At WAIC 2026, the scene was unmistakable: Agent products were everywhere. From the subway ads for Baidu's 'Baidu Dazi' to the packed halls, the battle for the next platform entry point had begun. The big model had stepped off center stage, replaced by the Agent as the new protagonist. IDC's first DAA research report, released during the conference, projected global active agents to surge from 28.6 million in 2025 to 79.4 million in 2026—nearly a threefold increase in just one year. Yet, as the author notes, this is a supply-side boom, not a demand-side one. User growth is slow, willingness to pay is weak, and business models remain fragile.
Why? One practitioner put it bluntly: 'AI is more like manufacturing than the internet. It doesn't get cheaper with more users; every call burns compute, and the more complex the task, the higher the cost.' Kai-Fu Lee, founder of 01.AI, echoed this in his WAIC keynote: the industry has shifted from competing on model strength to competing on application value and real-world impact. He warned that the general-purpose model race is a winner-take-most game, and companies without big-tech backing should avoid that homogeneous path. When making an Agent becomes easy, the scarce resources are user entry points, orders, and delivery.
Big tech's anxiety is palpable. Baidu, lacking a super app, sees search as its most valuable asset—and search is exactly what Agents threaten. So Baidu is betting on a universal Agent entry point, pushing 'Baidu Dazi' as a front-end that captures user intent before they decide which tool to use. Tencent, having lagged in the model race, is densely embedding Agents into its WeChat, documents, and office ecosystem, showcasing WorkBuddy, CodeBuddy, and Marvis. Alibaba leverages its cloud advantage, enabling users to build their own tools on the cloud, with products like Qoder and Miaowu, and launching an Agent-native cloud. StepFun, a newcomer without legacy constraints, is embedding Agents directly into phones and operating systems with STEPX Neo and Amoo.
Why the rush? Everyone sees Agent as the next platform, like cloud ten years ago. But no one knows the winning form—a single killer app, multi-agent orchestration, or an Agent OS—so the safest strategy is to hedge across all layers. This explains the dense product matrices on display.
Under the shadow of big tech, startups are even more crowded. H4 hall, which selected only 158 out of 1,200 applicants, was packed with vertical Agents: legal, research, recruitment, document writing, marketing, video creation, even fortune-telling. One exhibitor noted that legal is one of the few AI scenarios with real revenue, attracting many teams. Capital has grown cautious about general Agents, favoring vertical products with clear customers and payment scenarios. A participant observed that the window for small companies to set their own rules is closing; they must enter specific workflows or provide indispensable underlying capabilities.
Startups are taking three paths: direct vertical solutions (e.g., Midu's document tools, Codar for research), development platforms that lower the barrier to building Agents (e.g., AutoAgents.AI, AgentMa), and infrastructure providers (e.g., SoMark, PPIO) offering document parsing, sandboxing, and hosting. The last may hold long-term value, but the first is the most common.
Despite the crowded booths, demand remains cold. A founder of a vertical Agent product admitted that their commercial path focuses on B2B, with C-end products for branding only. Their membership is priced at 99 yuan per month, but 80% of revenue goes to token and document service costs, leaving almost no profit. This means even paying users are essentially working for the model vendors. B2B isn't easier: each enterprise requires custom data integration, system access, and workflow adaptation, turning products into bespoke services.
Kai-Fu Lee summarized the profitability dilemma: model vendors outside the top tier struggle to cover training and inference costs; To C companies face high customer acquisition costs; To B companies fall into project-based traps. When asked, most exhibitors could detail what their products do but avoided answering three key numbers: stable paying customers, repeat usage of the same task, and profit per task after model calls, human review, and delivery costs. These numbers matter more than demos.
Product homogeneity amplifies the problem. Many vertical Agents are essentially the same: a general model, a vector database, RAG, and a preset workflow wrapped in an industry UI. One exhibitor said, 'These endless Agent products are just reinventing the wheel.' Vertical alone is no longer a moat.
Yet demand isn't entirely absent. 01.AI showcased 'Boss AI,' 'Top Sales AI,' and 'Investment Officer AI,' targeting decision-making scenarios where enterprises are willing to pay, benchmarking Palantir. According to Lee, internal use has increased order value by 5x, with a goal to IPO by 2027 as China's first profitable AI company.
For founders and operators, the takeaway is clear: AI's unit economics demand a manufacturing mindset. Focus on repeatable tasks, reduce delivery costs, and build a moat beyond the model. The winners will be those who control the user entry point and deliver real, repeatable value—not just another wrapper.
Source: 雷峰网.