The Humanoid Race Just Became a Factory Race
Figure's line now builds a robot an hour and Boston Dynamics committed its entire 2026 Atlas run to two customers. The hard problem stopped being whether the robots work — and became how many you can actually make.
For years the humanoid-robot question was a capability question. Can it walk on uneven ground, can it manipulate a deformable object, can the policy generalize past the demo. Those questions haven't been fully answered — but in 2026 they stopped being the interesting ones. Two announcements this spring quietly reframed the entire race around a problem that's far less glamorous and far harder to fake: how many of these things can you actually build?
The clearest signal came from Figure AI. Its BotQ facility is now producing the Figure 03 at a rate of roughly one robot per hour, and the line is tooled to turn out about 12,000 units a year. That is not a research output. One unit per hour is a manufacturing cadence — the language of automotive plants and appliance lines, not of a robotics lab hand-building prototypes. The moment a humanoid company starts quoting throughput in robots-per-hour instead of capabilities-per-demo, it has declared that it believes the science is close enough to settled to bet a factory on.
Two opposite bets on the same belief
What makes the moment legible is that the field's two most credible players are responding to that belief in exactly opposite ways.
Figure is building capacity. BotQ is a bet on volume — a line designed to scale, priced for a world where humanoids ship by the thousand and then the ten-thousand, where the winner is whoever can flood factory floors fastest and drive unit cost down the curve. The 12,000-a-year tooling is a statement that the bottleneck Figure most fears is its own ability to manufacture, not customer demand.
Boston Dynamics is doing the inverse. Its entire 2026 production run of the fully electric Atlas is already committed to just two customers: parent company Hyundai Motor Group and research partner Google DeepMind. There is no general availability, no waitlist for the curious. The whole year's output is spoken for before it's built. Backed by Hyundai's $26 billion U.S. investment plan and a new factory targeting 30,000 Atlas units a year by 2028, Boston Dynamics is treating early production as scarce — channeling every unit into a vertically integrated manufacturing partner that can deploy them on real lines and a frontier lab that can push the intelligence layer.
These are not small strategic differences. They are bets on what the next two years reward. Figure is wagering that the market is a land grab and breadth wins — get the most robots into the most hands and let scale economics do the rest. Boston Dynamics, through Hyundai, is wagering that depth wins first — that a humanoid is only as valuable as the deployment around it, and that a tightly controlled rollout inside a manufacturer that builds 30,000-a-year capacity beats a wide release into customers who don't yet know how to use the things. One is racing to sell robots. The other is racing to use them.
Why throughput is the real benchmark now
The reason this shift matters is that manufacturing is the part of robotics that can't be demoed away. A viral video of a robot folding laundry tells you a model worked once, under conditions someone chose. A line producing a unit an hour tells you the company solved supply chains, actuator yield, assembly tolerances, quality control, and cost — the unsexy industrial problems that have killed more hardware companies than any algorithm ever did. Throughput is the benchmark that's hardest to fudge, because it shows up on a balance sheet and a loading dock.
It also resets the competitive map. When capability was the differentiator, the leaders were whoever had the best demo reel. When throughput is the differentiator, the leaders are whoever can finance a factory, secure the actuator and sensor supply, and run a line at automotive discipline. That favors players with manufacturing DNA — which is precisely why the Hyundai–Boston Dynamics pairing looks so different from a pure-play robotics startup, and why Figure's first real flex was a building, not a behavior.
None of this means the capability questions are solved. The hands are still the hard part, generalization across messy real-world tasks remains brittle, and a robot that builds at one-per-hour can still be a robot that struggles the moment the task drifts from its training distribution. Manufacturing capacity is necessary, not sufficient — you can build 12,000 robots that aren't useful enough, and that would be an expensive way to learn a lesson.
But the center of gravity has clearly moved. The question that defines 2026 isn't whether a humanoid can do the job. It's whether you can make enough of them, cheaply enough, to find out at scale — and whether you should flood the market or feed a single deployment until the playbook is proven. Figure and Boston Dynamics just gave the two cleanest answers in the industry, and they point in opposite directions. The next two years will reveal which kind of factory was the right one to build.