Baidu Intelligent Cloud Splits Platform Product Division, Agent Business Becomes Independent Unit
nashnova research
Baidu AI Cloud broke its Platform Products Division into three standalone units in August, with the headline move being a dedicated Agent division — this means Baidu has formally elevated its general-purpose agent business from a side function to a strategic bet on par with infrastructure.
One division into three — how does it split?
The former Platform Products Division becomes three peer units: Infrastructure (Hou Zhenyu), Agent (Yin Dawei), and Intelligent Applications (Ruan Yu). All three report directly to Shen Dou, Baidu's EVP and head of the AI Cloud group.
MaaS — model-as-a-service, packaging large models into callable cloud APIs — was folded into the Infrastructure Division, sitting alongside compute and chip resources.
This means → customers who previously had to coordinate with separate compute and model teams can now get end-to-end delivery from chip scheduling to model inference in a single division.
Why push MaaS into infrastructure?
A person close to the restructuring put it bluntly: "MaaS is essentially infra" — model calls ultimately compete on resource utilization, inference cost, and deployment scale, all infrastructure metrics.
After the merger, the pipeline from Kunlun chips (Baidu's in-house AI silicon) to Qianfan (Baidu's model-serving platform) is unified. The goal: cut cost, raise efficiency.
In plain terms = model serving is not a standalone business; it is the packaging layer on top of the compute business. Managing them separately only created friction.
The Agent division is independent — but where is the revenue?
The Agent Division focuses on general-purpose agent products — Baidu Dazi, Miaoda, and Famou — with core KPIs of DAU, MAU, retention, paying customers, and end-to-end task-completion rate.
Baidu Dazi cited AI product-chart data showing roughly 6.74 million PC MAU in July, up 1,063.79% month-on-month — but insiders say the team cares more about whether agents can genuinely enter professionals' and enterprises' production workflows.
This means → user volume is the shop window; what management actually tracks is DAA — daily active agents, measuring how many agents are truly used each day. Traffic looking good is not the same as monetization working.
What do the financials reveal?
Baidu's Q2 AI Cloud Infra revenue hit RMB 7.3 billion, up 50% year-on-year; GPU Cloud surged 283%. AI Applications revenue was RMB 2.5 billion, up only 3%.
In plain terms = infrastructure is earning real money; agents and vertical apps are still burning cash to find a path. The underlying logic of this split is scale what makes money, let what burns money iterate light and fast.
The Data Platform unit did not follow MaaS into Infrastructure. It was renamed Data Intelligence and moved into the Intelligent Applications Division — placing data teams closer to industry clients in verticals like transport and finance.
Big-tech agent wars — is Baidu moving fast enough?
Since July, Alibaba consolidated QoderWork, Wukong, and MuleRun into Qianwen Office; Tencent merged QClaw operations into WorkBuddy; ByteDance further integrated its Feishu and Doubao product teams.
Baidu has raised the organizational priority of general-purpose agents, but has not yet consolidated multiple office-agent entry points across the group — a step behind rivals in integration.
This reflects a broader shift: big-tech agent competition has moved from "who ships a product first" to "who sorts out the org chart first." After the org lands, the real test is whether any of them can deliver answers on end-to-end task-completion rate and paid conversion.
市场有风险，内容仅供研究参考，不构成投资建议。