A robot-vision founder says peers faked revenue — one is named
China's embodied-AI money trail just got audited by a competitor, and the auditor is a supplier that went public twelve days ago.
Shao Tianlan, founder of robot-vision component maker Mech-Mind, posted on WeChat Moments that several Chinese embodied-AI companies — including, in his comment replies, Galaxy General Robot — manufacture fake, unsustainable revenue through so-called "data collection centers" and related-party deals with local governments, investors and suppliers. As Red Star Capital Bureau reported it, Shao called the practice illegal, unethical, and "not smart," and said the trap is self-reinforcing: a company that books this kind of revenue one year has to keep producing more of it the next. Galaxy General Robot answered on September 10 with an official statement titled "Embodied intelligence is a long race; we only race against time," declining a shouting match and citing two years of routine commercial operation across industrial, retail and pharmacy settings. The timing gives the spat teeth: Mech-Mind listed in Hong Kong on September 1 at HK$101.7, was roughly 3,835 times oversubscribed in the public tranche, then slid more than 20% below its issue price within four sessions. Meanwhile a 36Kr survey of the field logged more than ¥46 billion of disclosed Chinese embodied-AI funding in the first half of 2026, about seventy percent of it to the top twenty companies, with at least four — Galaxy General Robot among them — valued above ¥20 billion. Regulators can smell the same air: the NDRC warned in late August against "blind following and a rush in" in robotics. Our read — this is what a public market does to a private narrative. The valuations need revenue, the revenue needs customers, and the customers turn out to be the investors' other portfolio companies. A rival supplier with an IPO price to defend is the one actor with an incentive to check the math. We argued last month that the sector's problem is capability, not capital — The Take — Unitree's rout isn't a bubble. It's the brain lagging the body.
Past 7,000 thinking tokens, reasoning models start un-solving their own answers — and the bill shows up before the insight does. A Nanjing University study, "When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling" (Findings of ACL 2026), tracked flip events across compute budgets: the ratio of negative flips (correct → incorrect) to positive corrections crosses 1.0 at about 7,000 tokens, hits 1.42 at 8,000, and degrades to 7.55 by 16,000 — with generating 8,000 tokens costing sixteen times a 500-token answer. Of the negative flips, 67.5% are genuine second-guessing, and among 312 naturally long runs, the 71% containing explicit reconsiderations ("wait", "actually", "let me reconsider") scored 12% worse than those without. Jiqizhixin is now framing overthinking as the defining deployment problem of 2026, and the pricing angle has its own literature: a Microsoft Research study found a lower-listed model costs more in practice in 21.8% of model-pair comparisons, with reversals up to 28x. The take — thinking length has quietly become a cost line, not a quality line. The next benchmark worth watching isn't accuracy at maximum budget; it's accuracy per token, and "knowing when to stop" is the mechanism nobody has productized well yet.
What to watch: whether Galaxy General Robot publishes auditable revenue detail — and whether any lab ships a stopping criterion, not just a thinking budget slider.
Is your team paying for thinking tokens that flip right answers into wrong ones — and would you trust a model to decide when to stop? Tell us in the comments.
