In the 2025 embodied intelligence market, companies are avoiding buying robots due to rapid technological iteration, shifting to leasing models to mitigate risk, emphasizing paying for functionality rather than hardware.
Toward the end of 2025, the mood in the embodied intelligence circle was strange in a way that was hard to fathom.
Over on the supply chain side, it was practically a contest of “who has it worse.” To clear inventory, some manufacturers were willing to slash prices down to the floor, creating the illusion that robots were about to become as ubiquitous and cheap as cabbages. By all logic, that should have been the moment for the market to rush out and snap up the bargains, right?
But an utterly ironic scene unfolded on December 22.
At the Shanghai Robot Rental Ecosystem Summit, I didn’t see a downturn—I saw a crowd so thick you couldn’t even get through the aisles. That stark contrast created a moment of magical realism: everyone was paying lip service to “technology for all,” yet their actions were voting with their feet—nobody dared to buy.
That glaring red figure on the summit’s big screen—“next year’s market size aims for 10 billion yuan”—caught my eye, but honestly, I had my doubts. As a veteran observer, I know that number carries plenty of water. But after chewing through tens of thousands of words of meeting notes and talking in depth with more than a dozen practitioners at the event, from VPs to small-business owners, I picked up on a shared yet unspoken sentiment among B-side bosses: “risk aversion.”
This isn’t some high-minded thing you can dress up as “value reshaping,” and don’t even try to sell it as “shared economy” idealism. Peel back those slick PPTs, and the core logic is pretty raw, just one sentence:
Robots today iterate too fast—buying one makes you the fool.
The “Brutal Reenactment” of Moore’s Law in the Physical World
A couple of years ago, the reason companies didn’t buy robots was simple: they were too expensive. A humanoid robot could cost hundreds of thousands or even millions of yuan. Aside from research labs and companies using them as showpieces for their lobbies, no normal business could make the numbers work.
But 2025 is different—prices have come down for real. The strange part is, this year, people aren’t buying because they “don’t dare.”
That fear was palpably real at the event. Deng Taihua, a partner at Zhiyuan Robotics, told a joke on stage, clearly meant to lighten the mood. He said: “This industry moves so fast now. The product we released at the start of the year walked like an old lady, wobbling and afraid of falling. By the end of the year, the upgraded version of the same model can sing, dance, rap, and even do a backflip for you.”
The audience laughed and applauded. But sitting there, I looked at several attendees around me who were clearly from B-side businesses—like restaurant chains or property management companies—and their faces showed not a hint of amusement. Instead, their brows were furrowed.
For business owners actually writing the checks, this wasn’t funny at all. It was a horror story.
Think about it. You’re the CFO of a company. Today, you spend a fortune to purchase a batch of robots as fixed assets, expecting them to pay for themselves over three years. Three months later, a competitor down the street gets a new model through leasing.
Even if the shell looks the same, everything inside has changed. Their joint motors use the latest torque-control technology, outclassing yours in response speed and load capacity. Their batteries use a new solid-state solution that runs half a day without charging, while yours dies after two hours and needs to find a charging station. Worst of all, the dexterous hands—yours can only grab a large apple, but theirs can already pinch a sewing needle.
At that point, the batch of equipment you bought just three months ago as the pinnacle of “high tech” instantly becomes “industrial waste.” Besides collecting dust in a warehouse, or selling it at a loss to secondhand dealers who don’t even want to take it, what else can you do with it?
Right now, embodied intelligence is in its “adolescence,” the stage where hardware architectures are the least stable. Unlike the phone industry’s incremental “milking” of minor innovations, this is Moore’s Law being violently reenacted in the physical world.
I talked to a boss at a servo motor company, and he vented to me: “The solutions really change every quarter. In March we were still grinding on coreless cup motors, by June everyone had moved to linear actuators, and by the end of the year they were researching bionic muscles. At that pace, who dares to stockpile inventory? Who dares to buy outright?”
In times like this, “holding” is the biggest risk.
So, it’s less that the leasing model is so advanced and more that it was forced into existence. Companies would rather pay a seemingly wasteful monthly fee than carry the heavy burden of “obsolete technology.” And this model is purely a self-protection measure—a “platform ticket” they buy so they can get off the train at any moment.
Paying for “Capability,” Not for “Scrap Metal”
Beyond the fear of obsolescence, there’s another painful realization behind this wave of leasing adoption—one that many companies learned only after paying tuition: a robot without content is just a pile of scrap metal.
Think back to 2024, when a bunch of so-called “budget robots” hit the market. They waved the banner of democratization and pushed prices very low. Many bosses, tempted by the novelty, bought a few.
And then what? They brought them home and discovered these things had “empty heads.”
Sure, it could walk. Sure, it could avoid obstacles. But then what? Nothing.
Take a scenario like Haidilao or similar restaurant chains. Are they really introducing robots just to carry plates? If it were just about moving plates, a delivery robot costing a thousand yuan would be far more reliable than a humanoid. What they want is the “wow factor” a humanoid brings—that emotional value.
They need the robot to sing “Happy Birthday” to customers celebrating a birthday, to bust out a viral dance routine like “Subject Three” for bored guests in the waiting area, or even to pick up a calligraphy brush and write a “Fu” character as a blessing during the Spring Festival.
These features look simple in videos, but getting them to actually work on a specific machine is a chasm for non-tech companies.
Where is a hotpot restaurant owner or a mall operator supposed to find an engineer who understands ROS (Robot Operating System)? Where do they find a prompt engineer who knows how to tune large language models? Where do they find an expert who can write motion-control algorithms? Even if they could, the salary cost alone might be ten times the price of the robot.
It’s completely unrealistic.
So, the reason platforms like “Qingtian Rent” are taking off now isn’t fundamentally because they have lots of robots—it’s because they’ve packaged all these “extras” into a turnkey solution. It’s a bit like phones today: the hardware itself isn’t worth much anymore, sometimes it’s even free. What’s valuable is the App Store on top—the WeChat, Douyin, and Honor of Kings you can download and start using instantly.
I caught a very telling detail at the summit. A client who does exhibition planning was talking to a rental provider about their needs. He didn’t ask a single question about torque, or how many Tops of computing power the robot had. All he asked was: “I’m putting on an anime convention. Can this robot dance to ‘Guren no Yumiya’? Can it recognize cosplayers’ outfits and greet them?”
The rental provider simply pulled out a tablet, opened the backend, and showed him: “No problem. We’ve connected to Zhiyuan’s Lingchuang platform. This skill pack was updated yesterday. Let me show you a demo.”
See the point? Businesses now rent robots the way they order food at a restaurant.
That “ready to use out of the box” satisfaction is the real reason B-side customers open their wallets. As for whether it’s the “Lingchuang platform” or some large model behind it, the customer doesn’t care. They only care whether it can immediately attract customers and generate revenue right now.
The “Shared Bike Moment” for Embodied Intelligence
The industry used to love throwing around a term: Sim2Real. It sounds super geeky and cutting-edge, but in the practical world of commercial deployment, it’s basically a nightmare.
In 2023 and even early 2024, if you wanted to rent a robot for a mall event, the process went like this:
A week before the event, the technical team had to haul the robot over. First, map the environment—even for a modest atrium of a few hundred square meters, they had to wheel the robot around multiple times. Then tune the parameters: Is the floor slippery? Will the lighting interfere with the cameras? Is there reflective glass nearby?
After a week of that, the techs were exhausted, and all of it was just for maybe four hours of actual use on event day. And during that time, if the mall moved a plant or changed a carpet, the robot could lose its mind and spin in circles.
With that kind of efficiency and cost, anyone who rented would be driven to despair. What’s called “poor generalization” shows up on a business spreadsheet as a pathetically low ROI.
During the group interview session, I specifically cornered Zhiyuan’s Jiang Qingsong with a very blunt question: “Can the current technology actually eliminate the debugging phase entirely? If not, leasing is a false premise.”
Jiang didn’t bury me in obscure algorithm theory. He directly brought up two concepts: the “cloud brain” and “plug and play.”
He said the logic has changed. Robots no longer need to memorize every map or tweak parameters for every floor material. Today’s VLA (Vision-Language-Action) large models give robots general knowledge. “Take the same dance. Different clients want it. The robot just connects to the linked platform, clicks one-click download, and the robot automatically learns the capability. Then it can execute it on command—out of the box, ready to use.”
As a journalist, I remain somewhat skeptical about truly “zero debugging”—on-site network conditions are always a wild card—but based on the live demos at the summit, it’s definitely a world away from the old “babysitting a diva” level of maintenance.
If embodied intelligence can truly become like a shared bike—scan the code, unlock, ride away, whether it’s concrete or asphalt—then this market will have finally arrived. Otherwise, it’s just a toy stuck in the lab.
Putting a Price on “Uncontrollable Fear”
Of course, the biggest landmine with leasing isn’t actually the technology—it’s liability. And that was the most discussed topic among B-side bosses outside the venue.
In a crowded mall, kids running wild everywhere. If a few-hundred-pound hunk of metal suddenly malfunctions and plows into someone, who settles that account? The lessee? The platform? The manufacturer?
Conversely, if a robot gets knocked over by the crowd, or someone spills Coke on it and it shorts out, who eats the loss on that valuable piece of equipment?
In the past, these were endless arguments. Contracts could run dozens of pages filled with disclaimer clauses. But one interesting takeaway from this summit: insurance companies finally stepped in.
And not with some vague handshake partnership. They’ve actually launched dedicated “robot body insurance” and “third-party liability insurance” products specifically for embodied intelligence.
I think that’s more important than any algorithm breakthrough.
Why? Because when insurance companies—the shrewdest, most risk-averse institutions on the planet—start being willing to price the question of “will a robot cause trouble,” it means reliability has finally reached a passing grade. It means their actuaries ran the numbers and concluded this thing is probably safe, and even if something happens, the odds are within a manageable range.
Plus, rental providers have wised up. They know clients are afraid of liability, so they’ve set up a strict “access mechanism.”
Jiang Qingsong mentioned they have a registration and promotion system for “robot operators.” Just like you need a license to run an excavator, operating a humanoid robot requires certification.
Combined with backend data monitoring, the safety factor is maxed out—even if the vibration curve on a joint motor deviates slightly, the cloud often detects it before anyone on-site does.
This combination of “insurance backing + licensed operators + data monitoring” doesn’t guarantee perfection, but it transforms “unknown fear” into “calculable cost.” For business owners, that’s enough. They can sign on the dotted line.
Final Thoughts
After all this, it might seem like leasing is the perfect answer for embodied intelligence deployment, right?
Actually, no.
As I sorted through my notes and looked back at those exciting numbers, something kept nagging at me. As someone who believes “tech shouldn’t run cold,” I feel a responsibility to throw some cold water on the hype while everyone’s heads are hot.
Leasing is great—it solves the cost problem and the technology anxiety—but it’s also causing companies to lose a form of “control.”
And this isn’t just about money.
Think about it. When the robots you rent are running around your factory floor or your retail store every day, where do the images captured by their cameras, the data collected by their sensors—like your production cadence, your foot-traffic heat maps, even your employees’ work patterns—ultimately end up?
On your servers? The rental platform’s? Or the robot manufacturer’s?
In current leasing contracts, the definition of data ownership is often murky, sometimes even one-sided take-it-or-leave-it terms. Platforms frequently default to claiming usage rights over the data under the guise of “algorithm optimization.”
If one day you decide not to renew, or the platform goes under—which is all too common in the startup world—can you take that accumulated business-critical data with you? Or will you find yourself “held hostage” by the platform?
It’s like using SaaS software for years only to discover you can’t export your data. That sense of helplessness is devastating.
Nobody’s talking about this right now. Everyone’s busy grabbing market share, chasing scale, and chasing that 10-billion KPI. But beneath the noisy “billion-dollar expectations,” this vague handling of data ownership—and the fragility of a supply chain so heavily dependent on cloud services—could be the biggest landmine ahead.
In 2025, robots have gone from “exhibits” to “laborers.” There’s no arguing with that. But who does this laborer actually answer to? And in whose hands does its “soul” reside?
That’s probably the question that deserves the most discussion next year—and the one most likely to be overlooked.