Shifting to the B-end and appointing Sun Tianxiang, how many hidden cards does Baidu AI have left?
With the 2026 calendar year already past its midpoint, the spotlight in the large language model (LLM) track has long shifted to new leading players.
When industry discussions center on Tencent's Workbuddy, Alibaba's Qwen, ByteDance's Doubao, DeepSeek V4, and Kimi K3, few people still bring up Baidu's Ernie Bot in side-by-side comparisons anymore.
That "pioneer" which debuted in March 2023 bearing the title of the world's first large tech company's general-purpose large language model has now been left behind by both major tech giants and startups alike.
From the perspective of AI applications, as of June 2026, Doubao firmly holds the top position with 382 million monthly active users, followed by Qwen with 167 million and DeepSeek with 130 million. Meanwhile, the once "top contender" Ernie has seen its standalone app's monthly active users steadily decline, now dropping to less than 5 million — a 7-fold gap separating it from the first tier of products.
In this track where product iterations are measured in days, "falling a single step behind" means being forgotten by users. A former Baidu employee claims the iteration pace of the Ernie App can no longer keep up with the industry rhythm.
The decline that started and faded from public user attention dragged on through June and July this year, until Baidu began to regain public notice.
Recently, Baidu consolidated Ernie's multiple entry points, and leading LLM technical expert Sun Tianxiang officially joined the company. Shen Dou, President of Baidu Intelligent Cloud Business Group, has repeatedly emphasized the agent strategy on various internal and external occasions. At the World Artificial Intelligence Conference, the agent product "Baidu Buddy" became the core offering. Shortly afterward, Baidu disclosed its capital plan to pursue a dual primary listing, attempting to reshape its valuation...
From patching technical gaps and adjusting product positioning to building momentum in capital markets, Baidu has fully adopted an all-in stance this time. However, compared to the widespread acclaim that greeted Ernie Bot's debut in 2023, this all-in push appears somewhat strained.
The LLM industry has evolved from "competing on parameter counts" to "competing on real-world implementation". The "new BAT" (ByteDance, Alibaba, Tencent) landscape has taken initial shape, the "four dragon" companies (Zhipu AI, Minimax, Moonshot AI, Stepfun) each hold their own ground, and DeepSeek stands out uniquely. Baidu's market capitalization has even been surpassed by Zhipu AI.
The game has reached its midgame. Can the pioneer that dealt the first cards three years ago still reclaim its rightful ticket to the game? This is a question that even Baidu itself has no clear answer to at the moment.
Two rounds of adjustments in half a year, one short-lived turnaround: Baidu's AI rides a rollercoaster
In November 2025, Baidu carried out an internal organizational restructuring focused on its AI business. This time, the internal realignment split the TPG (Technology Platform Group) into the Basic Model Research Department (BMU) and the Applied Model Research Department (AMU), initially overseen by Wu Tian, Baidu Group Vice President, and Jia Lei, Baidu Speech Chief Architect, respectively.
After this structure operated for six months, Baidu made further adjustments in May 2026: establishing the Model Committee (BMC) to coordinate the two major R&D departments, streamlining end-to-end decision-making from technical roadmap to real-world deployment.
Less than two months later, leading LLM technical expert Sun Tianxiang officially joined Baidu, taking over as head of the Basic Model Research Department (BMU) and concurrently joining the Model Committee, filling the core leadership position in the fundamental R&D line.
Nearly all mainstream LLM vendors are frequently adjusting their organizational structures to boost model R&D efficiency, and Baidu established its Basic Model Research Department even earlier than some other major tech giants. The results show this round of adjustments has been positive. In the interval between the two structural overhauls, Ernie briefly seized an opportunity for a comeback. The Ernie 5.0 large language model officially released in January 2026 once restored the reputation among many lapsed users.
Multiple long-time Ernie users told Tech Planet that for high-frequency office scenarios such as text writing and document organization, Ernie 5.0 delivers sufficiently competitive output stability and rationality.
But this positive feedback only lasted four months before Ernie 5.1's launch triggered renewed user complaints. A Baidu insider said, "After 5.1 went live, we could frequently see users complaining on social platforms that the model 'got dumber and stupider'."
What further alienated long-time users was the crude integration of commercialization. Some regular Ernie users reported that after the 5.1 update, whether asking questions or requesting content revisions, responses would inexplicably include ads. "For example, when just inquiring about the compensation process after a traffic collision, the response would directly push contact information for a law firm at the end."
The repeated fluctuations in product reputation are essentially a direct reflection of internal organizational turmoil and adjustments.
An employee who recently resigned from Baidu's LLM division attributed Ernie 5.1's poor reputation to the continuous attrition of the core team. In his view, although Baidu was the first major tech company across the entire industry to go all-in on large language models and elevated the business priority to the highest group level, organizational restructuring and project changes kept the team in a constant state of adaptation.
This continuous adaptation process even runs through Baidu's entire AI product strategy.
In June this year, Baidu officially merged the Ernie Bot web version, the Ernie App web version, and the Baidu Ernie Assistant web version. After the consolidation, the history of Baidu having multiple AI assistant brands on its web platform came to an end, making the Baidu Ernie Assistant web version the only To C AI assistant website under Baidu.
Prior to this, Baidu did not concentrate its resources on unifying the entry point, allowing different business lines to develop naturally, which quickly led to a proliferation of AI entry points with overlapping positioning and fragmented responsibilities.
The most typical chaos occurred on the To C side. Ernie Bot, initially launched as the public face of Baidu's LLM, was originally a general-purpose dialogue entry point designed to compete with ChatGPT, but was later renamed "Wen Xiaoyan". Meanwhile, the search division was simultaneously advancing its own AI transformation, successively launching Baidu AI Assistant, Baidu AI Search, and other similar products around the search box — churning through names repeatedly while their core functions remained highly overlapping.
The costs of this chaos were ultimately passed on to users, shaping the external perception of Baidu as having scattered entry points, ambiguous positioning, and inconsistent user experiences. Even a Baidu employee admitted that not to mention ordinary users, she herself could not keep track of how many different AI product names the company had launched.
After bringing all AI entry points under unified management, Baidu can now concentrate its efforts in one direction, reducing internal friction and eliminating the need for users to jump between multiple products. This is a positive signal for Baidu.
With the entry points unified, Baidu hopes to shore up the shortcomings in its foundational model capabilities, but the results of these efforts remain uncertain. Right now, hopes of turning the situation around rest on Sun Tianxiang, who has just taken over the Basic Model Research Department. The day he delivers a new version of the Ernie large language model may be the moment the public reassesses Baidu's AI progress from its organizational structure to its product offerings.
Rather than betting heavily on the C-end, Baidu wants to use AI to extend the lifespan of its search business
A mature company of a certain scale often faces internal obstacles when exploring new businesses. These obstacles stem partly from resource competition between different departments and teams over new initiatives, and partly from the power struggles between traditional pillar businesses and new ventures.
During the 2026 Spring Festival AI red envelope war, JD.com, Alibaba, Tencent, ByteDance, Baidu and other companies all joined the fray. But the difference was that ByteDance brought Doubao, Alibaba brought Qwen, and Tencent brought Yuanbao — all three being AI-native applications. In contrast, Baidu deployed the Baidu App, with Ernie Assistant embedded inside it.
This move revealed Baidu's final strategic choice: unlike Alibaba, Tencent, and ByteDance, which are building standalone AI applications, Baidu ultimately decided to position AI as the fuel for an "upgraded search engine".
In other words, this veteran internet giant with search as its core business ultimately defined Ernie Bot's role within Baidu Search. For this reason, some Baidu employees joked that Ernie Bot ended up "living" as a plugin for search.
However, an insider revealed that Ernie Bot was once highly anticipated on the C-end, but the misalignment between market shifts and internal decision timelines prevented it from becoming a hit standalone product like Doubao.
In its early launch phase, Ernie Bot immediately rolled out commercialization attempts. But user traffic at that time was not stable enough, and the monetization efforts were soon reversed, derailing revenue plans and causing dissatisfaction among paying users.
Yet beyond the market criticism of being "not decisive enough", an industry insider offered a different perspective: this is precisely the best choice Baidu could make at the moment.
In his view, retreating its main position behind search and other mature businesses allows Ernie Bot to avoid the risk of siphoning off Baidu App's existing search traffic if it grew large enough as an independent product. Meanwhile, embedding AI capabilities deep into its core search business injects new growth potential into Baidu Search.
Data from StatCounter shows that as of April 2026, Baidu holds approximately 44.6% of China's general search market share, still ranking first but down drastically from its peak of over 80%, and continuously facing competition from "decentralized search" platforms like WeChat, Douyin, and Xiaohongshu.
By leveraging AI to simultaneously drive the brand revival of the Baidu App, new possibilities for the search business might yet emerge. According to QuestMobile data, Ernie Assistant reached 360 million monthly active users in the first quarter of this year. Riding on Baidu Search's 700 million monthly active users, Ernie is no longer a marginal offering.
Once the strategic direction was clarified, Baidu accelerated its integration efforts. After TPG was split last year, MEG (Mobile Ecosystem Group) consolidated search resources across PC and mobile terminals to establish a "search-recommendation convergence" team.
Earlier this year, Baidu Wenku and Baidu Netdisk were spun off from MEG to form the new PSIG (Personal Super Intelligent Group), led by Wang Ying, the head of Baidu Wenku and Netdisk, who reports directly to Baidu founder Robin Li.
Since then, AI has begun accelerating its integration into Baidu's internal operations. According to Baidu's disclosures, Ernie Assistant now connects to Baidu's in-house ecosystem services including Baidu Maps and Baidu Health, and links to external platforms such as JD.com, Ctrip, and Meituan.
Another reason for de-prioritizing heavy bets on the C-end may lie in capital constraints. Take the wildly popular Doubao as an example: Guolian Securities estimates that even at the lowest cost per query for Doubao's model, its daily free AI service expenses reach 132 million to 240 million yuan.
For Baidu, using its limited available funds to gamble on an uncertain commercialization prospect at this stage would be overly extravagant.
Notably, Baidu Wenku represents Baidu's most successful use of AI to revamp its existing C-end business. Wang Ying, the head of both Baidu Wenku and Baidu Netdisk, was promoted from Baidu Group Vice President to leader of a core business group in January this year. The standout commercial performance of Wenku and Netdisk, with their paying user base exceeding 40 million and high profit margins, was the key factor behind her promotion.
At this year's World Artificial Intelligence Conference, the featured agent product "Baidu Buddy" is built on the technical foundation formed by the integration of Wenku and Netdisk. However, the revenue from Wenku and Netdisk accounts for less than 10% of Baidu's core revenue.
Can the B-end and chips sustain Baidu's future?
When AI to C cannot single-handedly carry Baidu's current performance, the burden naturally shifts to the B-end. An employee confirmed to Tech Planet that Baidu's internal focus on AI has now largely shifted to the B-end.
The B-end is closest to commercial realization. This is also reflected in Baidu founder Robin Li's 2026 OKRs, where the number of ambitious targets has been raised to eight, all centered on driving the implementation of AI technologies across business scenarios to unlock commercial value.
In the first-quarter earnings report, a notable data point drew widespread attention: AI business revenue reached 13.6 billion yuan, accounting for 52% of Baidu's general business revenue, crossing the 50% threshold for the first time. Robin Li stated on this occasion that AI has become Baidu's core growth driver.
But breaking down this revenue figure reveals the composition of Baidu's AI revenue: in Q1 2026, intelligent cloud infrastructure revenue hit 8.8 billion yuan, up 79% year-on-year, with GPU cloud revenue surging 184% year-on-year; in the same period, AI application revenue (Baidu Wenku, Baidu Netdisk, and digital employees, etc.) reached 2.5 billion yuan.
This means AI cloud infrastructure already makes up well over half of total AI business revenue. It would be more accurate to say Baidu has reaped the dividends of AI infrastructure than to claim AI has become Baidu's core growth driver.
In a sense, this can be interpreted as AI not yet deeply integrated into Baidu's core businesses to generate substantial revenue, but rather allowing Baidu to become an early "shovel seller" in the AI gold rush.
Looking at Baidu's overall performance, in Q1 2026, Baidu's online marketing service revenue was 12.6 billion yuan, down 22% year-on-year, with its proportion of general business revenue dropping to 48% from 63% in the same period last year — marking the first time search advertising's share fell below 50%. During the same period, Baidu's net profit attributable to shareholders was 3.445 billion yuan, plummeting 55.36% year-on-year from 7.717 billion yuan a year earlier.
AI has not yet reinvigorated the search business. Instead, AI and search have temporarily become two businesses that see their fortunes wax and wane in inverse correlation.
Baidu is still searching for new opportunities. Two months ago, Robin Li proposed the new concept of DAA (Daily Active Agents) at Create 2026. While the public joked Baidu is living in an era of coining new buzzwords, agents are indeed the core focus of Baidu's current efforts.
An insider stated that concentrating AI on the B-end means identifying more specific production scenarios, and the core "tool" for addressing these scenarios points directly to agents. Already a series of agents including Baidu Buddy, Miaoda, Famou, Yijing, and GenFlow have emerged from within Baidu.
However, similar to the high hopes once placed on Ernie Bot, Baidu's emphasis on "agents" carries a hint of urgency.
What is slightly more optimistic is that Baidu has now launched the full-stack "Chip-Cloud-Model-Agent" system. On the chip front, Kunlunxin is about to list in Hong Kong; on the cloud front, Baidu still maintains advantages in the scale game; on the model front, the Model Committee has been established, and a cohort of young researchers including Sun Tianxiang have been pushed to the forefront.
As for how far the "agent" layer that ultimately reaches production scenarios can go, it remains to be seen whether the preceding three layers can truly be fully integrated.
Nevertheless, when it comes to chips as the foundational infrastructure of the AI era, Kunlunxin is no slouch. Its external revenue already exceeds 50% of total revenue, with clients including China Mobile and Tencent. In terms of performance, Kunlunxin especially emphasizes cluster efficiency and engineering implementation: its 32,000-card cluster achieves an effective training rate of 97%, and 256-card super nodes have already been commercialized. During the 2026 WAIC (World Artificial Intelligence Conference), Baidu showcased 32/64-card and 256-card super nodes, drawing many visitors to its booth.
From Ernie Bot to agents and chips, Baidu keeps betting its future on the next card to play. This time, the cards are already on the table — it remains to be seen whether Baidu can play them right.