Li Yanhong and Baidu May Finally Be About to Catch the "Last Train" of AI

11/19 2026 429

Source | Business Light Year

Over the past few years, whenever Baidu and AI come up in conversation, there has always been a slightly awkward question hanging in the air: Baidu got started very early, so why is it that when the market talks about this round of large models, the first names that come up are usually OpenAI, DeepSeek, ByteDance and Alibaba?

On November 12, Li Yanhong spoke about "intelligent organizations" at Baidu World 2026. Four days later, Baidu released its third-quarter report: AI business revenue reached 13.2 billion yuan, accounting for 53% of Baidu's general business revenue—a share that has now exceeded half for three consecutive quarters. AI cloud infrastructure revenue came in at 8.2 billion yuan, up 94% year on year.

Looking at the conference and the financial report together, our assessment of Baidu has shifted somewhat. Li Yanhong and Baidu do have a chance to catch up this time, and some businesses have even already secured a seat in the "carriage." The basis is not which ranking a particular model holds on a leaderboard, nor how many new terms were unveiled at a conference. Customers have begun paying for computing power and model calls, and AI revenue has begun to occupy a major position in the financial statements. These changes are more tangible.

Baidu is still carrying heavy pressure from its old businesses. In the third quarter, online marketing services revenue fell 22%, and the company's total revenue has still not returned to growth. It is obviously too early to say that Baidu has completed its transformation. A more appropriate statement is that years of investment have finally shown signs of sustainable commercial returns, and Li Yanhong has waited for a clearer opportunity than in the past.

01

Baidu Got on the Train Very Early; the Hard Part Is Turning Technology into Business

It cannot be said that Baidu has just started catching up in AI. At the 2016 Baidu World Conference, the theme already included "AI is the new electricity." Over the following decade, from autonomous driving and Kunlun chips to deep learning frameworks and the ERNIE large model, Baidu gradually built a technology architecture spanning chips, cloud, large models and agents.

This long-term investment has left Baidu with technological accumulation, but it has also brought prolonged doubt. There has always been a gap between technology announcements and company growth. Autonomous driving needs to pass tests of safety, cost and urban operations; enterprise AI projects have to face deployment and delivery; and large models require continuous purchases of computing power. Each of these can prove that the technology has value, but none of them automatically turns into large-scale profit.

Baidu's situation is therefore somewhat special. It did not miss AI, yet it has repeatedly been asked to prove that it can make money from AI. What the market really doubts is the speed of commercialization and whether Baidu can build another sufficiently large business beyond search advertising.

This round of large models has changed an important condition. In the past, when enterprises bought AI, they often started with a single project—building a recognition system, setting up a knowledge base, or trying an intelligent customer service tool. Demand was fragmented, delivery cycles were long, and revenue was hard to replicate quickly. After the spread of large models, computing power rental, model calls and agent development have gradually become more standardized services. The cloud, chips and model capabilities that Baidu built early on have finally encountered demand that more easily forms continuous revenue.

This year's conference stated this change more directly. Shen Dou proposed that traditional cloud serves enterprise IT systems, while intelligent cloud serves specific tasks. On this basis, Baidu AI Cloud upgraded its new full-stack AI cloud, dividing products into Agent Infra and AI Infra. The former handles agent operation, sandboxes, model services and enterprise system calls, while the latter undertakes training and inference. The previously scattered technology investments have begun to be rearranged around the tasks customers need to complete.

Therefore, Baidu's opportunity today cannot simply be understood as technology suddenly catching up with peers. Only when technology, cost and enterprise demand reach the same point in time can past investments have a chance to connect. For Li Yanhong, this window has not come early, but it is not so late that Baidu has lost eligibility.

02

The Most Important Thing in the Third-Quarter Report Is Not Turning a Profit, but the Change in Revenue Structure

Baidu's total revenue in the third quarter was 30.8 billion yuan, down 1% year on year. Looking at this number alone, it is hard to conclude that an AI company is accelerating. What is truly worth breaking down is Baidu's general business.

Revenue from this part was 25 billion yuan, up 1% year on year. Among it, AI business revenue was 13.2 billion yuan, up 37% year on year, and its share rose from 39% in the same period last year to 53%. Online marketing services revenue was 11.9 billion yuan, down 22% year on year, and its share fell from 62% to 48%. Putting the two sets of data together, Baidu's revenue structure has clearly changed.

AI cloud infrastructure was the fastest-growing item. Revenue of 8.2 billion yuan corresponded to a year-on-year growth rate of 94%, further accelerating from 50% in the second quarter. GPU cloud revenue grew 191% year on year and 43% quarter on quarter. Token call revenue from external customers on the Qianfan platform grew 1,159% year on year. This percentage is affected by the base and cannot be directly converted into an equivalent scale of revenue increment, but it at least shows that external calls are increasing.

The financial report also mentioned that AI cloud's customer base is more diverse, customer prepayments have increased, and contracts continue to shift toward longer terms. For the cloud business, this information is crucial. Customers being willing to pay in advance and sign longer contracts usually says more about whether demand is stable than a single trial does. On the earnings call, Li Yanhong said that the profit and profit margin of AI cloud infrastructure improved both year on year and quarter on quarter, and that the investment payback period is shortening. This is management's statement, and it still needs to be verified later by profit and cash flow.

This also explains why Baidu now speaks about AI with more confidence than before. In the early years, when it talked about technological prospects, the outside world could only believe that there would be revenue in the future. Today, cloud revenue, customer prepayments and contract terms can already provide an observation window. Baidu still has to invest a lot of money, but at least it can now see where the money will come back from.

The new products announced at the conference correspond to the next round of investment. The Tianchi 1024-card super node appeared at the conference. The company said training performance is 2.5 times that of the previous generation and inference performance is 3 times higher, and the product is expected to launch in 2027. The Kunlun chip M100 has already been launched, and the M300 is planned for launch next year. Announcement, launch and revenue generation are different stages, and the commercial contribution of these products still needs to be verified by shipments, customers and profit.

In the third quarter, Baidu's operating cash flow was 2.8 billion yuan, remaining positive. AI cloud is still just one business, but its contribution to overall profit is increasing. For a company that has long been asked about AI returns, this is more convincing than a technology demonstration at a launch event.

03

Why Li Yanhong Changed to Talking About Intelligent Organizations

The theme of this year's Baidu World Conference was "The Next Act: Intelligent Organizations." Li Yanhong compared AI to new electricity and called intelligent organizations the "new factory" of the AI era. When an enterprise buys a large model, it is equivalent to installing a new motor. If the production line, positions and ways of collaboration do not change, the new motor will hardly bring the imagined efficiency.

This formulation responds to one of the most common things enterprises encounter when using AI. Employees have tried chatbots and have written several versions of copy and code, yet the company's costs have not fallen noticeably. The problem often lies in AI not entering the complete workflow. It does not know the company's internal materials, cannot call business systems, no one checks the results after it is done, and it is even less able to leave experience for the next task.

For this reason, Li Yanhong proposed the "three principles of organizational evolution": breaking job boundaries, precipitating tacit knowledge, and designing closed-loop processes. Jobs are the unit of the industrial age, while tasks are regarded as the unit of the intelligent age; the experience in employees' minds should become callable organizational Skills; and after an agent completes a task, it should continuously improve through iteration around measurable indicators. He summarized the last point as "measurable + Looping = continuous evolution."

Whether this method works depends on whether it can truly improve the business, not on how many agents an enterprise has deployed. Baidu first ran an experiment on itself. The company disclosed that over the past year, the number of internally daily active agents increased from 5,000 per day on average to 37,000, the number of effective tasks completed per week increased from 84,000 to more than 1 million, 4,300 Workspaces have been built, and more than 20,000 organizational Skills have been accumulated. These figures come from Baidu itself and cannot yet represent that other enterprises can achieve the same results, but they at least show that "intelligent organizations" have a set of actual operating objects inside Baidu.

External cases provide another perspective for observation. A dispatching institution of China Southern Power Grid connected the experience of veteran masters to "Dawate Tianxuan Partner," and Baidu said the efficiency of self-healing fault analysis in distribution networks increased 12 times. Lanhai Heishi used three agents to connect ERP, OA and WeCom. A procurement process that originally required at least 10 people from two teams and a cumulative 100 work hours each week was shortened to less than 1 hour with 2 people. The quarterly profit increment in the latter case is still a calculation made by the enterprise based on current results and cannot be treated in advance as realized income.

Products such as Baidu Dazi, Famou, Yijing and Kuku are all moving toward specific tasks. The video of the digital human Xijiajia interviewing Li Yanhong shown at the conference was generated by Baidu Yijing; Famou 3.0 emphasizes being responsible for final business indicators. Product demonstrations do not mean customers will definitely renew, but Baidu has at least advanced its sales pitch from "how smart the model is" to "how the task is completed."

04

Behind Token Exchange, Baidu Wants to Switch to a Broader Business

The more models there are, the harder it becomes for enterprises to choose. The same type of model may come from different providers, and price, stability, response speed and service responsibility are not always consistent. At the conference, Baidu launched the large model trading platform "Token Exchange," which does verification and credit reporting: it verifies resources through model fingerprint comparison, capability testing and performance monitoring, and then conducts continuous spot checks on providers and establishes credit ratings.

After one month of trial operation, Token Exchange has already onboarded more than 100 supply resources, with average daily calls exceeding 50 million and average daily token calls exceeding 700 billion. On November 14, Baidu further disclosed that more than 100 suppliers applied to join on the first day of launch, and more than 1,000 customers were newly connected. Applications and connections are only the starting point; subsequent active calls, payments and retention are more valuable.

It is still too early to conclude whether this business can form an independent scale. The direction it reveals is clear. Baidu hopes to participate in the entire process of enterprises choosing models, calling models and transforming processes, and revenue does not have to come only from the ERNIE model.

The conference also released ERNIE 5.5 Flash Preview, the ERNIE Lightning deep thinking model and the ERNIE Yijing micro-expression large model. Models remain Baidu's technology foundation, but the commercial focus is no longer limited to selling its own models. Qianfan provides model services, Baihe provides computing power, Dazi connects professional agents, and Token Exchange connects different suppliers. Baidu wants to turn "chip, cloud, model and agent" into a set of enterprise services that charge continuously, rather than updating technological terms once a year.

This path also raises issues of balancing interests. The platform both operates its own models and provides trading and credit services for different models. How to ensure neutral evaluation, data security and customers' right to choose will affect whether external partners are willing to stay for the long term. It is easy to shout openness, but whether an ecosystem can be built depends on whether partners can truly obtain customers and revenue.

05

Baidu Has Caught the Window, but It Still Has to Keep Answering the Profit Question

If one looks only at the conference and the growth rate of AI cloud, it is easy to write Baidu as a story of completed turnaround. The income statement is not so relaxed.

In the third quarter, Baidu's AI application revenue was 2.6 billion yuan, basically flat year on year. AI-native marketing services revenue was 2.4 billion yuan, down 15% year on year. The company explained that at the current stage of search's transition to AI, priority is given to improving user experience, and short-term monetization is placed later. This choice is reasonable; if users are unwilling to use the new search method, it will also be difficult to maintain ad slots and paid products. It also means that Baidu's most important user entrance is still bearing transformation costs.

The decline speed of legacy businesses also cannot be ignored. Online marketing services revenue fell 22% year over year, and AI business growth is currently mainly offsetting this decline and helping the general business return to slight growth. There is still some distance to go before it can drive the company's overall revenue back into acceleration.

Profit also needs to be broken down. In the third quarter, net profit attributable to Baidu was 1.4 billion yuan, compared with a huge loss in the same period last year. The same period last year included 16.190 billion yuan in long-lived asset impairment, so the year-on-year turnaround cannot be entirely understood as a sudden improvement in operating capability. Under non-GAAP, net profit attributable to the parent company this quarter was 2.4 billion yuan, which, calculated from announced data, is down about 36% year over year. Operating profit was 626 million yuan, and the operating margin was only 2%.

The larger the AI business becomes, the more specific the cost problem becomes. Baidu's cost of sales in the third quarter was 19.9 billion yuan, up 9% year over year, which the company explained was mainly due to increased cloud-related costs and traffic acquisition costs. R&D expenses were 4.8 billion yuan, down 7% year over year. When AI demand is growing rapidly, how to balance R&D investment, computing power procurement and profit will directly affect how much money this business can retain.

In the coming quarters, the revenue growth rate and profit margin of AI cloud are worth looking at together. AI applications also need to provide a second growth point. Wenku, Wangpan, digital humans and industry agents all have user or customer foundations. After product upgrades, whether paying users, renewals and average revenue per customer can improve is more important than how many new features are added. Apollo Go's covered cities and cumulative mileage are still growing, and its commercial value also needs to be calculated one by one in combination with fleet costs, order density and operating cities.

Despite retaining these questions, we are still willing to give Baidu a more positive judgment than in the past. AI revenue has accounted for more than half of general business revenue for three consecutive quarters, cloud revenue is close to doubling, customers have begun to prepay and sign longer contracts, and agents are also undertaking real tasks inside the company. The simultaneous appearance of these things shows that Baidu's years of investment have begun to connect with market demand.

The so-called last train will not suddenly close its doors one night. Foundation models will still change, applications will still be reshuffled, and enterprise customers' choices are far from settled. Baidu cannot obtain a permanent seat by relying on its early qualifications. What it has obtained now is an opportunity to prove itself again through revenue and profit.

This time, besides technological judgment, Li Yanhong also has a financial report that has begun to show changes. For Baidu, this is already a new starting point important enough.

Solemnly declare: the copyright of this article belongs to the original author. The reprinted article is only for the purpose of spreading more information. If the author's information is marked incorrectly, please contact us immediately to modify or delete it. Thank you.