Key points
- The funding: On August 28, 2026, Sharpa—a company founded in late 2024 by Hesai Technology's three co-founders (CEO Li Yifan, CTO Xiang Shaoqing, Chief Scientist Sun Kai)—disclosed a financing round of over 4.5 billion RMB at a 22 billion RMB post-money valuation. Investors include Alibaba, Meituan, Tencent, JD.com, Transsion, Sequoia China, Qiming Venture Partners, Meituan Longzhu, and Guanghe Ventures.
- The milestone: On August 29, a Sharpa humanoid robot went live at a DQ (Dairy Queen) store on Shanghai's Wujiang Road, independently completing all 55 steps of making a Blizzard ice cream—taking orders, operating the ice cream machine, opening lids, adding toppings, stirring, and delivery—using the store's original equipment with zero modifications. This is a rare "zero-retrofit, fully autonomous" deployment in Chinese chain food service.
- Sharpa Wave: a 1:1 human-hand-size dexterous hand with 22 active degrees of freedom and 1,000+ tactile units per fingertip.
- Sharpa North: a 70-DOF full-body humanoid platform.
- CraftNet: a 0/1/2 hierarchical omni-modal manipulation foundation model unifying reasoning, vision, and touch.
- "Effective data" = edge cases a robot encounters in real deployment, solves, and feeds back—not the scale of teleoperated collection.
- Cost logic: if the scenario is chosen well, data comes from building a useful product (Tesla-style), not from spending money on collection.
- The DQ scenario took under five months of integrated training; the robot currently runs at half human speed. Li's benchmark: "The first scenario can cost 100,000 hours—as long as the second costs 100."
- Speed gaps close fast (at the 2nd World Humanoid Robot Games, the 400m record dropped from 1:28.03 to 45.66s; standing high jump rose from 0.96m to 3.40m), but manipulation advances far slower because running has physical optima and mocap data; "scooping toppings with that store's spoon" has neither.
- Unitree listed on the STAR Market in August 2026 (IPO price 150.80 RMB, ~60.99B RMB market cap), peaked above 444.9B RMB, then fell 44% to 248.8B RMB by Aug 27 close. Per its prospectus, 73.6% of humanoid revenue in the first nine months of 2025 came from research/education; industrial applications were only ~9%.
- Counterpoint Research: global humanoid shipments exceeded 22,000 units in H1 2026 (+~300% YoY), with full-year forecasts above 50,000. Zhiyuan led with 43% (~9,700), Unitree 31% (~7,000); the top five are all Chinese firms holding 86% of global share. Yet entertainment/performance plus research/data-collection still exceed 60% of deployments; smart manufacturing is only 13%, warehousing 5%.
- Unitree founder Wang Xingxing at WRC 2026: mass adoption awaits the "two 80%s"—completing ~80% of tasks in 80% of novel scenes; "as fast as 2–3 years, as slow as 5–10." Galbot's Wang He: current embodied foundation models are at a "GPT-2 stage," reaching GPT-3.5–GPT-4 level around 2028.
Li Yifan's three rules for "real-world usability"
1. Never modify the environment. "When the staff leaves, the robot walks in and works." No mounts, handles, or machines were changed in the DQ store. 2. Finish the entire task. If a robot does 9 of 10 steps and a human does 1, the human must stand by permanently—so the replacement rate is zero. True replacement is determined by the last remaining human-touchpoint (the bottleneck), not the percentage of automated steps. 3. Choose high-value, complex tasks. 55 sequential steps test tactile perception, computation, real-time error correction, and dexterity simultaneously.
He also mocks "manufacturing jobs for robots"—e.g., a coffee robot whose only job is handing over coffee made by an automatic machine placed far away on purpose.
Why dexterous hands
Li's argument: "Every machine can use cameras, but what about problems cameras can't solve? Watching 3 hours of soccer daily gives you a strong visual world model, but you still can't play—because you need to feel the ball under your foot." Contact signals (touch, force, proprioception) never appear in internet videos—you can find a million egg-cracking videos but no record of the fingertip pressure curve the moment the shell breaks.
Data flywheel and generalization
Industry context: the valuation anchor is shifting
"Physical AI has no OpenAI opportunity—only an Apple opportunity"
Li argues Digital AI could have a model-layer-only winner because its infrastructure layer is mature. Physical AI's bottom layer—real-world data, compute, dexterous hands—is severely incomplete, so a full-stack player is required. His endgame view: market size 5–10x the auto industry, but consolidation to no more than 5 general robotics players, with layered moats (users → data → generalization → efficiency → ecosystem).
Restraint
Li refuses orders when feasibility scores under 90/100: "A customer's willingness to pay may reflect fantasy about the product, not a solved problem." His aphorism: "AI problems are climbing a mountain; engineering problems are stopping halfway and walking sideways." Home deployment is targeted from 2028: "If you can't finish an ice cream commercial scenario, how can you enter homes?"
Three things to watch
1. Actual uptime of the DQ store post-Aug 29—the ratio of autonomous orders to human interventions. 2. Training hours for the second scenario—a three-orders-of-magnitude drop (100,000 → 100 hours) would validate the technical route. 3. Hesai's actuator module shipments: over 10,000 cumulative units by end of Q2 2026, ramping to ~10,000/month, targeting six-digit annual output by 2027; full-joint modules SOP in H2 2026.
Li's closing metaphor on the hype cycle: with so few real deployments, "everyone thinks they're in the first tier—it's like arguing about who runs fastest inside a bathhouse." The counter to narrative inflation: 55 steps, zero retrofit, full autonomy—a yardstick that is hard to fake.
References
1. LatePost exclusive interview, 2026-08-28 2. Lieyunwang, 2026-08-28, https://www.163.com/dy/article/L5EIAG9O05118HJE.html 3. Robot Insight, 2026-08-28, https://www.sohu.com/a/1068882430_122054251 4. AInvest, 2026-08-28, https://www.ainvest.com/news/sharpa-robotics-secures-4-5b-funding-22b-valuation-physical-ai-2608 5. STAR Market Daily / Counterpoint Research, 2026-08-27, https://www.chinastarmarket.cn/detail/2466860