Robin Li stood before an audience at Baidu Create 2026 in Beijing on May 13 and declared a transition. The competition among AI companies, he argued, had shifted from building more capable models to building systems that can actually get things done. “What users are willing to pay for,” Li said, according to TechNode, “is no longer whether AI can think or not, but whether it can get things done.”
The conference, themed Agents at Scale, was Baidu’s clearest signal yet that China’s AI industry is moving from the model-capability phase, characterized by benchmark racing and large-model releases, into a deployment phase built around autonomous agents that persist, reason, and execute across enterprise workflows. That shift has direct implications for infrastructure demand.
The Agentic Inference Difference
A conventional AI chatbot session consumes a bounded number of tokens: a user sends a prompt, the model responds, the session ends. An agent operates differently. It breaks tasks into subtasks, calls external tools, manages state across multiple steps, and may remain active across extended time horizons. As OpenRouter’s December 2025 State of AI report noted in its analysis of more than 100 trillion real-world inference interactions, agentic usage patterns generate sustained compute demand across longer sessions. Where one user query might previously have triggered a single model call, an agent handling the same task may trigger dozens.
Baidu launched three agent products at Create 2026 that illustrate what this looks like in practice. DuMate is a general-purpose agent designed for customer service resolution, data analysis, and content generation. Miaoda is a code-generation agent that, according to TechNode, generates approximately 90 percent of its own code. Baidu YiJing is positioned as an agentic platform for enterprise workflows. The company also described an agent self-evolution loop incorporating verification, error correction, and iterative optimization.
A New Metric for the AI Era
Li proposed what amounts to a redenomination of how the AI era will be measured. Token consumption, meaning the number of words processed per day, reflects language model deployment but does not capture agentic complexity or economic value delivered. Li argued that daily active agents could become the defining metric, and reportedly predicted that the global count could eventually exceed 10 billion.
This framing matters strategically. If China’s AI industry succeeds in shifting the conversation from model benchmarks to deployment reach, the competitive scoreboard changes. Agents embedded in enterprise workflows, manufacturing operations, and customer-facing systems create stickier relationships and more recurring infrastructure demand than consumer chatbot usage. Baidu AI Cloud, which according to AASTOCKS Financial News has served more than 1,000 AI hardware companies, is positioned as the infrastructure layer for this shift.
The Infrastructure Implication
If daily active agents scales as Li projected, the compute requirements are qualitatively different from chatbot usage. A chatbot session may generate hundreds of tokens per exchange. An enterprise agent resolving a customer service case, analyzing financial documents, and generating a compliance report across a multi-hour workflow may generate tens of thousands of tokens while calling external tools, writing and executing code, and querying databases. That multiplier effect, distributed across millions of concurrent enterprise agents, represents a step-change in inference infrastructure demand.
China’s AI cloud providers, including Baidu AI Cloud, Alibaba Cloud, and Tencent Cloud, are all positioning agentic infrastructure as an enterprise service layer. The competition is not only between AI models but between orchestration platforms, data integration capabilities, and the enterprise relationships that determine which cloud gets the agent workloads. Baidu’s reported service to more than 1,000 AI hardware companies gives it a node in the device and robotics ecosystem that could differentiate its agent infrastructure from pure-play cloud competitors.
Baidu’s agent-era push represents a response to competitive pressure from DeepSeek and other Chinese labs that have eroded ERNIE’s model-capability differentiation. By reframing the contest around execution rather than reasoning scores, Li is betting that enterprise integration and agentic workflow depth will prove more durable moats than benchmark performance. Whether the infrastructure demand this creates accrues primarily to Baidu’s cloud or disperses across the Chinese AI stack remains an open question. What is clear is that China’s agentic AI policy framework is already taking shape around it.
The deeper question is whether Baidu’s agent strategy succeeds in converting ERNIE’s broad enterprise touchpoints into a defensible position before competitors establish competing agent defaults. Robin Li’s prediction of 10 billion daily active agents globally is a vision, not a roadmap. But it names the contest he believes the next phase of Chinese AI competition will be decided by, and it stakes Baidu’s future on being well-positioned when it arrives.