“Hidden Costs” and “Visible Anxiety” in Robotics Earnings Reports

Boom, but not without bruises

China’s robotics sector is booming on paper — and the latest crop of earnings reports make that plain. UBTech (优必选), Geek+ (极智嘉) and Yushu Technology (宇树科技) all posted double‑digit revenue growth in 2025, while venture rounds keep pouring in: it has been reported that Galaxy General (银河通用) closed a ¥2.5bn round and Songyan Power (松延动力) nearly ¥1bn in Series B. Yet the upbeat top‑line masks a more fragile reality. On the other side of the Pacific, it has been reported that several high‑profile U.S. and European robotics names have faltered — K‑Scale Labs dissolved pre‑production and open‑sourced tech with only $400,000 left on the books; Rethink Robotics and Aldebaran reportedly shuttered operations; iRobot filed for bankruptcy protection — underscoring how hard commercialization remains even for well‑known brands.

Revenue growth, but cash and profit problems

The headline numbers are eye‑catching: UBTech’s revenue rose to ¥2.001bn (up 53.3%), Geek+ ¥3.171bn (up 31.6%), and Yushu reported a 335% jump to ¥1.708bn. But profitability tells a different story. Most listed players remain loss‑making: UBTech posted a ¥790m loss, Yuejiang (越疆) showed an ¥84m net loss, and cloud‑native Yunji Technology (云迹科技) has lost more than it earned over several years. Spending is still high — combined sales, admin and R&D at UBTech topped ¥2.561bn, eclipsing its revenue — and accounts receivable are swelling. UBTech’s receivables hit ¥1.842bn, exceeding 92% of revenue, with bad‑debt provisions creeping up to ¥539m. Growth is happening, but cashflow quality and unit economics are far from settled.

From “small brain” motion to “big brain” models

A striking theme in these reports is where R&D dollars are headed. Hardware gaps are narrowing; now the race is for embodied intelligence — big models that let robots understand and act in messy, physical worlds. Yushu allocated nearly half of its IPO proceeds to embodied model development; UBTech spent over ¥500m on R&D in 2025 and plans to lift that to ¥700m in 2026, earmarking large shares for “world models” and embodied LLMs. Even profitable firms like Geek+ are setting up new units focused on these systems. Why? Because superior locomotion no longer guarantees differentiation. The harder problem is end‑to‑end semantic understanding, generalization and safe physical interaction — and that needs something mostly absent: high‑quality real‑world interaction data.

Data is the next strategic moat — and a geopolitical one

Who owns the physical interaction logs will likely decide the next phase of the race. Unlike text on the internet, embodied data — force feedback, tactile events, long visual sequences tied to precise motor commands — is expensive and hard to scale. As Yuejiang’s founder warned, industry‑useful data hours are still very limited; firms are answering differently: Yushu “buys” data via research platforms, UBTech trades real factory deployments for training data, and Yuejiang aims for scale through high unit shipments to create a data loop. This is not just a commercial contest; it has geopolitical dimensions. Cross‑border data flows, export controls and supply‑chain restrictions will shape who can assemble the multi‑modal datasets needed to train generalizable embodied models. In short: robots can now move, but can they think — and who will own the data that teaches them? The next two to three years should answer that.