It suddenly struck me that aside from occasionally checking if something can be found, I haven't actively searched through Baidu in a long time. When was the last time you voluntarily opened the Baidu app? Many people probably can't even remember. For answers, people ask DeepSeek; for chatting, they use Doubao; for scrolling content, they go to Douyin or Xiaohongshu—the search box was once the gateway to the internet, but for a growing number of people, it has become just another icon on their phone that they rarely tap.
This isn't just a subjective feeling. It's transforming into cold, hard numbers in earnings reports and an unavoidable crisis for Robin Li over the past two years.
One: The Feeling of 'Nobody Uses It' Is Not an Illusion
Baidu's search market share has already dropped from 86.8% in 2021 to 55.9%. The monthly active users of the Baidu app have fallen from 679 million at the end of 2025 to approximately 640 million as of June this year—for a company that was built on user traffic, its presence in the consumer sector is genuinely weakening. According to QuestMobile statistics, as of June 2025, Baidu's "media influence index" had fallen to ninth place, trailing platforms like Douyin, Taobao, WeChat, Kuaishou, and Xiaohongshu. The way people access information has changed: more and more users are accustomed to finding answers directly through social media or AI chat tools rather than opening a search box first. Baidu itself has sensed this shift. In July this year, Baidu upgraded its search box to a "smart box," with 70% of mobile search result pages now embedding AI-generated content in an attempt to retain users with AI. But this self-rescue effort came a bit late—the moat of the search business has already been breached in more than one place by social platforms and a wave of AI applications.
Two: Earnings Reports Confirm: Actions Speak Louder Than Words
Users vote with their feet, and the market subsequently votes with its money. Baidu's second-quarter earnings report, released on August 18, showed total revenue down 4.2% year-over-year, marking the fifth consecutive quarter of decline; net profit attributable to shareholders was RMB 2.319 billion, a sharp drop of 68.33% year-over-year, with profits halved for three consecutive quarters. On the day the report was released, Baidu's US-listed shares plunged more than 12% during trading. Key metrics for the quarter: total revenue RMB 31.325 billion (-4.2%, declining for 5 straight quarters); net profit attributable to shareholders RMB 2.319 billion (-68.33%, halved for 3 consecutive quarters); search market share 55.9% (was 86.8% in 2021); Baidu app MAU approximately 640 million (was 679 million at the end of 2025).
Baidu has pinned its hopes on its AI business—AI-related revenue accounted for more than half of its general business revenue for the second consecutive quarter. But even so, this RMB 12.5 billion in AI revenue is down from RMB 13.6 billion in the first quarter, and it's not yet enough to fully offset the slowdown in search advertising. The story of digital growth has been told, but the story of profit realization is still unfinished.
Three: Early to Rise, Late to Arrive: Robin Li's Recurring Fate
Looking at Baidu's AI layout over the years, a recurring pattern emerges: almost every time, Baidu charged ahead first; almost every time, someone else ended up with the results. The most typical example is ERNIE Bot. In February 2023, less than three months after ChatGPT went viral, Baidu announced it would become "the first Chinese company to release a ChatGPT-like product." At the time, there were internal disagreements—the technical lead wanted more time to polish the product, but Robin Li insisted on launching first, worried that the crisis in the core search advertising business would shake secondary market sentiment. On March 16, ERNIE Bot was hastily launched, with Li and executives admitting the product was "not yet mature," and Baidu's stock price subsequently fell. The first-mover advantage ultimately failed to translate into long-term leadership.
By 2025, few in the industry building large models considered Baidu a primary benchmark—Tongyi Qianwen's presence was clearly stronger; on Apple's App Store efficiency charts, latecomers like Doubao and Kimi had also surpassed ERNIE Bot to become leading large-model conversational apps. Frankly speaking, in my experience, Baidu's ERNIE is the worst AI, and it even feels less like AI and more like a censorship system—often when you ask a question, ERNIE first tells you that you're wrong. Autonomous driving follows a similar script: Baidu started R&D in autonomous driving in 2013, and its Apollo project started earlier than almost anyone in China. Apollo Go remains one of the largest commercialized robotaxi platforms in the country. AI chips also got an early start—as early as 2018, Baidu was already trial-producing its first-generation cloud AI chip, the "Kunlun 1," which was the only cloud AI chip in China at the time that had undergone large-scale internet core algorithm testing. But whether in global influence in autonomous driving or the commercial scale of AI chips, Baidu has yet to turn "waking up early" into "running far."
Four: The AI Talent He 'Gave Away': Many of Today's US AI Leaders Were Trained by Baidu
The most ironic episode in this cycle took place in Silicon Valley a decade ago. In 2014, Baidu hired "Godfather of Google Brain" Andrew Ng as chief scientist and established the Baidu Silicon Valley AI Lab (SVAIL), which gathered some of the industry's top research talent at the time. Among this group was a relatively unremarkable research scientist focused on refining speech recognition models—he was Dario Amodei, the future founder and CEO of Anthropic. According to Amodei's own recollection, it was during his speech recognition research at Baidu that he informally observed a pattern: the more data and computing power fed to a model, and the more thoroughly it was trained, the better its performance. This empirical rule, later named the "Scaling Law," has now become the foundational logic of large model development. After leaving Baidu, Amodei joined Google Brain, then became an early member of OpenAI, and eventually founded Anthropic in 2021. His former colleagues at Baidu SVAIL have also become prominent names in the industry: Sherjil Ozair became a founding team member of OpenAI; Bryan Catanzaro, who led the Deep Speech 2 project, is now Vice President of Deep Learning Applied Research at NVIDIA; and Jim Fan, an intern he personally recruited into his team, is now head of NVIDIA's embodied intelligence project.
The list could go on: Peng Jun, a founding member of Baidu's US research institute, co-founded autonomous driving company Pony.ai with Lou Tiancheng; Tony Han, former chief scientist at Baidu's autonomous driving division, founded self-driving company WeRide. Even Andrew Ng himself, upon leaving, was generous in his praise for Robin Li, calling him one of the first big-company CEOs to understand the value of deep learning—but Ng still left Baidu in 2017 to found Landing AI and DeepLearning.AI. In other words, a significant portion of the people occupying pivotal positions in today's Silicon Valley AI landscape have "Baidu" written somewhere on their resumes. Baidu once held a winning hand—data, talent, and an early technological edge—but failed to parlay those cards into a truly global AI giant. Instead, it became something like a "Whampoa Military Academy," training and then watching waves of talent depart for others.
Five: Robin Li's Curse of Waking Up Early
Search, content, autonomous driving, AI chips, large models—in almost every track, Baidu was not a latecomer; it was often the first to raise its hand. The problem has never been that Baidu can't see the direction; it's that every time it sees the direction, someone else ends up with the results: ERNIE Bot's first-mover advantage was ceded to Doubao and DeepSeek; the talent cultivated at its Silicon Valley AI lab was ceded to OpenAI and NVIDIA; and the user habits built around the search portal were ceded to Douyin and Xiaohongshu. This might be Robin Li's real crisis: it's not a lack of vision, but rather that every vision ultimately becomes someone else's achievement. Now, Li is pinning his hopes on the IPO of Kunlun Chip—but with so many small and micro chip GPU and CPU players domestically, is this just another story of waking up early?
Source verification: Baidu's Q2 2026 earnings data comes from the company's official announcement (released August 18, 2026); search market share, app MAU, and media influence index data are compiled from media citations of third-party institutions like QuestMobile via sources such as Sohu and 36Kr; the ERNIE Bot launch timeline and internal decision-making details are compiled from media reports including Jiemian News; Dario Amodei's experience at Baidu SVAIL and the "Scaling Law" narrative are compiled from multiple reports citing Amodei's own public statements via sources like Tencent News, Sina Finance, and Zhihu columns, with specific details subject to Amodei's public interviews.
Comments