At Baidu CEO Robin Li’s first internal address of 2026, he reiterated a point he has emphasized over the past year: in the artificial intelligence era, applications matter most. This time, he sharpened the focus around a specific concept, agents.
For the first time, Li clearly categorized Baidu’s four types of agents:
- Search agents, which he described as the company’s primary focus.
- “Digital human” agents, branded under Huiboxing.
- Coding agents, distributed through Miaoda.
- “Self-evolving” AI agents, codenamed Famou.
At the corporate level, Baidu also clarified the relationship between its core search business and Ernie Bot.
Traditional search follows a familiar structure: a user sends a query and receives a list of hyperlinks. In the AI era, however, search functions as an agent powered by a large language model, or LLM, which is a system trained on vast datasets to generate and understand human language.
Baidu is experimenting along two paths. One is Baikan, which integrates a headline image and an LLM-generated answer directly into search results. The other is Ernie Bot, accessed through a central button at the bottom of the Baidu app, where it functions as a chatbot interface.
The broader backdrop is notable. For the first time in nearly a decade, the founders of Baidu, Alibaba, Tencent, and ByteDance have stepped back into prominent roles, directly shaping their companies’ AI strategies. Their renewed involvement indicates that 2026 could prove to be a decisive year in China’s AI competition.
Jack Ma has reemerged at Alibaba, framing AI around what he calls a “golden triangle” of large models, cloud computing, and chips, through Tongyi Lab, Alibaba Cloud, and T-Head Semiconductor. Meanwhile, the Qwen app has replaced Quark as Alibaba’s primary vehicle in the competition for what industry observers describe as a superapp entry point.
At Tencent, Pony Ma has revived the “red envelope” playbook. Yuanbao is offering RMB 1 billion (USD 140 million) in digital red envelopes alongside a new AI-driven social feature called Yuanbao Pai, in an effort to recreate the virality of WeChat’s 2015 Lunar New Year campaign.
Liang Rubo of ByteDance has adopted an equally ambitious tone, focusing resources on expanding Doubao and Dola, its overseas AI assistant.
Building on an existing channel
While competitors are racing to build new channels to capture users, Baidu has chosen a different approach: integrate AI into an existing one.
That strategy reflects lessons from the past two years. In 2024, newcomers such as Kimi gained attention through aggressive marketing and product experimentation. During the 2025 Lunar New Year period, DeepSeek reshaped the competitive landscape. By the end of that year, Doubao had surpassed 100 million daily active users, according to company disclosures. Around the same time, Alibaba said Qwen had exceeded 100 million monthly active users. On January 20, WSJ reported that Ernie Bot’s monthly active users had surpassed 200 million.
The competition has since narrowed to three primary contenders: Doubao, Ernie, and Qwen.
In 2026, competition among major technology companies has translated into what industry observers describe as a RMB 6 billion (USD 840 million) spending race. The logic is familiar in China’s internet sector: prioritize user acquisition and market share before monetization.
For more than two decades, this playbook has defined sectors ranging from group buying to bike sharing and, more recently, quick commerce. In the AI era, that pattern continues, with companies focusing on scale first and profits later.
Ernie Bot appears to be taking a different route. Rather than persuading users to download a standalone app, Baidu aims to remove that step entirely. Users can access AI functions directly within the Baidu app.
With more than 700 million monthly active users, according to company data, the app serves as the anchor. The 2026 Ernie brand upgrade centers on a single objective: users should not have to seek out AI. Instead, AI should surface within search when relevant.
Rather than requiring users to download an app and register before use, Baidu allows them to switch from traditional search to Ernie with a single tap. By reducing friction, it aims to shift user behavior gradually.
This approach resembles the strategy behind Google’s Gemini. While ChatGPT requires users to visit a website or download an app, Google integrated Gemini into Gmail and Android. According to company statements, internal integration significantly expanded Gemini’s Android user base compared with its standalone app. When new capabilities are embedded in high-frequency products, switching costs decline.
Baidu’s advantage is structural. Its 700 million monthly active users and search-driven use cases provide a built-in testing ground for AI agents.
To support this push, Baidu is consolidating resources. Rather than pursuing multiple entry points, it is concentrating on embedding Ernie within the Baidu app and directing traffic internally at minimal marginal cost.
Native AI applications are also emerging. OpenClaw, for example, has attracted attention. Its founder, Peter Steinberger, said that 80% of apps on smartphones are “dead,” though users may not yet recognize it. The claim reflects a broader debate about whether AI-native tools will displace traditional mobile apps.
Li appears mindful of these risks. On January 1, Baidu subsidiary Kunlunxin filed for a main board IPO on the Hong Kong Stock Exchange. If approved, it could become one of the first major Chinese technology firms to spin off its chip unit.
Li has previously said internally that the purpose of training models is inference, meaning the process of generating outputs after a model has been trained. While some companies focus on scaling computing power for training, Baidu is emphasizing inference efficiency.
According to estimates from Longbridge Securities, based on an average price of RMB 60,000 (USD 8,400) per Kunlunxin P800 chip, shipments in 2025 reached about 60,000 units. Orders from major clients, including Baidu, exceeded 200,000 units, leaving 140,000 pending delivery because of Samsung’s capacity constraints. If production increases and those orders are fulfilled, the backlog alone could generate RMB 8.4 billion (USD 1.2 billion) in revenue, representing 140% year-on-year growth. These figures are based on brokerage estimates and have not been independently verified.
On the cloud side, Baidu disclosed AI revenue for the first time in its third-quarter earnings: RMB 4.2 billion (USD 588 million), up 33% year on year. Subscription revenue from AI high-performance computing infrastructure rose 128%. Recently, Baidu AI Cloud raised its 2026 AI-related revenue growth target from 100% to 200%, according to company statements.
At the model layer, Baidu reorganized internally. The basic model unit focuses on achieving global state-of-the-art benchmarks. The applied model unit prioritizes smaller models, faster responses, and cost efficiency.
That structure was reflected in the January 22 release of Ernie 5.0. With 2.4 trillion parameters and native multimodal modeling, meaning it can process and generate text, images, audio, and video within a unified system, the model supports multiple input and output formats. Baidu said that across more than 40 benchmarks, its language and multimodal capabilities surpassed Gemini 2.5 Pro and GPT-5 High.
As one of the few Chinese companies with a technology stack spanning chips, cloud services, AI models, and agents, Baidu argues that in-house development improves cost control and supply chain resilience. By aligning chip-level instruction sets with Ernie’s algorithms and coordinating scheduling systems at the cloud level through Famou, it integrates hardware, cloud computing, and inference.
In a geopolitically fragmented environment where computing resources are constrained, that level of control offers operational security. Kunlunxin supports chip supply. Baidu AI Cloud enables low-latency distribution. Ernie 5.0 provides the model capabilities that allow Ernie Bot to serve Baidu’s extensive user base.
Will this approach help Baidu come out on top in the AI race? That remains an open question. For now, however, its full-stack strategy gives it more cards to play and, potentially, the power to define the rules of the game.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Xiao Xi for 36Kr.