Two numbers were on display at WRC 2026, and most readers will have filed them under the same heading. UBTech booked over 2 billion yuan in 2025 revenue. Sunway New Materials logged 210 million yuan for its Qiyuan robots. Both got applause. Only one is money the company is allowed to call money.
That distinction is the whole story of Chinese embodied AI right now, and the industry has every incentive to keep it fuzzy.
Start at the AiMOGA booth, because it is the emblem the sector has chosen for itself. More than 30 screens cover one wall, and not one shows a robot dancing.
Each carries a live feed from a different city: a robot in traffic-police uniform at an intersection, another in a car dealership, another out on patrol. "Our robots are already working in these cities," said Zhang Guibing, AiMOGA's general manager and Chery International's president. The traffic-police units are on duty in 31 cities across seven provinces.
It is a deliberate anti-dance statement, and it lands. But look at what a live feed actually proves. It proves a robot is somewhere. It does not prove anyone paid to keep it there, or that they will pay again next quarter.
A live feed proves a robot is somewhere. It does not prove anyone paid to keep it there, or that they will pay again next quarter.
That is the only question that separates the four kinds of "traction" crowding the WRC floor, and it maps almost exactly onto how an accountant would sort them.
Take UBTech first, because it is the one showing the hard version. In 2025 the company reported over 2 billion yuan in revenue and sold 13,838 robots. Inside that, full-size embodied units brought in 820 million yuan across 1,079 machines.
The load-bearing detail is not the headline number. It is where those 1,079 went: over 80 percent into automotive, logistics, semiconductor, aerospace and industrial-data sites — the harsh, continuous, unglamorous floors where a robot has to work day after day.
UBTech built no display set this year. "We moved the scenes we are actually deploying onto the show floor," vice president Jiao Jichao said. The Cruzr units ran continuously from the 9 a.m. open to the 5 p.m. close, no timed demos, no human intervention.
This is recognized revenue against real customers. A factory paid, the robot stayed, and — the phrase that matters — it was still running the next day. Huawei and Foxconn began buying units for their lines this year.
And the deployment feeds a loop that is genuinely hard to fake. Real-machine data now makes up 60 to 70 percent of UBTech's training mix; the company generates thousands of hours of raw data a day and has started selling that data to head embodied-model customers. Deployment produces data, data improves the model, the better model wins more deployments.
That loop is the actual moat. It is not shipment count, and it is certainly not booth applause.
Now Qiyuan. Sunway New Materials opened public pre-orders for the Qiyuan Q1 and T1 on the last morning of WRC, and the headline was 210 million yuan.
Read the financial report and the number changes character entirely. That 210 million is advance receipts. It sits on the liability sheet. It has not been recognized as revenue, the report states plainly that no operating income has been generated, and the two products remain in "R&D and application-scenario verification."
Delivery is only promised for September. So the 210 million is not a P&L line — it is a stack of IOUs, 1,000-yuan deposits from customers who have been shown a robot and told to wait.
The offer itself is telling: pay 1,000 yuan, deduct 2,000 from the final price, and receive an autographed letter from Peng Zhihui, a founding-user card, and a limited-edition figurine. That is a consumer pre-sale, and pre-sales are precisely the thing that has not yet been tested against the one question — will anyone keep it running the next day and pay for it in full?
This is where the scrutiny rhymes with the Anthropic gross-versus-net debate playing out elsewhere in the model-lab world. The mechanism differs — one is a hardware balance sheet, the other is a services P&L — but the instinct is identical: a large, real, impressive number is presented one layer earlier than the number that would actually settle the argument.
Then there is Qianxun, which represents a third and more honest category. It did something odd at WRC: it re-ran the exact "tidy the living room" demo from WAIC a month earlier, prompting a reasonable visitor to ask whether it was reheating leftovers.
The answer was in the deltas. Over under 30 days, the autonomous-reasoning success rate for putting cola into the fridge rose from about 80 percent to over 99 percent; bowls into the dishwasher went from roughly 90 to over 99; navigation target-confirmation time dropped nearly 50 percent.
Qianxun calls this "time density" — how fast a system can absorb a failure, retrain, and push the fix back to the real machine. It is genuine engineering progress, measured on the metric that matters for long-horizon tasks, where a single early slip cascades down the chain.
But be precise about what it is. It is a controlled iteration snapshot — a booth under stress-test, with visitors deliberately blocking the robot's path. It is not a signed P&L. Qianxun does have industrial deployment (its Moz1 is reportedly on a CATL battery-pack line above 99 percent success), but the WRC evidence itself is an improvement curve, not a customer keeping the machine.
Line the four up and the hierarchy is clear. A live feed proves presence. A success-rate delta proves engineering velocity. An advance receipt proves interest. Only recognized industrial revenue proves someone paid to keep the robot working.
The industry blurs these together because "traction" is a word that flattens all four into one. WRC's own framing — "from the stage to the shop floor" — invites the flattening.
But the real dividing line WRC surfaced was never dancing-versus-working. As UBTech's own closing pitch put it: the coming gap concentrates on one question — whose robot gets kept by the customer to work again the next day.
That reframing has teeth for founders and investors. It says pre-orders and consumer figurines cannot shortcut the moat, because the moat is the deployment-to-data-to-iteration loop, and you only enter that loop once a robot is on a real floor generating real failures.
Qiyuan's 210 million may well convert. Unitree's IPO closed up 460 percent the day WRC opened, and the capital is clearly there. But conversion is a future event, and the balance sheet is honest about that even when the press release is not.
So when you see a wall of 30 screens, admire the engineering and then ask the accountant's question. Somewhere is not the same as paid. Working today is not the same as kept tomorrow.
Right now, exactly one Chinese embodied-AI company has audited proof of the second thing. Everything else is a snapshot or an IOU — and knowing which is which is the entire discipline this sector still has to learn.