AI's Economic Shift: From Efficiency to Incremental Value

Introduction: The Shift from Demo to Deployment

The 2026 Inclusion·Bund Conference in Shanghai gathered 300+ exhibitors across 15,000 sqm, including 40+ embodied intelligence companies. Unlike previous years where robots performed backflips, danced, or played piano, this year's exhibits focused on practical deployment: a robot navigating an 80-cm pharmacy aisle to pick medicine, a wheeled dual-arm robot inserting a screw into a 3-mm hole on a washing machine backplate, and a quadruped robot autonomously patrolling GPS-denied, dark, electromagnetically noisy underground tunnels.

The contrast signals a collective industry mindset shift: "Demo does not buy a second-year budget unless it enters real workflows." The evaluation criteria moved from motion performance (backflips, gait human-likeness) to workflow embeddability (zero store retrofit, overnight operation, peak-load resilience) and supply-chain completeness (brain → body → payment → scenario).

Scene 1: The Night-Shift Pharmacist in an 80-cm Aisle

Ant Group's Lingbo brain and Shanghai Guoda Pharmacy co-deployed a robot in actual stores. The robot receives an order, navigates to the target shelf, picks the correct box from 3,000+ visually similar drugs in dense packing, and delivers it to the pickup window in ~1 minute. Key constraints: 80-cm aisle width, transparent packaging glare, random placement.

Two design details stand out: (1) the robot's waist bends 90° to reach bottom shelves and can switch hands after grasping; (2) zero store retrofit required —no shelf relocation, QR codes, or re-planning. This eliminates the integration cost that typically extends ROI from 18 months to 5 years.

Counter-intuitively, the robot first replaces night shifts , not day shifts. Night shifts are expensive, hard to staff, and error-prone at 3 AM. The robot is slower than a skilled pharmacist (30 sec vs 60 sec per order) but "its selling point was never speed—it was 'someone is there'." This exemplifies "incremental" value: not accelerating existing flow but filling a previously unfillable time slot.

The same Lingbo brain (LingBot-VLA 2.0) also drives logistics parcel sorting (partner Leju) and industrial part loading/unloading (partner Tita). The model pre-trained on 17 mainstream robot brands and 20 configurations (single-arm, dual-arm, biped, wheeled). This signals VLA models evolving from single-robot firmware into cross-body operating systems , shifting business model from one-time hardware sales to recurring brain licensing.

Scene 2: Two Arms at a 3-mm Hole

Fudan-spawned Miaoshen Intelligence's wheeled dual-arm robot identifies bolt structures and assembly holes on washing machines, then autonomously grasps, aligns, inserts, and tightens screws. Chief Scientist Prof. Chen Tao calls screw assembly "the touchstone for industrializing dexterous manipulation" because it is the most fundamental, core industrial process and tests force control and spatial precision.

Chen outlines four generations of embodied intelligence: (1) pre-programmed kinematics/control; (2) VLA + imitation learning from human demos; (3) World Model + VLA; (4) World Motion Model (WMM) —injecting world knowledge/mechanics into the robot brain so it works with understanding, not just data fitting. In May 2026, the team released the first spatiotemporal unified WMM.

Two industrialization-critical achievements: (1) Data cost : EGM framework learns from internet video at 1/12 the usual data cost, achieving SOTA humanoid motion tracking. (2) Edge deployment : 60× compute acceleration shrinks the brain to run on onboard chips, eliminating cloud round-trips.

Miaoshen already works with Midea (washing machines) and Bright Group, achieving 3-mm precision rear-cover insertion and screwing. H1 orders near ¥100M. The lesson: factory adoption hinges on driving both hardware and data costs into customer-acceptable ranges —1/12 data cost and 60× edge speed matter more than any demo video.

Also in this zone, Qinglang's commercial service robot (KOM 3.0 brain) "thinks before acting"—predicting physical outcomes to avoid failure, turning "trial-and-error execution" into "predict-decide-execute". Targets flexible object manipulation and precise grasping in unstructured environments. Deployed at Shangri-La hotels for reception, delivery, cleaning, laundry.

Scene 3: Where GPS, Light, and Comms Fail

Chenhang Yuelue's flying agents and quadrupeds enter nuclear plants, subway tunnels, water plants— no GPS, no light, strong EMI —for autonomous cruise, obstacle avoidance, defect detection. Erxing's quadrupeds target substations and factories for long-distance, high-risk inspection.

These solve "humans can't do it" vs. the previous two solving "humans won't do it". Nuclear containment, deep tunnels, urban pipe networks share: navigation failure, extreme personnel risk, huge accident cost. Traditional quarterly manual shutdown inspections become on-demand; the increment is enabling a service that previously did not exist .

Similar: Fudan's multi-gait peristaltic pipeline robots (origami peristaltic, snake-peristaltic) for corroded/cracked/blocked gas pipes; ShiHang's "Orca" ocean robot (0–10,000 m depth) with proprietary ocean embodied brain, autonomous in complex currents, dark, no comms. Served 1,000+ vessels at 20+ core ports globally; entered CNNC Tianwan demo for underwater autonomous charging and networked recovery.

Economic Perspective: Why "Efficiency" Is No Longer Enough

Peking Univ. Prof. Zhao Bo decomposes AI's growth mechanism into three: lowering skill acquisition cost, expanding market participants, lowering innovation trial cost. None points to "speeding up existing processes" but to "letting previously excluded people and businesses enter".

His widely cited judgment: AI's true economic value is not making existing people do existing things faster, but making previously too-expensive, too-hard, non-scalable things doable, purchasable, tradable for the first time.

Three progressive layers:

Doable —technical feasibility (night shift, deep-sea desilting, pipe inspection).

Purchasable —cost lowered to willingness-to-pay (robot night shift becomes a viable business).

Tradable —standardized billing turns custom projects into a market.

Other forum views:

Shanghai Univ. of Finance President Liu Yuanchun: huge divergence in AI productivity estimates (Acemoglu 0.07–0.1 pp, OECD 0.25–0.6 pp, Aghion 0.68 pp task automation + 0.4 pp idea generation = 1.08 pp). "AI economy shape set, but AI economics not started."

Peking Univ. Dean Huang Yiping: two historical parallels—Solow Paradox (tech diffusion precedes productivity data) and Engels' Pause (output rises, wages stagnate). Flags "strong supply, weak demand" structural risk.

Nobel laureate Sargent: current AI at "Kepler stage"—finds data patterns but not structural causes; massive investment ≠ high returns.

Ant Research Dean Li Zhenhua: upstream compute ROI clear; mid-stream models and downstream apps under commercialization pressure.

Ant CEO Han Xinyi: enterprises must move from "cost reduction" to "value creation"—new supply, new demand.

Together they frame "incremental" as a harder problem: efficiency optimizes within an existing production function; incremental requires building a new one. The former fits an ROI spreadsheet; the latter demands answering "why didn't this business exist before?"

Industry Judgments: Three Structural Shifts

Selling brains scales better than selling bodies. Bodies are one-time revenue, heading for commoditized price wars. Brains naturally reuse across bodies; once adapted configurations and workflows create network effects, model shifts from hardware sales to licensing. Lingbo's 17 brands/20 configs and Miaoshen's "brain must not be body's slave" both point to whoever owns the brain layer defining interface standards.

Standardized delivery determines whether one store success replicates to 1,000. Zero retrofit in 80-cm aisles matters more than demo success rate. It means controllable marginal deployment cost—otherwise every new store is a re-survey, re-debug, re-acceptance cycle, and the scale math never closes.

Agent commercialization needs trust infrastructure. As agents autonomously execute commerce, "who is transacting" blurs. Ant launched APASS for agent commercial trust: identity (who), representation (for whom), permission (allowed actions), behavior trust (auditable). Gartner predicts 90% of B2B procurement mediated by AI agents by 2028, driving $15T+ through agent platforms—without identity/authorization foundations, those numbers never materialize. Embodied robots as physical agent endpoints face the same gate.

Conclusion: The Robots That Don't Dance

The answer to "bubble or new economy?" sits in the exhibition hall. The robots didn't change the world; they simply started appearing where no one was before—3 AM pharmacy, 80-cm aisle, nuclear containment, 40-m seabed.

Past two years' keyword: "efficiency", metric: headcount saved, cost cut. Next keyword: "incremental", metric: a simpler yet harder question— "Did this thing exist before?"