Home / Transcripts / Baidu, Inc. (BIDU) · November 16, 2026

Baidu, Inc. (BIDU) Earnings Call Transcript

November 16, 2026

NASDAQ US Communication Services Interactive Media and Services earnings

Earnings Call Speaker Segments

Juan Lin executive
#1

Good day, everyone, and welcome to the earnings conference call for Baidu's third quarter of 2026. Baidu's earnings release went out earlier today, and a copy is available on our website and through newswire services. Joining us on the call today are Robin Li, our Co-Founder and CEO; Julius Rong Luo, our EVP in charge of Baidu Mobile Ecosystem Group, MEG; Dou Shen, our EVP in charge of Baidu AI Cloud Group, ACG; and Henry Haijian He, our CFO. Following our prepared remarks, we will hold a Q&A session. Please be aware that today's discussion will include forward-looking statements made under the safe harbor provisions of the U.S. Private Securities Litigation Reform Act of 1995. Such forward-looking statements carry risks and uncertainties that could cause actual results to differ materially from our current expectations; for detailed discussions of these risks and uncertainties, please refer to our latest annual report and other filings with SEC and Hong Kong Stock Exchange. Except as required under applicable law, Baidu does not undertake any obligation to update any forward-looking statements. Our earnings press release and this call include discussions of certain unaudited non-GAAP financial measures. A reconciliation of the unaudited non-GAAP measures to the unaudited most directly comparable GAAP measures is contained in our press release, which is available on our IR website at ir.baidu.com. As a reminder, this conference is being recorded. In addition, a webcast of this conference call will be available on Baidu's IR website. I will now turn the call over to our CEO, Robin.

Robin Li executive
#2

Hi everyone. During Q3, Baidu's general business went back to revenue growth versus a year earlier, rising 1% to RMB 25.0 billion, mainly on the back of strong momentum in AI Cloud Infra. Baidu core AI-powered business brought in revenue of RMB 13.2 billion, which represented 53% of Baidu general business revenue and keeps expanding as a portion of the total. AI Cloud Infra revenue almost doubled compared with a year earlier, and growth sped up sharply, from 50% in the prior quarter to 94%, well ahead of the broader market. Inside AI Cloud Infra, GPU Cloud kept up triple-digit revenue growth versus a year earlier and rose 43% quarter-over-quarter, which reflects robust and sustained demand. On the MaaS side, momentum was also strong: external customers' token consumption on our Qianfan MaaS platform generated revenue that rose by over 12x compared with a year earlier in Q3. This quarter's results further demonstrated that our long-term bets on full-stack AI are converting into meaningful business expansion, and AI Cloud Infra is now becoming a major force behind Baidu's revenue growth overall. We think our full-stack AI strengths—spanning chips and cloud infrastructure through to models and applications—leave us well placed to serve rising demand and seize fresh opportunities. Looking ahead, we anticipate that AI Cloud Infra will serve as a progressively more significant growth engine for Baidu as we continue to build on this momentum. Let's take a closer look now at AI Cloud Infra. This quarter we delivered strong revenue growth there, made gains in profitability, expanded our customer base, and locked in longer-term contracts. GPU Cloud drove most of the revenue increase for AI Cloud Infra. GPU Cloud revenue in Q3 was up 191% year-over-year, and that came despite a tough comparison base, while sequential growth sped up to roughly 43%. That shows how strong demand is for scalable AI computing in the public cloud, and it shows we can satisfy that demand with dependable, high-performance capacity at scale. We're confident AI Cloud Infra will keep up its strong growth momentum over the coming quarters. Beyond the top-line increase, AI Cloud Infra posted bigger profits and wider margins on both a year-over-year and sequential basis, and it made up a considerably larger portion of Baidu's total profit in this quarter. That gain came on the back of a better revenue mix, as higher-margin products — GPU Cloud, Kunlunxin, and MaaS — now represent most of AI Cloud Infra's revenue. As those offerings take a larger slice of revenue, margins keep getting better, and resource utilization rises, overall profit and margin expansion follow. At the same time, demand stays robust, our scale keeps growing, and profitability improves steadily, so payback periods on cloud infrastructure investments are getting shorter — which makes AI Cloud Infra a healthier and more sustainable business. There are a number of important drivers behind the ongoing expansion of our AI Cloud Infra. To begin with, the need for AI computing kept speeding up, and genuine business requirements are what underlie it. Both training and inference workloads showed robust demand, and inference in particular is expanding quickly now that AI applications and agents are being rolled out across an ever-broader set of business use cases. All the same, we think inference remains at an early point in its scaling journey, leaving considerable space for additional growth. In addition, our customer base expanded considerably in Q3 and turned more diverse in terms of sectors and business sizes; we continued to add new clients, while existing ones increased both how much they used and how much they spent. Momentum stayed solid in important areas including e-commerce, internet, autonomous driving, and embodied AI. More and more, customers are putting AI at the heart of how they operate, which generates recurring and expanding need for AI computing; take, for instance, a top e-commerce firm that relies on our cloud infrastructure to train and refine multimodal foundation models underpinning its retail, logistics, and financial services operations. Every model refresh brings new training requirements, and wider rollout fuels continued expansion in inference usage, plus added need for data processing, storage, and networking services. In embodied AI, a top robotics firm employs our GPU Cloud to train the models that drive its robots. Every new task or environment calls for new data gathering, model training, simulation, and testing. As that firm moves into additional tasks, robot models, and deployment situations, its computing requirements rise at each development stage. Once deployed, the robots create continuous inference demand and return real-world operating data into model development. Combined, these workloads generate persistent need for GPU computing and associated cloud services. Third, we are observing stronger and longer-term commitments from customers. Clients are mapping out computing needs further in advance and increasingly pursuing long-term partnerships so they can lock in stable, dependable, and scalable capacity as they expand. Our industry-leading technology, service quality, and demonstrated track record give them confidence to place these needs with us. Within GPU Cloud, more and more customers are prepaying to lock in capacity for later and committing to contracts that run longer. As an example, each of our three largest GPU Cloud customers measured by contract value renewed with us this year and moved to multiyear terms, and one of those renewals was for a term six times longer than before. Commitments of that kind give us clearer sight of upcoming demand and support better cash flow. Turning to MaaS, appetite for model inference continued to accelerate. Qianfan, our MaaS platform that leads the industry, offers ERNIE along with the latest editions of other top open-source models from China, so customers get quick access to the most advanced capabilities. During Q3, we further tuned inference to boost throughput and cut latency, while enhancing resource scheduling to balance workloads and lift computing utilization. These upgrades make model inference on Qianfan more dependable, stable, and cost-efficient; in Q3, revenue from external customers' token usage on Qianfan climbed more than 12 times year-over-year, and margins kept improving. As MaaS grows and accounts for a bigger portion of AI Cloud Infra revenue, we believe it will contribute ever more meaningfully to overall profit growth. Now, let me move to the model layer. Furthering ERNIE stays a strategic priority. At Baidu World last week, we introduced ERNIE 5.5 Flash Preview, the newest foundation model in our flagship ERNIE family. It has 280 billion total parameters, yet just 14 billion get activated for each token, and that gives it impressive coding and agentic abilities, so it can coordinate professional software, tools, and external models to handle complicated tasks from start to finish; thanks to these abilities, users can generate usable, high-quality outputs in knowledge work, game development, 3D modeling, and video creation. On the most recent Code Arena WebDev leaderboard, ERNIE 5.5 Flash Preview sits in second place among open-source models worldwide. Looking ahead, we intend to speed up model development throughout the ERNIE family and keep strengthening what it can do. Our model development approach stays application-driven. Every iteration is shaped by what our applications require. ERNIE Lite DeepThinking model is a solid illustration, since it was first built for AI Search and is now offered as well to customers and partners via our Qianfan MaaS platform. It is excellent at grasping user intent, reasoning through intricate queries, and invoking the correct tools, all while returning quick responses at scale. Feedback from Search lets us keep refining the model, and those refinements help Search provide smarter answers and accomplish more for users. Tuning across our entire AI stack drives inference costs down further, which lets us extend these capabilities to additional users. This is how model innovation turns into improved user experiences and bigger cost savings. Across the wider market, appetite for model services is rising rapidly.

Robin Li executive
#3

Therefore, to make it easier to find and utilize high-quality model services, we launched Token Exchange at Baidu World. This marketplace is neutral, transparent, and trusted, linking providers of AI models with customers. Buyers can evaluate model services alongside each other, seeing straightforward pricing and performance information, and access everything through a single destination. Meanwhile, quality providers obtain a route to reach additional customers. As time goes on, we anticipate that Token Exchange will help activate underutilized compute and make the AI and foundation model ecosystem healthier and more efficient. Shifting to AI applications, we continue broadening adoption and reinforcing product capabilities. General work agents are still a very promising space, and adoption keeps rising among individuals as well as enterprises. DuMate — the general work agent we launched earlier this year — has expanded quickly. Its MAU in September rose over 30-fold compared with the prior quarter. That increase happened without big-budget marketing pushes or being plugged into an existing super app. Next-day retention has also been strong, which suggests users are having a good experience. Since the enterprise edition debuted in June, DuMate has picked up considerable momentum with businesses, helping it move into a wider set of enterprise and industry-specific workflows. During this quarter, we keep building out DuMate by bringing in capabilities from throughout Baidu's AI ecosystem, such as our internal workplace AI agent Dudu, our AI coding assistant Comate, and our BI coding platform Miaoda. Combining these proven features into a single agent allows users to finish harder, multi-step jobs with greater efficiency. At the same time, we keep building specialized agents for fields where demand is strong and our models are tuned to perform well. In August, GenFlow — created jointly by Baidu Wenku and Baidu Drive — was upgraded and given the new name KuaKua, an AI-driven platform for documents and content creation. Leveraging the broad content resources and user ecosystems of Baidu Wenku and Baidu Drive, KuaKua unites publicly available information, professional databases, and users' own content, helping them research, write, and produce high-quality artifacts for work and life. Baidu Yijing is our hyper-realistic digital human agent. The company's tech is the industry's most advanced, letting it reliably equal or surpass human hosts in e-commerce livestreaming. During this quarter, we rolled that technology out at scale on our own platform, while also boosting its cost efficiency and the speed of production. Starting in the fourth quarter of 2025, we have reduced the cost of producing each hyper-realistic digital human by roughly 80%, and shortened production time by more than tenfold; thanks to these gains, our hyper-realistic digital humans are becoming more accessible and scalable. Adoption of Yijing by customers keeps expanding. One big client has been ramping up how much it relies on our digital human tech for multiple quarters running, which shows it still trusts what we've built. During this quarter we also stretched into more verticals, bringing on board fresh customers in financial services, automotive, apparel and various other areas. On the international side, we built on the partnerships we already had with TikTok and Shopee and grew our client list further by signing Lazada. Yijing is still in development, and in our view it carries substantial potential to reshape how businesses engage with their customers across numerous industries. We're additionally applying frontier AI to hard operational challenges in the real world. At Baidu World we introduced Farmo 3.0, which lifts its performance further and broadens how it gets used in tough industrial settings. An industrial laser equipment manufacturer, one of our customers, relies on Farmo to work out how four robotic arms coordinate to cut a large car body panel with intricate curves. Every plan has to spell out how the work is divided, the sequence of cuts, and the route each arm takes, keeping all four arms synchronized and preventing collisions. By learning from production data, Farmo independently explores a huge set of potential combinations, assessing and repeatedly honing plans until it locates the most efficient one. Creating such a plan once required 4 to 8 hours; with Farmo, that customer now completes it in roughly 20 minutes. Across our AI applications portfolio, we continue to expand what we offer, enhance what those products can do, and simplify how people use them. Monetization remains early, yet we anticipate considerable upside as these apps become part of additional workflows and provide value users can measure. In Q3, we continued improving the experience on AI Search. AI answers became better organized, tighter, and easier to move through, helping users find and understand what they need more efficiently. We also brought AI Search and ERNIE Assistant into closer integration, improving how well we capture user intent and handle more complicated requirements. ERNIE Assistant leverages pertinent follow-up questions to help people clarify and dig into whatever they need, extending the search experience into conversations that feel more natural and go deeper. Thanks to its agentic abilities, it handles complicated tasks on users' behalf without friction. During September, daily average conversation rounds for ERNIE Assistant increased 126% compared with a year earlier, and both engagement and stickiness among users also rose. Looking ahead, we'll continue prioritizing user experience over near-term monetization as we advance our AI Search transformation. While that will pressure our online marketing revenue and overall profitability in the near term, we're convinced it's the correct route to bolster our long-term competitiveness and unlock the full potential of AI Search. On the physical side, Apollo Go—our autonomous ride-hailing business—is bringing smart mobility to a growing number of users worldwide. With technology continuing to improve and our operations spreading, we're moving into a new stage of development, concentrating on cities where demand and commercial potential are greatest; in those places, our focus is on deepening market penetration and expanding operations in an efficient, sustainable way; in Shenzhen, which is now one of our biggest and fastest-growing markets, fully driverless operational rides increased roughly 170% quarter-over-quarter in Q3, just a few months after we ramped up our efforts there. Over the course of the quarter, we broadened the areas we serve and introduced additional pick-up and drop-off locations throughout the city, while also making operations more efficient and enhancing how riders experience the service; because of these steps, we were able to satisfy more demand and deliver a service that is more convenient and easier to access. We believe there is considerable space for additional growth as we keep strengthening our footprint in Shenzhen. In Dubai, we are continuing to grow our commercial operations that run fully without a driver and to widen the areas we serve. The Apollo Go fleet there has doubled in size, which further reinforces our standing as the largest provider of genuinely driverless ride-hailing in Dubai. Since rides can be booked through the Apollo Go app as well as the Uber app, more riders are gaining access to our service, and our commercial presence in that market keeps expanding. Apollo Go likewise maintained its industry lead in making progress across major right-hand drive, left-hand traffic markets; in London, we steadily enlarged the areas where we conduct open-road testing alongside Uber and Lyft, building greater familiarity with local roads, traffic, and operating conditions. In Hong Kong, fully driverless testing moved forward at a steady pace on Airport Island, where we accumulated additional mileage to underpin the next phase of development. With a global footprint that now reaches 30 cities, our attention is fixed on strengthening our presence in priority markets. Safety will stay our foremost priority as we widen service coverage, raise fleet utilization, and improve the rider experience.

Robin Li executive
#4

Through efficient, reliable, and sustainable operations that we build, our goal is to support Apollo Go's long-term growth and to make it a trusted, integral part of everyday mobility. To sum up, this quarter showed meaningful progress in turning our full-stack AI investments into scaled growth, and AI Cloud Infra served as a key driver. We are also putting growing emphasis on international expansion, bringing our AI capabilities to users and enterprises around the world. We believe the foundation we have built puts us in a good position to capture expanding opportunities ahead. With that, let me hand the call over to Henry to go through the financial results.

Henry Haijian He executive
#5

Thank you, Robin, and hello to all of you. During Q3, revenue from Baidu's general business went back to growth on a year-over-year basis, rising 1%. The share of Baidu general business revenue coming from Baidu core AI-powered business kept expanding, hitting 53%. AI Cloud Infra stood out as the strongest performer in this quarter, with revenue climbing 94% year-over-year. That swift expansion, along with its growing contribution to revenue and profit, helped bring Baidu general business revenue back to growth, which offers clear proof that our second growth curve is taking shape. As we carry this momentum forward, we will keep a disciplined approach to investment, steering capital toward opportunities where demand is clear and expected returns are attractive. We're increasing how much we use the capacity we already possess to help drive revenue growth. Bigger prepayments from customers also accelerate our cash collection, giving us greater flexibility to put money into our cloud and AI businesses. On top of that, we maintain sizable cash reserves, a solid balance sheet, and continuous cash generation from our current businesses. During Q3, Baidu's operating cash flow remained robust at RMB 2.8 billion, representing our fifth quarter in a row of positive operating cash flow; these financial resources, together with improved capital efficiency, enable us to expand our cloud and AI businesses rapidly. Our investment pace will continue to match demand and the cash we generate, which supports growth that is both sustained and profitable. On the topic of returning money to shareholders: we value the long-term backing we receive from them and remain dedicated to giving capital back. As we said before, we anticipate declaring our first cash dividend payment before year-end, assuming the board approves and market conditions allow; beyond that, we plan to continue buying back shares at a suitable scale through our existing program. Going forward, we will keep weighing growth investment against shareholder returns so as to build long-term value for our shareholders. With that, let me take you through the specifics of our third quarter 2026 financial results. Our top line reached RMB 30.8 billion, reflecting a 1% decline from a year earlier and a 2% decrease versus the prior quarter. Baidu general business contributed RMB 25.0 billion, which was up 1% year-over-year but down 1% quarter-over-quarter. iQIYI generated RMB 6.0 billion, marking a 10% drop year-over-year and a 4% sequential decline. Cost of revenues totaled RMB 19.9 billion, rising 9% compared with last year and 4% from the previous quarter. That annual increase stemmed primarily from greater cloud business expenses and traffic acquisition costs, partially offset by reduced content costs. Operating expenses amounted to RMB 10.3 billion, a 13% year-over-year decrease driven mainly by lower expected credit loss expenses and channel spending expenses, while quarter-over-quarter they climbed 12%, largely on account of higher expected credit loss expenses. Operating income came in at RMB 626 million, with an operating margin of 2%. On a non-GAAP basis, operating income reached RMB 1.7 billion and the operating margin was 6%. Net total other income was RMB 1.2 billion, versus RMB 1.9 billion in the same period a year ago and RMB 184 million in the prior quarter. The year-over-year decrease primarily reflected weaker earnings pickup from equity method investments and a smaller fair value gain on long-term investments, along with a bigger net foreign exchange loss stemming from movements in the RMB and U.S. dollar exchange rates, partially offset by exchangeable bonds swinging from a fair value loss to a fair value gain. That sequential increase largely came from a swing on trading debt investments, which moved from a fair value loss to a fair value gain. Income tax expense was RMB 427 million, compared with an income tax benefit of RMB 1.8 billion in the same period last year and income tax expense of RMB 1 billion in the prior quarter. Net income attributable to Baidu came to RMB 1.4 billion. Baidu's net margin was 5%, and diluted earnings per ADS were RMB 4.35. Non-GAAP net income attributable to Baidu amounted to RMB 2.4 billion. Non-GAAP net margin for Baidu was 8%, and non-GAAP diluted earnings per ADS were RMB 6.78. Total cash and investments is how we describe cash, cash equivalents, restricted cash, short-term investments, net long-term time deposits, held-to-maturity investments, and adjusted long-term investments. On September 30, 2026, total cash and investments amounted to RMB 276.6 billion. Operating cash flow was RMB 2.8 billion. As of September 30, 2026, Baidu general business had roughly 26,000 employees. With that, operator, we're ready to take questions on the call.

Alicia Yap analyst
#6

Hello, and thanks. Good evening, management. I appreciate you taking my questions. I'd like to ask about AI Cloud Infra. How does management see the growth trajectory for AI Cloud Infra over the coming few quarters, and what are the main drivers behind that? In addition, could you give us more detail on the revenue mix, and also talk about how sustainable those growth drivers are over the longer term? Thank you.

Dou Shen executive
#7

Thanks, Alicia. It's Dou here, and I'll take that question. During this quarter, revenue for AI Cloud Infra rose 94% compared with a year earlier, outpacing the market yet again. The biggest growth contributor was GPU Cloud, which posted robust gains both year over year and quarter over quarter. We have confidence that AI Cloud Infra will keep up its momentum over the quarters ahead. That confidence stems not just from robust demand and an expanding customer base, but also from enhancements in our revenue mix and customer commitments. First, demand is very strong and accelerating for both training and inference. Companies are no longer merely experimenting with AI; they are embedding it into their core businesses, which turns compute into a recurring need. Agent adoption is still at an early stage. As agents grow more capable and become simpler to deploy, additional businesses will employ them in daily operations, and we believe that will fuel sustained growth in demand. Second, we are converting rising demand into more and deeper customer relationships. During this quarter, we brought in customers spanning more industries and company sizes, and we are expanding our work with existing ones. Take a leading autonomous driving chip company as one example. Its training needs have risen swiftly as it iterates on its models. Our collaboration actually began with GPU Cloud and then grew into storage, networking, and more. Today we are its largest infrastructure provider. This example shows how growing AI demand opens opportunities beyond GPU compute. As customers scale their AI workloads, they require a wider range of cloud services. Our full-stack capabilities put us in position to address those needs, grow alongside our customers, and capture a larger portion of their cloud spending over time. Third, our revenue mix keeps getting better. GPU Cloud, Kunlunxin, and MaaS combined represent over half of our AI Cloud Infra revenue this quarter. GPU Cloud's share climbed further, powered by robust demand for AI compute in the public cloud. MaaS remains small but is expanding quickly, and we expect it to turn into a more meaningful growth driver as customer usage grows. A substantial portion of AI Cloud Infra revenue is also recurring, which gives us a steadier revenue base. Looking ahead, we expect these trends to persist and to support healthier, more sustainable growth. In addition, we are encouraged that more customers are opting for longer contract terms and making prepayments, particularly in GPU Cloud. Many customers extend their contract terms when renewing, and some major customers renew for terms that are several times longer than their original contracts. Customers now see their long-term compute needs more clearly and want to lock in reliable capacity ahead of time. Their experience working with us has also given them confidence in our technology, services, and delivery, so they're willing to commit to longer-term partnerships. These longer contracts and prepayments greatly enhance the visibility and predictability of our future revenue. They also help us anticipate demand, so we can plan capacity and support customers as they expand. Taken together, these factors give AI Cloud Infra a solid foundation for sustainable growth. Thank you.

Alicia Yap analyst
#8

Thank you.

Lincoln Kong analyst
#9

Thanks to the management team for taking my question. I'd like to ask about the model side of things. Could you bring us up to date on ERNIE's roadmap and how it's positioned within China's AI landscape? What role do you expect it to play across Baidu's business, and how should we think about the returns on our continued investment there? Thank you very much.

Robin Li executive
#10

Hello, Lincoln. Robin here. Foundation models represent a crucial layer of our full-stack AI architecture, and we'll keep investing in ERNIE and pushing the frontier of intelligence, which we regard as vital to Baidu's long-term competitiveness. We've just launched ERNIE 5.5 Flash Preview, the newest foundation model in our ERNIE family. It delivers strong coding and agentic capabilities across knowledge work, game development, 3D modeling, and video creation; on the latest Code Arena WebDev leaderboard, ERNIE 5.5 Flash Preview ranks second among open-source models globally. Remember, our model is only about 1/10 the size of the leading one in terms of number of parameters; we aim to accelerate the pace of model iteration and look forward to sharing more progress soon. Our approach to model development remains application-driven. Foundation models span a wide set of capabilities, yet since we operate so many applications, we witness directly what users require and which model enhancements will deliver the greatest impact; in the domains that count most for our applications, we aim for our models to lead. Consider Yijing, our hyper-realistic digital human agent. Driven by ERNIE, our digital humans now achieve highly nuanced, natural-looking visuals. Every smallest facial expression and the faintest gaze shift appears lifelike. We regard these capabilities as among the strongest in the industry. Take ERNIE Lite DeepThinking as another case — a model tailored specifically for AI Search. In coding, reasoning, and agentic work it turns in best-in-class results, so Search can grasp complicated user intent and complete tasks, all while running quickly and at low cost. Those are precisely the abilities our applications require most, and they are where we aim to lead. As for investment, our commitment continues, carried out with discipline: we direct resources toward the highest-priority areas and keep raising training efficiency. Down the road, a more capable ERNIE pays off for us in two respects. First, stronger proprietary models let our apps provide an improved user experience more cheaply, which helps monetization and profitability. Second, via our Qianfan MaaS platform, we extend ERNIE to additional outside customers as they create AI applications. As those applications grow, we anticipate they will generate persistent demand for our model services, further releasing the commercial value of our full-stack AI capabilities and broadening AI Cloud Infra revenue streams. Thank you.

Alex Yao analyst
#11

Thank you, management, for taking my question. My question concerns the cloud margin. With your AI Cloud Infra business continuing to scale, how should we think about the margin outlook, and what are the primary drivers behind that margin? A follow-up to that: given the ongoing investment in compute infrastructure, how do you guys assess the returns on those investments? Thank you.

Dou Shen executive
#12

Thank you, Alex. During Q3, AI Cloud Infra posted year-over-year and quarter-over-quarter gains in both profit and margin, and it has grown into a meaningful contributor to Baidu's overall profitability; that's an encouraging signal that the business is becoming healthier even while scaling quickly. Looking further out, we believe there is room for additional margin expansion and better overall profitability, and four factors drive that. First, as I said a moment ago, our revenue mix keeps getting better. Revenue is increasingly coming from higher-margin segments — GPU Cloud, Kunlunxin, and MaaS — and every one of those gets more profitable as it scales. Gains stem from two places: the shifting mix and better profitability within each segment. Next, greater scale lets us extract more from our resources. As the business expands and workloads become more varied, we can pool capacity across different workloads, enhance scheduling, and smooth out demand peaks and troughs. The result is that the same infrastructure handles more demand while unit costs fall. Third, our full-stack AI capabilities reduce costs through end-to-end optimization. On the infrastructure front, Kunlunxin gives us added flexibility to optimize infrastructure cost and performance. When it comes to the MaaS layer, continuous gains in inference mean the same amount of compute can process additional tokens at reduced latency, all while keeping reliability intact. Fourth, having a complete stack positions us to offer services of greater value. In addition to cloud infrastructure and model services, that stack spans a wide array of agents — among them DuMate, Farmo, and Yijing, which Robin referenced a moment ago. This likewise enables us to address customer requirements across every layer through more comprehensive products, services, and solutions. As uptake grows, we anticipate this will generate greater value for customers and yield a bigger contribution from higher-value services. Together with profitability, we watch closely the net cash investment needed to sustain growth. In evaluating any customer or contract, we look at anticipated revenue and margins, along with payment terms and how long we expect payback to take. Our goal is to recoup the computing investment we make upfront more quickly and to channel cash coming in from customers into funding the following round of investment. The share of new contracts signed this year that carry prepayments is also rising rapidly. As demand grows, our business scales, and margins get better, we're observing shorter payback periods on the cloud infrastructure investments we make. With more prepayments coming in and payback happening faster, we can reinvest sooner and therefore scale more quickly. Taken together, these factors leave us confident about the long-term upside for AI Cloud Infra's margin and profitability. Thank you.

Miranda Zhuang analyst
#13

Thanks. Good evening, management. Thanks for taking my question. I would like to ask about the AI applications. How does management see the competition developing in AI applications and agents? How are companies' AI applications such as DuMate performing? What should we think about the monetization model and the longer-term revenue potential for these products? Thank you.

Robin Li executive
#14

First, I believe that ultimately AI's value has to come from applications. Without applications, then the models, cloud infrastructure, or chips are simply useless. The AI application market is also evolving very rapidly. With improvements in model capabilities, AI applications and agents can now tackle more complex tasks for a growing number of users across a wider variety of scenarios. Agentic products are emerging faster and faster. However the market shifts, we think long-term competitiveness ultimately hinges on the value an application delivers — whether it can complete tasks efficiently and dependably across many real-world applications, produce measurable productivity gains for users, and drive meaningful business outcomes for enterprises. That is why innovation at Baidu has always followed an application-driven approach, beginning with real-world use cases. At present, we have assembled a set of AI applications and agents covering both individual and enterprise customers, addressing many different tasks and scenarios. This set includes DuMate, our general work agent, along with specialized agents like Yijing, our digital human agent; Farmo, our self-evolving agent; and KuaKua, which is for creating AI artifacts. Consider DuMate: its user base is expanding quite rapidly, mostly organically, without significant marketing expenditure or traffic coming from our other products. Such signals are encouraging, showing that users see how effectively the product performs. Regarding monetization, AI-native applications remain at a very early stage, and the entire industry is still figuring things out. Products differ in their audiences, their functions, and the value they deliver—so it makes sense that monetization approaches won’t be uniform. One point we’re certain about: people pay when they get value, and delivering that value is the core purpose of applications. When a product reliably saves users time, boosts their efficiency, and resolves their problems, our view is that they’ll pay for it. And as AI applications tackle work that is more intricate and worth more, the readiness to pay should rise alongside it. The revenue opportunity here is substantial. At present, our priority is creating excellent products that draw in more users and bring them back repeatedly. We feel good about the opportunity over the long run, and we anticipate that AI applications will increasingly serve as a key driver of revenue growth for Baidu. Thank you.

Thomas Chong analyst
#15

Hi, good evening. Thanks to management for taking my question. I'd like to ask about capital allocation. How should we view your capex plans and investment priorities for this year and next, including returns on AI investment? What are your latest plans regarding dividends and share repurchase? Thank you.

Henry Haijian He executive
#16

Hello, Thomas. Henry here. In Q3, AI-driven business again made up more than 50% of our revenue, and that share keeps rising. AI Cloud Infra revenue grew nearly 100% versus a year ago. We'll continue putting money into full-stack AI capabilities with a lot of discipline, directing capital toward opportunities where demand is clear and returns look attractive. Regarding how fast we invest, you'll see that capex for Q3 came in 4 billion RMB below Q2, even though our cloud business kept up quite strong momentum. Two factors allow for that. First, we built up substantial computing reserves in advance, so when demand rises we're able to satisfy some of it using capacity we already hold. Second, we continue lifting utilization of the compute capacity we have, which releases additional capability to serve external customers. Looking ahead to Q4 and beyond into next year, we plan to get more out of the compute we already have in order to support growth, and the speed at which we make fresh investments will track customer demand along with the timing of cash collections from those customers. Meanwhile, we remain dedicated to disciplined spending and to improving returns. In our cloud business, three areas hold our focus. We will drive utilization higher so that the capital we have already committed earns stronger returns. Beyond that, we are leveraging our full-stack AI capabilities to grow both revenue and profit from what we offer. This quarter, GPU Cloud, Kunlunxin, and MaaS represented the majority of AI Cloud Infra revenue; those businesses generally come with higher margins, and as you see them scale, we anticipate growth that's more diversified and profitability that's stronger. Third, we're accelerating cash collection. In GPU Cloud particularly, we're seeing more customers prepay and sign long-term contracts, which cuts the cash tied up in procurement and delivery and gives us greater flexibility to back future growth. On foundation models, we'll remain application-driven instead of pursuing SOTA on each benchmark, and we'll keep investing with discipline, guided by genuine business and monetization needs, concentrating on the capabilities that matter most for our applications. Thomas, as you have probably seen, our cash position is quite large and our balance sheet remains very strong. At the close of this quarter, net cash stood at roughly 133 billion RMB, and long-term investments were approximately 43 billion RMB. The business we already operate continues to produce cash. Operating cash flow for the quarter was about 2.8 billion RMB, marking the fifth quarter in a row of positive operating cash flow; together with improved capital efficiency, those cash flows and resources will support further expansion of our cloud and AI business. Thomas, you also pointed out that returning value to shareholders remains a central priority in how we allocate capital. At the start of this year, we introduced our inaugural dividend policy, and we anticipate announcing the initial dividend before the year is out—assuming, naturally, that the board gives its approval and market conditions permit. During this year to date, we have repurchased roughly $373 million worth of stock through our buyback program. We will continue repurchasing shares at an appropriate pace under the authorization already in place, taking into account our eventual financial position and market conditions, and of course we intend to balance long-term growth with shareholder returns. Thank you, Thomas.

Ellie Jiang analyst
#17

Thanks, management, for fielding my questions. Can you bring us up to date on Kunlunxin's IPO plan? How do you view its role within your own AI Cloud infrastructure as that evolves, and what margin benefits would your in-house chip deliver to the AI Cloud over the longer term? Thank you.

Robin Li executive
#18

I'll take that one. Earlier this year, Kunlunxin confidentially filed its listing application with the Hong Kong Stock Exchange. We're continuing to advance Kunlunxin's listing process and will update the market as appropriate. Kunlunxin is the product of 15 years of steady investment in our own AI chips, and it is already running at scale in production settings at major enterprises spanning industries including financial services, telecommunications, energy, and the internet, among others. When selecting AI compute, companies consider more than just peak chip performance—they ask whether chips operate dependably and efficiently on actual workloads, which calls for compatibility with mainstream models and frameworks, straightforward migration and deployment, and stable functioning in large clusters. Kunlunxin has developed robust capabilities in all these areas, and customers are increasingly acknowledging its efficiency, especially for inference workloads; we are also broadening Kunlunxin's product lineup. With the P800 already selling, we anticipate the M100—tuned for inference at large scale—will soon be ready for volume production. Next year we intend to launch the M300, built for very large multimodal models. Between them, these offerings will address a wider spectrum of compute requirements. As agents enter real-world use and inference demand speeds up, we foresee substantial growth potential for Kunlunxin. Looking further ahead, with the domestic semiconductor supply chain maturing and as we continue enhancing our chips and expanding output, we are confident Kunlunxin will take on a growing role in AI Cloud Infra. Kunlunxin also forms a key part of our infrastructure layer. Having our own silicon lets us integrate and tune the full stack — chips, cloud infrastructure, models, and applications — end to end, and that steadily brings costs down. As Kunlunxin deployments grow and this optimization goes deeper, we expect additional room to cut costs. Over time, we think those gains will make AI Cloud Infra more competitive and help margins improve over the long run. Thank you.

Wei Xiong analyst
#19

Good evening, management, and thanks for taking my question. It's encouraging that Baidu general business went back to positive year-on-year revenue growth in the third quarter. How does management see the revenue trajectory across the coming few quarters, and over the longer term? Additionally, might management provide your perspective on how profitability and margins will trend moving ahead, and over the longer horizon too? Many thanks.

Henry Haijian He executive
#20

Thanks. Henry here. During this quarter, Baidu's general business achieved year-over-year revenue growth once again. The primary force behind that was AI Cloud Infra, which, as you've observed, has emerged as a fresh growth engine for Baidu while delivering a rising portion of both our top line and our profit. At this point, AI-driven core operations represent around 53% of the company's general business revenue — a clear signal that the makeup of our earnings has shifted. Over the near term, macro conditions, fierce competition, and the AI Search transformation we are still carrying out will keep substantial pressure on our advertising business. We were among the first companies anywhere to rebuild search around AI, we have pursued that aggressively, and deeper user engagement is already visible. Turning those gains into revenue, though, will take time, and instead of rushing toward monetization we are putting user experience first; that decision will keep weighing on near-term revenue, yet we believe it strengthens our competitive position and supports growth over the long run. AI Cloud Infra, meanwhile, keeps growing strongly, and we expect it to maintain rapid growth across the coming quarters. With that revenue share climbing, we anticipate it will increasingly counterbalance ad-related headwinds and bolster growth overall. Despite the short-term difficulties, we continue to believe our revenue path will strengthen over the medium and long term. Last, regarding profit: near-term pressure on both profit and margin will persist, given ongoing advertising weakness. We'll continue putting money into AI—cloud infrastructure, in-house chips, and foundation models included; those investments matter greatly for our long-term competitiveness, yet they'll limit how much profit can improve in the near term. Even so, this quarter AI Cloud Infra posted year-over-year gains in both profit and margin, and it added a bigger share of our total profit — a sign that its rapid expansion is converting into genuine earnings. Going forward, with higher-margin business accounting for a larger portion of revenue, we anticipate that AI Cloud Infra will play a bigger role in generating profit and will underpin profit growth across the medium to long term. On the whole, we'll keep up the investment that's required while emphasizing efficiency and returns on that investment, so that growth steadily turns into sustainable profit margins down the road. Thank you.

Operator operator
#21

That brings us to the end of today's conference, everyone. We appreciate your taking part. You're welcome to disconnect now.

Read the full transcript via the API

You're viewing the first half of this call. Get the complete Baidu, Inc. transcript - plus 131,000+ transcripts from 7,000+ companies, speaker segments and full-text search - through the EarningsAPI REST API or hosted MCP server.

Get an API key View API docs →

For developers and AI pipelines

Programmatic access to Baidu, Inc. earnings transcripts and 131,000+ others is available through the EarningsAPI REST API and the hosted MCP server. Quarterly plans from $162 - full transcripts, speaker segments, full-text search, and the /api/v1/transcripts/recent polling endpoint for ETL pipelines.