Author: Peng Zhao (Founder of Zhifang, Co-Founding Partner of Yunhe Capital) Original
Previously, in the article "After the Demo: The Next Trillion-Dollar Track for Humanoid Robots is Neither in the Brain Nor in the Body", I mentioned a core viewpoint: The true bottleneck for robots to leap from "laboratory demos" to "factory deployment" lies in continuous operation capability. Crossing this threshold relies absolutely not on adding a few more degrees of freedom, but on a whole set of industrial infrastructure hidden beneath the surface: training data collection, maintenance networks, spare parts supply chains, insurance, financial leasing, and more.
In today's article, we will delve into a particularly crucial yet rarely accurately calculated aspect of this infrastructure: energy replenishment infrastructure.
At the end of that article, I used the Ford Model T as an analogy. When the Model T rolled off the production line in 1908, it did not immediately transform the United States; what turned the U.S. into a "nation on wheels" was the quiet construction of gas stations, 4S dealerships, auto insurance, and driving schools in every small town over the following two decades.
In the industrial ecosystem of humanoid robots, what does the primary "gas station" look like? When robots are deployed on a large scale, will it exacerbate power shortages in factories? Let us calculate a few fundamental business accounts.
The First Account: What is Lacking is Not "Kilowatt-hours", but "Kilowatts"
In June this year, the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) jointly issued the "Notice on Jointly Carrying Out the Special Action for Real-scene Practical Training of Humanoid Robots and Embodied Intelligence in 2026", which explicitly proposed to "drive the formation of a deployment capability on the scale of ten thousand units" by the end of 2026.
Let us first calculate a macroscopic power consumption account based on this standard.
The power of current mainstream industrial humanoid robots: the official specification for Figure 03 is that a 2.3 kWh battery supports 5 hours of peak operation, translating to an average power of about 460 watts; Tesla, Inc.'s Optimus Gen-3 claims that the same 2.3 kWh battery can operate under light load for 10 hours, translating to about 230 watts.
Considering that on-site battery life is generally lower than the rated value, we take 500 watts as the calculation baseline. If operating 24/7, a single unit consumes about 12 kWh per day. If ten thousand units are deployed at full load, the annual power consumption is about 44 million kWh. The national annual power generation is on the scale of 10 trillion kWh, so 44 million kWh is only equivalent to the annual power consumption of a medium-sized data center.
From a macroscopic perspective, the power grid is under no pressure at all. But the real account book is hidden in the power distribution cabinets at the factory level.
The notice also provides two figures: over a hundred high-value application scenarios, and a deployment capability on the scale of ten thousand units. Dividing the two yields an average of a hundred units per scenario. For a hundred robots, the daily energy replenishment is about 1,200 kWh. At the domestic average commercial and industrial electricity price, these 1,200 kWh cost only 700 to 800 RMB.
The expensive part is hidden in the other half of the electricity bill.
Commercial and industrial users with a receiving transformer capacity of 315 kVA and above are subject to a two-part tariff system. In addition to paying for electricity by the kilowatt-hour, they must also pay a basic electricity fee based on transformer capacity or maximum demand. If a hundred robots charge during a concentrated window period, the newly added maximum demand will be between 100 and 200 kilowatts. According to the transmission and distribution electricity price table for the fourth regulatory cycle, the demand electricity price in the ten provinces and cities involved in the real-scene practical training ranges from 35 to 52 RMB/kW per month, corresponding to an additional basic electricity fee of about 40,000 to 120,000 RMB per year.
This amount of money has nothing to do with how many kilowatt-hours are charged, but only with the peak power at that moment.
A harder constraint is the transformer capacity.
If the factory's transformer load rate is already above 80%, any newly added load must go through the power grid capacity expansion approval process. The queuing cycle for capacity expansion projects is measured in months, while the delivery cycle for robots is measured in weeks.
This time gap implies an industrial reality: What determines whether a factory can deploy robots may not be the process sheet in the workshop, but the capacity sheet in the power distribution room.
The Second Account: Electricity Bills are Only the Surface, Batteries are Actually "Consumables"
Regarding how much it actually costs for robots to work, the industry often cites a misleading figure: a robot consumes about half a kilowatt-hour per hour, and the electricity cost is less than 0.4 RMB. Compared to the hourly wage of 20 to 30 RMB for human workers, the economic viability of robots seems self-evident.
But this account misses the largest hidden cost: battery depreciation.
Robot battery packs are generally in the range of 1 to 2 kWh, with a single pack supporting about 2 hours of operation. This means:
If operating at full load 24/7, a single robot will consume 12 battery packs a day.
Even calculating based on a two-shift system of 16 hours, it still requires the consumption of 8 battery packs.
If equipped with three battery packs for rotation, each battery pack must undergo 3 to 4 full charge-discharge cycles per day. Assuming the cycle life of lithium iron phosphate (LFP) batteries is 3,000 cycles, the service life of a single battery pack is only a short two to three years.
We amortize the battery loss into the cost per kilowatt-hour: the cost of a kWh-level quick-swap battery pack is estimated at 3,000 to 6,000 RMB, and the battery discharges about 2,700 kWh over its entire lifecycle. This means that for every kilowatt-hour discharged, the amortized battery loss is 1.1 to 2.2 RMB.
In other words, for every kilowatt-hour a robot uses, the money spent on battery wear is two to three times the electricity cost. Calculated over a year, the electricity cost for a single robot is about 2,000 to 3,000 RMB, but the battery loss reaches 4,500 to 9,000 RMB.
Therefore, for robots, batteries are logically closer to printer ink cartridges than to automobile parts.
NEVs (New Energy Vehicles) are charged once every two or three days; the battery is a durable asset with an eight-year warranty and residual value. Robot batteries undergo three or four cycles a day, are scrapped in two to three years, and their residual value approaches zero. The same electrochemical materials are assets when installed in vehicles, but consumables when installed in robots.
Assets focus on depreciation and residual value, while consumables focus on repurchase and gross profit.
The consumables business has its own gravity. Original equipment manufacturers (OEMs) have sufficient motivation to turn batteries into profit components and locking components: proprietary interfaces, proprietary battery compartments, and repurchase by the piece.
The Third Account: The Dispute over Energy Replenishment Routes and the "Free" Battery Swap Station
The current energy replenishment routes have clearly diverged into three factions. Putting these three factions together, we will find that they are betting on completely different business logics.
The first faction is the "charging faction", represented by Tesla, Inc.
Since its debut in 2022, Optimus has firmly integrated the 2.3 kWh battery pack into its torso, which is a typical centralized power distribution approach for automobiles. By Gen-3, the specification given by Tesla, Inc. is "10 hours of continuous operation + 10 minutes of fast charging", and it has applied for a patent for a standing charging station.
The charging faction's account is calculated very shrewdly: minimalist hardware, smaller space, lower initial capital expenditure, and stronger versatility. The charging faction's trump card is betting that the steady advancement of battery performance, coupled with fragmented task gaps, is sufficient to cover energy replenishment needs.
The second faction is the "battery swap faction", represented by UBTECH Robotics.
In July 2025, Walker S2 pioneered hot-swappable autonomous battery swapping globally, featuring dual-battery redundancy. The robot walks to the shelf by itself, swaps the battery by itself, and gets it done in 3 minutes. By the end of last year, they had already run through the mass production and rollout of a thousand-unit scale.
The battery swap faction does not believe in the "task gap" approach. In 24/7 continuous operation scenarios, uptime is revenue, and any waiting time for charging downtime is a deficit on the income statement.
The third faction is the "large battery chassis faction", such as Cross-Dimensional Intelligence.
Their W1 directly uses a wheeled chassis to carry a 20-kilogram giant battery, forcefully achieving an 8-hour battery life. This account is the simplest and most straightforward: trading the load-bearing capacity of the chassis for battery life, at the cost of giving up the versatility of the bipedal form in complex terrains.
If we zoom out, this route dispute is actually nothing new. In the warehousing and logistics track, unmanned forklifts and AMRs (Autonomous Mobile Robots) have long solved this problem with a combination of "automatic charging + manual battery swapping". Today, humanoid robots are merely challenging "hell mode": cramming batteries into the highly space-constrained humanoid torso, and making the robot "swap its own battery".
My judgment is: Ultimately, none of these three factions will eliminate the others; instead, they will undergo physical diversion based on scenarios. Single-shift scenarios with natural task gaps belong to the charging faction; 24/7 continuous operation heavy industry scenarios belong to the battery swap faction; and flat-ground transportation scenarios belong to the chassis faction.
Among them, the battery swap route conceals the most disruptive variable. It physically tears the energy out of the robot body itself, turning it into an independent "asset layer".
Let us compare this with NIO Inc.'s battery swap account book: the construction cost of NIO's battery swap stations dropped from about 3 million RMB for the first-generation station to about 1.5 million RMB for the third-generation station (excluding batteries); fully equipped with 21 batteries, the one-time investment per station is between 3 million and 5 million RMB, of which the major part of the equipment is the precision execution system for lifting, aligning, and locking.
But when we transpose this logic to the robot scenario, we find a fundamental mismatch: the automated mechanisms that NIO spent a lot of money to build have their costs reduced to zero in the robot scenario, because the robot itself is the execution mechanism. It walks to the shelf by itself and reaches out to swap the battery by itself.
What remains is a set of charging shelves plus a pool of batteries. The assets for a single scenario are at the million-RMB level, of which batteries account for 70% to 90%. This asset structure determines that the core of robot battery swapping is no longer "equipment operation", but pure "battery asset management".
In the future, there is a very high probability of replicating NIO Inc.'s Battery as a Service (BaaS) model: the battery bank holds the batteries, the fleet rents them monthly and pays per cycle, and the OEMs return to body manufacturing.
Conclusion
Beyond the above calculations, there are some hidden variables that have not yet entered the investment return models of various research reports. I condense them into two reefs and one gold mine:
Reef One: Fire Protection Compliance and Insurance Gap.
The battery swap room for a hundred robots is equivalent to centrally charging over a hundred lithium batteries inside the factory building. Currently, this is in a blank zone of fire protection regulations. Spraying and hazardous chemical plants may directly cancel robot deployment plans because of this; and property insurance companies may also increase premiums or even refuse insurance due to "centralized charging within the factory". Insurance is not only a new organ in the robot aftermarket, but also potentially the first veto player for battery swap rooms.
Reef Two: "Demand Electricity Fee" in the Policy Gap.
The current transmission and distribution electricity price policy stipulates that before 2030, centralized charging and swapping facilities implementing a two-part tariff system are exempt from demand (capacity) electricity fees. However, in-factory robot battery swap rooms are currently billed as ordinary industrial loads and cannot enjoy this exemption of nearly 100,000 RMB per year. This is both a real cost in cash and a window to observe policy trends.
Gold Mine: The Most Obedient "Virtual Power Plant".
Factories are most afraid of production shutdowns when meeting peak summer demand, but robot energy replenishment can naturally be shifted. If the daily 1,200 kWh energy replenishment for a hundred robots is entirely moved to the valley period of electricity prices, calculated at a peak-valley price difference of 0.6 RMB, a single scenario can save about 260,000 RMB a year, which is equivalent to having a built-in investment recovery channel. If connected to a virtual power plant, robots are not only not a burden to the power grid, but rather the most obedient flexible load in the factory.
This is essentially a problem of IoT (Internet of Things) platform scheduling and orchestration. Whoever's scheduling platform can simultaneously hold the three data streams of "electricity price curve, battery status, and production scheduling plan" will hold the operational entry point for the robot fleet.
In the robot era, the "gas station" that changes the industrial ecosystem is highly likely to look inconspicuous. It is a row of charging shelves in the corner of the workshop, a pool of batteries billed by the cycle, and an invisible AIoT scheduling software.