Almost every discussion of humanoid robots treats intelligence as the hard part. Foundation models learned to see, plan and manipulate, the argument goes, and the rest is engineering. That framing was reasonable three years ago. It is now wrong, and the evidence is sitting in the spec sheets of every platform actually deployed in a factory today.
The average humanoid in 2026 carries under 2.5 kWh. Unitree’s H1 runs a 0.864 kWh pack good for under four hours of static operation. Tesla’s Optimus Gen 2 carries roughly 2.3 kWh and manages about two hours of dynamic work. Across the field, runtimes cluster between two and four hours, and the best lithium-ion cells going into these machines sit around 280 to 300 Wh/kg, which is close to the practical ceiling for the chemistry. Everything downstream of that number, the unit economics, the deployment model, the addressable task list, is constrained by it.
The form factor fights the battery
An electric car is a flat chassis with thousands of cells laid into the floor. Weight sits low, cooling is straightforward, and adding capacity mostly adds range. A humanoid has none of that. The pack goes in the torso because that is the only volume available, which puts mass high and forces the balance controller to work harder. Bipedal locomotion is expensive even when nothing is happening, because the actuators fire continuously just to hold a standing pose.
Then there is the loop that makes the problem genuinely hard rather than merely annoying. Add a kilowatt-hour and you add mass. Extra mass increases the torque demand at the hip and the knee, which increases consumption per meter walked, which eats into the runtime the extra capacity was supposed to buy. Past a certain pack size the returns go flat and then negative. This is the structural difference between a humanoid and every other battery-powered product category, and no amount of cell-level improvement removes it. It only moves the inflection point.
Discharge behavior compounds it. These machines need burst current for grasping, catching, recovering from a stumble, lifting a heavy box. High-dynamic platforms want instantaneous rates in the 5C to 15C range with peaks well above that. Optimizing a cell for energy density and optimizing it for peak power are different exercises, and the humanoid duty cycle demands both simultaneously in a package with almost no airflow. Figure’s engineers have made the point publicly that repurposed EV cells do not work here. The discharge profile, the thermal envelope and the physical packaging all have to be custom.
The economics break at the duty cycle, not the sticker price
Every humanoid pitch deck eventually shows a price target. Tesla says $20,000 to $30,000 at volume. Unitree’s G1 already sells around $16,000. Those numbers are meant to be compared against an annual wage, and the comparison looks devastating.
It is also the wrong comparison. What a plant manager buys is coverage of a work lane, and coverage is a function of duty cycle. Figure bills BMW somewhere around $25 per robot-operating-hour at Spartanburg. If a unit works two hours and then needs an hour on a charger, staffing one lane across two shifts requires three or four robots, a charging bay, spare packs, floor space for all of it, and somebody to manage the rotation. The capital cost of the deployment is the unit price divided by the duty cycle, and at current runtimes that divisor is brutal. Agility’s CTO has framed the goal correctly: the target is not continuous runtime but matching charge added to work performed inside the operation’s cadence. That is a scheduling constraint dressed up as a hardware spec, and it is why long uninterrupted duty profiles still go to fixed industrial arms or wheeled platforms.
Notice which deployments are real. Figure 02 accumulated over 1,250 operating hours at BMW across multiple units running ten-hour days. That worked because the task was parts handling in a fixed cell inside a plant that could build charging infrastructure around it. It is not a general-purpose result. It is a result about a specific task with a specific power profile in a building with spare electrical capacity.
Better models make the problem worse
Here is the part the intelligence-first framing misses entirely. Onboard inference is a continuous parasitic load. Vision-language-action models running locally draw real power, and they draw it whether the robot is moving or standing. Every improvement in capability that comes from a bigger model or a higher control frequency or more camera streams is paid for out of the same pack that runs the actuators.
So the two curves point in opposite directions. Model capability is improving fast. Cell energy density is improving at a few percent a year. The smarter these machines get, the more of their fixed energy budget goes to thinking rather than working, unless the compute side delivers efficiency gains at a rate the silicon roadmap does not currently promise.
The workarounds are logistics, not solutions
Two approaches dominate. Hot-swapping is the mature one, with Agility’s Digit and Apptronik’s Apollo both designed to change packs without a reboot, which gets you close to round-the-clock uptime. It works. It also converts an engineering constraint into an operating cost: spare pack inventory, swap labor or a swap robot, charging infrastructure, and the safety procedures that come with handling live high-nickel packs on a plant floor. That cost is invisible in a $25,000 price tag and very visible in a total cost of ownership model.
The other approach is waiting for chemistry. Xpeng’s IRON, GAC’s GoMate and EngineAI’s T800 have gone to solid-state, and TrendForce projects humanoid-driven solid-state demand reaching 74 GWh by 2035, up more than a thousandfold from this year. The energy density case is real. The rate capability case is not settled. Solid-state has historically traded high-current performance and low-temperature behavior for density, which is exactly the wrong trade for a machine whose defining requirement is explosive multi-joint actuation. Assuming a clean substitution is optimistic.
Neither path is a reason to be bearish on humanoids. Both are reasons to be specific about where they land first. The near-term deployments will keep concentrating in environments with predictable task cadence, short cycle times and existing electrical infrastructure, which describes automotive plants and large logistics facilities and very little else. Home robots and untethered field work sit on the far side of a battery problem nobody has solved.
The interesting consequence is competitive. If intelligence converges toward parity across platforms, and there is decent reason to think it will, the differentiation moves to pack architecture, thermal design, swap mechanics and battery management. The company that wins humanoids may end up being the one with the best power system rather than the best policy model. That is a very different industry than the one currently being funded.
Leave a Reply