At WAIC 2026, the AI race shifted from benchmarks to delivery

WAIC 2026 ran July 17–20 across three Shanghai venues (the Expo, Zhangjiang, and West Bund) under the official theme "Intelligent Partners, Co-creating the Future." The unofficial subtext across booths and panels was the same question: when does AI start doing the work, not just passing the test? (36氪, a Chinese tech outlet's on-floor recap).

The shift showed up in three concrete places.

A headhunter Agent built for a workflow, not a benchmark. An Agent is a workflow-integrated AI assistant that takes actions inside a job, not just answers questions. 禾蛙 used the 氪话未来 livestream inside 36Kr's WAIC content program to debut Domi, an Agent for the recruiting industry, built on its in-house HWE headhunter model and a matching engine, with an agent-to-agent (A2A) trading network and a developer skill ecosystem on top. The company also reported platform scale: more than 54,600 partner firms, 190,800 HR consultants on the platform, 600,000 candidate recommendations a year, and 12,700 annual offers (36氪, Sina Finance, 清科, Sohu). Those are company-reported figures repeated across multiple syndications, not independently audited metrics. The CEO framed the bet directly in the livestream: "我们不是用AI替代猎头，而是让从业者骑着AI跑赢AI" ("We are not using AI to replace headhunters, we are letting practitioners ride AI to outrun AI") (36氪).

An embodied AI flywheel already running in pharmacies and hotel laundries. Embodied AI means artificial intelligence that controls physical robots, not just text or images. 穹彻智能 walked the floor talking about a flywheel the company calls "real-world data collection, embodied foundation model, deployed commercial scenarios," and showed RoboPocket 2.0 doing the same job in two different settings: dispensing in a pharmacy, and folding in a hotel back-of-house (36氪). The AI scientist 吕博, on the show floor, made the company's moat argument explicit: a single-robot demo is not the moat, a continuously evolving data-and-model flywheel is. The point is not that pharmacy robots are a market; the point is that the company is selling repeatable service hours, the way a logistics software company sells shipments.

A modular humanoid platform that treats the body as a system. 华沿机器人 unveiled what it called a "modular embodied nervous system," with standardized modules for data capture, teleoperation, perception, upper limbs, mobility, and tasks, alongside its HY/Echo seven-axis humanoid arms built on impedance control, spatial following, and VR teleoperation (36氪). The pitch was that a robot arm should be assembled from interchangeable parts, the way a server is built from a standard rack, and that the real product is the integration layer, not the metal.

The 36Kr piece frames the show's larger narrative turn explicitly: AI moving from model racing and leaderboard chasing toward real-scene delivery, system delivery, and measurable business value. The same "delivery contest" reading has been the dominant WAIC 2026 recap across the Chinese AI press. The question worth asking is whether the rhetoric matches the receipts. The conference's loudest reframe, "delivery over benchmarks," needs to be measured against next quarter's AI revenue reports, not against the next product launch. If the deployed workflow numbers move, the reframe sticks. If vendors go back to model-launch theater by Q4, the conference was mood, not market.

The watch item from the show floor is whether the same companies that spent 2024 and 2025 chasing parameter counts and benchmark wins actually book revenue from the workflows they demoed this week. Domi's headhunter Agent, 穹彻智能's pharmacy and hotel service hours, and 华沿机器人's modular arm platform are three test cases worth tracking over the next two quarters. The conference is over; the delivery contest has just started.