end of 2027, breaks ground on a piece of industrial land sized and powered for the work.