BIDU Baidu, Inc.
$73.94
Baidu, Inc. Q3 F2026 Earnings Call Transcript
AI Conference Call Analysis
Sign in or subscribe to read.Conference Call Operator
Operator
Hello, and thank you for standing by for Beidou's third quarter 2026 earnings conference call. At this time, all participants are in listen-only mode. After management's prepared remarks, there will be a question and answer session. Today's conference is being recorded. If you have any objections, you may disconnect at this time. I would now like to turn the meeting over to your host for today's conference, Jun Lin, Beidou's Director of Investor Relations.
Jun Lin
Director of Investor Relations, Baidu
Hello, everyone, and welcome to Baidu's third quarter 2026 earnings conference call. Baidu's earnings release was distributed earlier today, and you can find a copy on our website as well as on Newswire Services. On the call today, we have Robin Li, our co-founder and CEO, Julius Rong Luo, our EVP in charge of Baidu Mobile Ecosystem Group, MAMIGI, Dou Shen, our EVP in charge of Baidu AI Cloud Group, ACG, and Henry Haijian He, our CFO, After our prepared remarks, we'll hold a Q&A session. Please note that the discussion today will contain forward-looking statements made under the safe harbor provisions of the U.S. Credit Security Litigation Reform Act of 1995. Forward-looking statements are subject to risks and uncertainties that may 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 ICC and Hong Kong Stock Exchange, Baidu does not undertake any obligation to update any forward-looking statements except as required under applicable law. Our earnings press release and this call include discussions of certain annuitated non-GAAP financial measures. Our press release contains a reconciliation of the annuitated non-GAAP measures to the annuitated most directly comparable GAAP measures and 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
Co-Founder and Chief Executive Officer, Baidu
Hello, everyone. In Q3, Baidu General Business returned to year-over-year revenue growth, increasing 1% to RMB 25.0 billion, driven primarily by strong momentum in AI Cloud Infra. Baidu called AI-powered business generated revenue of RMB 13.2 billion, accounting for 53% of Baidu general business revenue and continuing to grow as a share of the total. AI Cloud Infra revenue nearly doubled year over year, with growth accelerating sharply from 50% last quarter to 94%, significantly outpacing the broader market within AI Cloud Infra. GPU Cloud sustained triple-digit revenue growth year-over-year and grew 43% quarter-over-quarter, reflecting robust and sustained demand. Our math business also saw strong momentum, with revenue from external customers' token usage on our Qianfan math platform growing more than 12 times year-over-year in Q3. This quarter's results further demonstrated that our long-term investments in full-stack AI are translating into substantial business growth, with AI Cloud Infra emerging as a key driver of Baidu's overall revenue growth. We believe the strength of our full-stack AI capabilities, from chips and cloud infrastructure to models and applications, positions us well to meet growing demand and capture new opportunities. Looking ahead, we expect AI Cloud Infra to become an increasingly important growth engine for Baidu as we continue to build on this momentum. Now, let's take a closer look at AI Cloud Infra, where we achieved strong revenue growth, improved profitability, broadened our customer base, and secured longer-term contracts this quarter. GPU Cloud was the primary driver of AI Cloud Infra revenue growth. In Q3, GPU cloud revenue grew 191% year-over-year despite a high comparison base and sequential growth accelerated to approximately 43%. This reflects robust demand for scalable AI computing in the public cloud and our ability to meet that demand with reliable and high performance capacity at scale. We are confident that AI Cloud Infra can sustain its strong growth momentum in the coming quarters. Alongside revenue growth, AI Cloud Infra delivered higher profit and margins both year-over-year and sequentially, contributing a meaningfully larger share of Baidu's overall profit this quarter. This improvement was supported by a more favorable revenue mix with higher margin offerings such as GPU Cloud, Kunlunxin, and Mass. now accounting for a majority of AI cloud infra revenue. Their growing share of revenue and continued margin improvements, together with higher resource utilization, are driving overall profit growth and margin expansion. Meanwhile, with demand remaining strong, scale expanding, and profitability steadily improving, we are seeing shorter payback periods on our cloud infrastructure investments. making AI Cloud Infra a healthier, more sustainable business. We see several key factors supporting continued growth in our AI Cloud Infra. First, demand for AI computing continued to accelerate, underpinned by real business needs. Demand was strong across both training and inference workloads. with inference growing particularly fast as AI applications and agents are deployed across a widening range on business use cases. Even so, we believe inference is still in the early stages of scaling with substantial room for further expansion. Second, our customer base expanded significantly in Q3 and became more diversified across industries and company sizes. We continue to win new customers while existing customers increase their usage and spending. Growth remains strong in key sectors such as e-commerce, internet, autonomous driving, and embodied AI. Customers are increasingly embedding AI into their core operations, creating recurring and growing demand for AI computing. For example, a leading e-commerce company uses our cloud infrastructure to train and iterate on multimodal foundation models. that power its retail, logistics, and financial services businesses. Each model upgrade creates new training demand while broader deployment drives ongoing growth in inference usage alongside additional demand for data processing, storage, and network services. In embodied AI, a leading robotics company uses our GPU cloud to train the models behind its robot. Each new task or setting requires fresh data collection, model training, simulation, and testing. As the company expands into more tasks, robot models, and deployment scenarios, its computing needs grow at every stage of development. After deployment, the robots generate ongoing inference demand and feed real-world operating data back into model development. Together, these workflows create sustained demand for GPU computing and related cloud services. Third, we are seeing stronger, longer-term customer commitments. Customers are planning computing needs further ahead and increasingly seeking long-term partnerships to secure stable, reliable, and scalable capacity as they grow. Our industry-leading technology serves excellence and Proven Track Record give them the confidence to entrust us with these needs. In GPU Cloud, in particular, more customers are making prepayments to secure future capacity and committing to longer contract terms. For example, our three largest GPU Cloud customers buy at contract value, all extended to multi-year terms when renewing with us this year. including one whose renewal term was six times longer than its previous term. These stronger commitments give us greater visibility into future demand and support better cash flow. On math, demand for model inference continue to accelerate. Our industry-leading math platform, Xianfan, offers both Ernie and the latest versions of other leading dramatic open source models. giving customers timely access to the most advanced capabilities. In Q3, we further optimize inference to increase throughput and reduce latency while improving resource scheduling to balance workloads and enhance computing utilization. These improvements make model inference on Tianfan more reliable, stable, and cost-efficient. In Q3, Revenue from external customers token usage on Qianfan grew over 12 times year over year and margins continue to improve as math scales and represents a larger share of AI cloud infra revenue. We believe it will become an increasingly meaningful contributor to overall profit growth. Next, let me turn to the model layer. Advancing earning remains a strategic priority. At Baidu World last week, we unveiled Ernie 5.5 Flash Preview, the latest foundation model in our flagship Ernie family. With 280 billion total parameters and only 14 billion activated per token, it delivers strong coding and agentic capabilities, enabling it to orchestrate professional software, tools, and external models to complete complex tasks end-to-end. These capabilities enable users to produce usable, high-quality outputs across knowledge work, game development, 3D modeling, and video creation. On the latest CodeArena web dev leaderboard, Ernie 5.5 Flash Preview ranks second among open-source models globally. Going forward, we plan to accelerate model development across the Ernie family and further enhance its capabilities. Our approach to model development remains application-driven. Each iteration is guided by the needs of our applications. A good example is Ernie Light Deep Thinking Model, a model originally developed for AI search and now also available to customers and partners through our TianfanMath platform. It excels at understanding user intent, reasoning through complex queries, and calling the right tools. while delivering fast responses at scale. Feedback from search helps us continually improve the model, and these improvements help search deliver more intelligent answers and get more done for users. Optimization across our full AI stack further lowers inference costs, enabling us to bring these capabilities to more users. In this way, model innovation translates into better user experiences and greater cost efficiency. Looking at the broader market, demand for model services is growing fast. To make high-quality model services easier to find and use, we launched Token Exchange at Baidu World. It is a neutral, transparent, and trusted marketplace that connects AI model providers with customers. Customers can compare model services side by side with clear pricing and performance data and access them in one place. Quality providers in turn gain a channel to more customers. Over time, we expect token exchange to help put underused compute to work and make the ecosystem around AI and foundation models healthier and more efficient. Turning to AI applications, We continue to expand adoption and strengthen product capabilities. General work agents remain a promising area with growing adoption among both individuals and enterprises. DuMate, our general work agent introduced earlier this year, has grown rapidly. In September, its MAU increased by more than 30-fold quarter-over-quarter. This growth came without large-scale marketing campaigns or integration into an existing super app. Strong next-day retention also points to a positive user experience. Since its enterprise edition launched in June, DouMate has gained strong traction among enterprises, supporting its expansion into a broader range of enterprise and industry-specific workflows. This quarter, we continue to enhance DouMate by integrating capabilities from across Baidu's AI ecosystem, including our internal workplace AI agent Dudu, AI coding assistant Comate, and vibe coding platform Yalda. Bringing these established capabilities together in one agent helps users complete more complex, multi-step tasks more efficiently. Meanwhile, we continue to develop specialized agents in areas where we see strong demand and that our models are optimized for. In August, GenFlow, jointly developed by Baidu Wenku and Baidu Drive, was upgraded and rebranded as Kuku, an AI-powered platform for documents and content creation. Join on the extensive content resources and user ecosystems of Baidu Wenku and Baidu Drive Kuku brings together publicly available information, professional databases, and users' own content to help users research, write, and create high-quality artifacts for work and life. Baidu Yijing is our hyper-realistic digital human agent. Its technology represents the state of the art in the industry, enabling it to consistently match or even exceed human hosts in e-commerce live streaming This quarter, we deployed this technology at scale across our own platform while further improving its cost efficiency and production speed. Since the fourth quarter of 2025, we have cut the production cost for hyper-realistic digital human by approximately 80% and production time by more than an order of magnitude These improvements are making our hyper-realistic digital humans more accessible and scalable. Customer adoption of Yijin continues to grow. A major customer has significantly expanded its use of our digital human capabilities for several consecutive quarters, reflecting its continued confidence in our technology. This quarter, we also broadened our reach across industries adding new customers in financial services, automotive, apparel, and other sectors. Internationally, building on our existing collaborations with TikTok and Shopee, we further expanded our customer base by adding Lazada. As Yijing continues to evolve, we believe it has significant potential to transform how businesses engage with customers across a wide range of industries. We are also applying frontier AI capabilities to complex operational challenges in the real world. At Baidu World, we introduced Phamo 3.0, further strengthening its capabilities and expanding its use in demanding industrial settings. One of our customers, an industrial laser equipment manufacturer, uses Phamo to plan how four robotic arms work together to cut a huge car body panel with complex curves. Each plan must specify how to split the work. The order of cuts and each arms path. Keeping all four arms in sync and avoiding collation. Learning from production data, FAMO autonomously searches through a vast number of possible combinations. Evaluating and iteratively refining plans to find the most efficient one. Designing such a plan used to take 4 to 8 hours. With FAMO, the customer now does it in around 20 minutes. Across our AI applications portfolio, we continue to broaden our offerings, strengthen their capabilities, and improve usability. While monetization remains at an early stage, we see significant potential as these applications become embedded in more workflows and deliver measurable value to users. Turning to AI search, we continue to improve user experience in Q3. AI Answers became more organized, more concise, and easier to navigate, helping users find and understand the information they need more efficiently. We also further deepened the integration between AI Search and Ernie Assistant, improving our ability to understand user intent and respond to more complex needs. Through relevant follow-up questions, Ernie Assistant helps users clarify and explore their needs. Extending the search experience into more natural in-depth conversations and with its agentic capabilities, completing complex tasks seamlessly on users' behalf. In September, early assistance daily average conversation rounds grew 126% year-over-year, while user engagement and stickiness also increased. Looking ahead, we will continue to prioritize user experience over near-term monetization as we advance our AI search transformation. While this will weigh on our online marketing revenue and overall profitability in the near term, we believe it is the right path to strengthen our long-term competitiveness and unlock the full potential of AI search. In the physical world, ApolloGo, our autonomous write-hating service, is bringing smart mobility to more users worldwide. As our technology continues to advance and our operations grow, we are entering a new phase of development with our efforts concentrated on cities with the strongest demand and commercial potential. In these cities, our priority is to deepen market penetration and scale up operations efficiently and sustainably. In Shenzhen, now one of our largest and fastest-growing markets, fully driverless operational rides increased by approximately 170% quarter-over-quarter in Q3, just a few months after we stepped up our efforts there. During the quarter, we expanded our service coverage and added more pickup and drop-off points across the city while improving operating efficiency and the rider experience. These efforts helped us meet more demand and made our service more convenient and accessible. We see significant room for further growth as we continue to deepen our presence in Shenzhen. In Dubai, we continue to expand our fully driverless commercial operations and broaden service areas. Our ApolloGo fleet in the city has doubled, further strengthening our position as Dubai's largest are all fully driverless ride-hailing service providers. With rides available through both the ApolloGo and Uber apps, our service is becoming accessible to more riders and our operating presence in the market continues to grow. ApolloGo also continues to lead the industry in advancing across key right-hand drive, left-hand traffic markets. In London, we progressively expanded our open road testing areas in partnership with Uber and Lyft, gaining more experience with local road, traffic, and operating conditions. In Hong Kong, fully driverless testing progressed steadily on Airport Island, with additional mileage accumulated to support the next stage of development. With a global footprint now spanning 30 cities Our focus is on deepening our presence in priority markets. Safety will remain our top priority as we broaden service coverage, improve fleet utilization, and enhance the rider experience. By building efficient, reliable, and sustainable operations, we aim to support the long-term growth of HoloGO and make it a trusted, integral part of everyday mobility. In summary, this quarter demonstrated meaningful progress in translating our full-stack AI investments into scaled growth, with AI Cloud Infra serving as a key driver. We are also placing increasing emphasis on international expansion, bringing our AI capabilities to users and enterprises globally. We believe the foundation we have built positions us well to capture expanding opportunities ahead. With that, let me turn the call over to Henry to go through the financial results.
Henry Haijian He
Chief Financial Officer, Baidu
Thank you, Robin, and hello, everyone. In Q3, Baidu general business revenue returned to year-over-year growth, increasing by 1%. Baidu core AI-powered business accounted for a growing share of Baidu general business revenue, reaching 53%. AI Cloud Infra was the standout performance this quarter, with revenue up 94% year-over-year. Its rapid growth and rising contribution to both revenue and profit helped return Baidu general business revenue to growth, providing clear evidence that our second growth curve is taking shape. As we build on this momentum, we will maintain a disciplined approach to investment, directing capital toward opportunities with clear demand and attractive expected returns. We are improving the utilization of our existing capacity to support revenue growth. Higher customer prepayments also accelerate cash collection, giving us greater flexibility to invest in our cloud and AI businesses. In addition, we have substantial cash reserves, a strong balance sheet, and continued cash generation from our existing businesses. In Q3, operating cash flow for Baidu remained strong at RMB 2.8 billion, marking our fifth consecutive quarter of positive operating cash flow. These financial resources, together with improved capital efficiency, allow us to grow our cloud and AI businesses rapidly. We will continue to pace new investments in line with demand and cash generation to support sustained profitable growth. On shareholder returns, we appreciate our shareholders' long-term support and remain committed to returning capital to them. As previously communicated, We expect to declare our first cash dividend payment before year-end, subject to board approval and market conditions. We also plan to continue repurchasing shares at an appropriate scale under our existing program. Going forward, we will continue to balance investment in growth with shareholder returns to create long-term value for our shareholders. Now let me walk through the details of our third quarter 2026 financial results. Revenue was RMB 30.8 billion, decreasing 1% year over year and 2% quarter over quarter. Revenue from Baidu General Business was RMB 25.0 billion, increasing 1% year over year and decreasing 1% quarter over quarter. Revenue from IGE was RMB 6.0 billion, decreasing 10% year over year and 4% quarter over quarter. Cost of revenues was RMB 19.9 billion, increasing 9% year-over-year and 4% quarter-over-quarter. The year-over-year increase was primarily due to increases in costs related to cloud business and traffic acquisition costs, partially offset by a decrease in compound costs. Operating expenses were RMB 10.3 billion, decreasing 13% year-over-year, primarily due to decreases in expected credit loss expenses and channel spending expenses, and increasing 12% quarter over quarter, primarily due to an increase in expected credit loss expenses. Operating income was RMB 626 million and operating margin was 2%. Non-GAAP operating income was RMB 1.7 billion and non-GAAP operating margin was 6%. Total other income net was RMB 1.2 billion compared to RMB 1.9 billion for the same period last year and RMB 184 million last quarter. The year-over-year decrease was primarily due to decreases in pickup of earnings from equity method investments and fair value gain from long-term investments and an increase in net foreign exchange loss arising from exchange rate fluctuations between RMB and U.S. dollars. partially offset by a shift from fair value loss to fair value gain on exchangeable bonds. The quarter-over-quarter increase was primarily due to a shift from fair value loss to fair value gain on trading debt investments. Income tax expense was RMB $427 million compared to income tax benefit of RMB $1.8 billion for the same period last year and income tax expense of RMB $1 billion last quarter. Net income attributable to Baidu was RMB 1.4 billion. Net margin for Baidu was 5%, and diluted earnings per ADS was RMB 4.35. Non-GAAP net income attributable to Baidu was RMB 2.4 billion. Non-GAAP net margin for Baidu was 8%, and non-GAAP diluted earnings per ADS was RMB 6.78. We defined total cash and investments as cash, cash equivalents, restricted cash, short-term investments, net, long-term time deposits, and held to maturity investments and adjusted long-term investments. As of September 30, 2026, total cash and investments were RMB 276.6 billion. Operating cash flow was RMB 2.8 billion. Baidu General Business had approximately 26,000 employees as of September 30, 2026. With that operator, let's now open the call to questions.
Conference Call Operator
Operator
Ladies and gentlemen, we will now begin the question and answer session. If you wish to ask a question, please press star 1 on your telephone and wait for your name to be announced. If you wish to cancel your request, please press star 2. Your first question comes from Alicia Yap of Citigroup.
Alicia Yap
Analyst, Citigroup
Hi, thank you. Good evening, management. Thanks for taking my questions. I have a question for AI Cloud Infra. How does management view the AI Cloud Infra trajectory over the next few quarters, and what are the key drivers of that? And also, could you provide more colors on the revenue mix and also discuss how sustainable these growth drivers are over the longer term? Thank you.
Dou Shen
Executive Vice President, Baidu AI Cloud Group
Thank you, Alicia. This is Dou. I will take your question. This quarter, AI Cloud Infra revenue grew 94% year-over-year, once again outpacing the market. GPU Cloud was the main growth driver with strong growth both year-over-year and quarter-over-quarter. We are confident that AI Cloud Infra will sustain its momentum in the coming quarters. That confidence comes not only from strong demand and a growing customer base, but also from improvements in our revenue mix and the customer commitments. First, a demand for both training and inference is very strong and accelerating. Companies are moving beyond experimenting with AI and building it into their core businesses. which makes compute a recurring need. And we are still in the early days of agent adoption. As agents become more capable and easier to deploy, more business will use them in their day-to-day operations. And we believe this will drive sustained demand growth. Second, we are turning growing demand into more and deeper customer relationships. In this quarter, we added customers across more industries and company sizes, and we are doing more with existing ones. One example is a leading autonomous driving chip company. As it iterates on its models, its training needs have grown quickly. Our work together actually started with TPU Cloud, and then expanded into storage, networking and more. And we are now its largest infrastructure provider. This example illustrates how rising AI demand creates opportunities beyond GPU compute. As customers scale their AI workloads, they need a broader range of cloud services. Our full-stack capabilities position us to meet these needs, grow with our customers, and capture more of their cloud spending over time. Third, our revenue mix continues to improve. Depew Cloud, Kunting, and Miles together accounts for more than half of our AI infra revenue this quarter. GPU cloud share rose further, driven by strong demand for AI compute in the public cloud. Mass is still small, but growing rapidly, and we expect it to become a more meaningful growth driver as customer usage expands. A large share of AI cloud infra revenue is also recurring, giving us a more stable revenue base. Going forward, We expect these trends to continue and support healthier, more sustainable groups. Furthermore, we are encouraged to see more customers choosing longer contract terms and making prepayments, particularly in GPU cloud. Many customers extend their contract terms at renewal, and some major customers renew for terms several times longer than their original contracts. Customers now have a clearer view of their long-term compute needs and want to secure reliable capacity in advance. So their experience working with us has also given them confidence in our technology, services, and delivery. So they are willing to commit to longer-term partnerships. These longer contracts and the prepayments significantly improved the visibility and predictability of our future revenue. They also helped us anticipate demand so we can plan capacity and support customers as they expand. Together, these factors give AI Cloud Infra a solid foundation for sustained growth. Thank you.
Miranda Zhang
Analyst, Bank of America Securities
Thank you.
Conference Call Operator
Operator
Your next question comes from the LinkedIn call with Goldman Sachs.
Lincoln Kong
Analyst, Goldman Sachs
Thank you, management, for taking my question. So I want to ask about model side. Could you update us on earnings roadmap and positioning in China's AI landscape? And how did you see this grow across Baidu's business? And how are we thinking about the return on our continuing investment here? Thank you very much.
Robin Li
Co-Founder and Chief Executive Officer, Baidu
Hi, Lincoln. This is Robin. Foundation models are a critical layer of our full-stack AI architecture, and we will continue to invest in learning and push the frontier of intelligence, which we see as essential to Baidu's long-term competitiveness. We just launched Earning 5.5 Flash Preview, the latest foundation model in our Earning family. It delivers strong coding and authentic capabilities across knowledge work, game development, 3D modeling, and video creation. On the latest CodeArena WebDive leaderboard, Ernie 5.5 Flash Preview ranks second among open source models globally. and remember our model is only about one-tenth 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 cover a broad range of capabilities But because we run so many applications, we see firsthand what users need and which model improvements will make the biggest difference. In the areas that matter most to our applications, we want our models to be the best. Take our hyper-realistic digital human agent Yijing, for example. Powered by Ernie, our digital humans now support really nuanced, natural-looking visuals. Even the tiniest facial expression and the slightest shift in gaze come across as lifelike. We believe these capabilities are among the best in the industry. Another example is Ernie Light's Deep Thinking, a model we built especially for AI search. It delivers industry-leading performance in coding, reasoning, and identity tasks. That means it can really help search understand complex user intent and get tasks done while remaining fast and cost-efficient to run. These are exactly the capabilities our applications need most and that's where we want to be the strongest. On investment, we all keep investing and we all do it in a disciplined way. Focusing resources on the areas that matter most and continually improving training efficiently. Over the long term, a stronger earning benefits us in two ways. First, better in-house models enable our applications to deliver better user experience at lower cost, supporting monetization and profitability. Second, through our ChampaMath platform, we bring earnings to more external customers as they build AI applications. At this application scale, we expect to drive sustained demand for our model services, further unlocking the commercial value of our full-stack AI capabilities and diversifying AI cloud infrastructure revenue streams. Thank you.
Conference Call Operator
Operator
Once again, if you wish to ask a question, please press star 1 on your telephone and wait for your name to be announced. Your next question comes from Alex Yao, JP Morgan.
Alex Yao
Analyst, J.P. Morgan
Thank you, management, for taking my question. So my question is about the cloud margin. As your AI cloud info business continues to scale, how should we think about the margin outlook, and what are the key drivers of the margin? A follow-up question to that is with the continued investment in compute infrastructure, how do you guys evaluate the returns on these investments? Thank you.
Dou Shen
Executive Vice President, Baidu AI Cloud Group
Thank you, Alex. In Q3, AI Cloud Infra improved both profit and margin year-over-year and quarter-over-quarter. becoming a meaningful contributor to buy this overall profitability. This is an encouraging sign that the business is getting healthier, even as it scales rapidly. Over the longer term, we see room for further margin expansion and improved overall profitability, driven by four factors. First, as I just mentioned, our value mix continues to improve. Higher margin businesses such as GPU Cloud, Kunming, and MESS are contributing a growing share of revenue, and each is becoming more profitable as it scales. So the upside comes from both the next shift and improving profitability within each business. Second, greater scale helps us improve resource utilization. As our business grows and workloads become more diverse, we can pool capacity across different workloads, improve scheduling, and balance demand peaks and crops. This allows the same infrastructure to serve more demand, lowering unit costs. Third, our full-stack AI capabilities help reduce costs through end-to-end optimization. At the infrastructure level, Kunmingxin gives us greater flexibility to optimize infrastructure cost and performance. At the math layer, continuous inference improvements enable the same computing capacity to process more tokens with lower latency while maintaining reliability. Fourth, our first tank, also positions us to deliver higher-value services. Beyond cloud infrastructure and model services, our stack includes a wide range of agents, such as Jumeis, Phamo, and Yijin, just mentioned by Robin. This also allows us to meet customer needs at every layer with more comprehensive products, services, and solutions. As adoption grows, we expect this to create more value for customers and expand the contribution from their value services. Alongside profitability, we also pay close attention to the net cash investment needed to support growth. When we evaluate a customer or contract, we look at expected revenue and margins. and also at payment terms and the expected payback period. Our goal is to recover our upfront compute investment faster and use cash received from customers to fund the next fund of investment. Among new contracts signed this year, the share with prepayments is also rising quickly. The demand grows Our business skills and margins improved. We're seeing shorter payback periods for our cloud infrastructure investments. With more prepayments and faster payback, we can reinvest sooner and then scale faster. So taken together, these factors give us confidence in the long-term upside in AI Cloud Infra's margin and profitability. Thank you.
Conference Call Operator
Operator
Your next question comes from Miranda Zong with Bank of America Securities.
Miranda Zhang
Analyst, Bank of America Securities
Thank you. Good evening, management. Thanks for taking my question. My question is about the AI applications. How does management see the competition in the AI application agents evolving? How are companies' AI applications such as stillmates performing? and how should we think about the monetization model and the longer-term revenue potential of these products? Thank you.
Robin Li
Co-Founder and Chief Executive Officer, Baidu
Okay, first, I think, ultimately, the value of AI has to come from applications. Without applications, the models or cloud infrastructure or chips are just useless. and AI application market is also moving very quickly. As model capabilities improve, AI applications and agents can take on increasingly complex tasks for more users across a broader range of scenarios. And new agentic products are emerging at an accelerating pace. However the market evolves, We believe long-term competitiveness comes down to the value an application creates, whether it can complete tasks efficiently and reliably across a wide range of real-world applications and deliver measurable productivity gains for users and meaningful business outcomes for enterprises. That's why our approach to innovation at Baidu has consistently been application-driven, starting from real-world use cases. Today, we've built out a portfolio of AI applications and agents that spans individual and enterprise customers, covering a wide range of tasks and scenarios. It includes our general work agent in DuMate, as well as specialized agents such as our digital human agent Yijin, our self-evolving agent Fimo, and Kuku for creating AI artifacts. Take DuMate as an example. Its user base is growing very fast, largely organically, without heavy marketing spend or traffic from are other products. These are encouraging signals that users recognize how well the product actually works. On monetization, AI-native applications are still at very early stage, and the whole industry is still exploring. Different products serve different users, handle different tasks, and create different kinds of value. So monetization models will naturally vary. But one thing is clear to us. Users pay for value. And creating that value is what applications are fundamentally about. If a product can consistently help users save time, work more efficiently, and solve problems, we believe they will be willing to pay for it. As AI applications take on more complex, fire-bellied work, users' willingness to pay will grow with it. The monetization upside is significant. So our focus right now is on building great products that more users choose and keep coming back to. We are optimistic about the long-term opportunity and we expect AI applications to become an increasingly important revenue growth driver provided. Thank you.
Conference Call Operator
Operator
Once again, if you wish to ask a question, please press star one on your telephone and wait for your name to be announced. Your next question comes from Thomas Chong with Jefferies.
Thomas Chong
Analyst, Jefferies
Hi, good evening. Thanks, management, for taking my question. My question is about capital allocation. How should we think about your CapEx plans and investment priorities for this year and the next, including returns on AI investment? And what are your latest plans for dividends and share repurchase? Thank you.
Henry Haijian He
Chief Financial Officer, Baidu
Hello, Thomas. This is Henry. In Q3, AI-powered business again contributed more than half of our revenue, with the share still rising. And AI cloud infra revenue grew nearly 100% year-over-year. We will keep investing in full-stack AI capabilities with very strict discipline, putting capital into opportunities with a clear demand and attractive returns. On the pace of the investment, as you'll notice that the CapEx in Q3 was 4 billion RMB lower compared with Q2, while our cloud business still maintained very strong momentum. So two things make this possible. First of all, we have built ample computing reserves earlier. So as the demand grows, we can meet part of it with the capacity we already have. Second, we keep raising the utilization of the existing compute capacity. which frees up more capability to serve external customers. Looking ahead to Q4 next year, we will make further usage of our existing compute to support growth and pace new investment in line with the customer demand and the cash collection cycles from the customers. Meanwhile, we remain focused on spending wisely and improving returns. In our cloud business, we are working on three areas. First of all, the utilization will be further extended to get better returns on what we have already invested. Second, using our full-stack AI capability to broaden our resource on revenue and profit. This quarter, the GPO Cloud, Kunlunxin, and MaaS made up majority of the AI Cloud infra revenue. Those business typically carry higher margins. So when you see they can scale, we expect more diversified growth and a stronger profitability. Third, speeding up the cash collection. In a GPU cloud especially, more customers were observing a prepaying and assigning long-term contracts, which reduced the cash tied up in the procurement and the delivery, giving us more flexibility to support future growth. On foundation models, we will stay application-driven rather than chasing thought on every benchmark, so we will keep investing with the discipline. guided by the real business and monetization needs and a focus on capabilities that matter most to our applications. So Thomas, as you probably noticed that we have very substantial cash reserves and a very strong balance sheet. Right now we have the net cash about 133 billion RMB and a long-term investment about 43 billion RMB by end of this quarter. Our existing business continue to generate cash. Our operating cash flow for this quarter is about 2.8 billion RMB, making it is the fifth consecutive quarter of positive operating cash flow. These cash flows and resources, together with improving capital efficiency, will support continued growth in our cloud and AI business. You also point out, Thomas, shareholder return remain a key capital allocation priority. So earlier this year, we introduced our first dividend policy, and we expect to declare the first dividend before the year end, and subject, of course, subject to the board approval and the market conditions. Year to date, we have repurchased approximately about 373 million US dollars on share buyback of the shares. We will continue to repurchase at appropriate scale on our existing authorization, taking into account our final financial position and market conditions while our intention is to balancing the long-term growth with the shareholder return of call. Thank you, Tom.
Conference Call Operator
Operator
Your next question comes from Ellie Zhang with Macquarie.
Ellie Zhang
Analyst, Macquarie
Thank you, management, for taking my questions. Could you update us on Kunlun Xing's IPO plan? How do you see its role within your own AI cloud infrastructure evolving, and what margin benefits would your in-house chip bring to the AI cloud over the longer term? Thank you.
Dou Shen
Executive Vice President, Baidu AI Cloud Group
Okay, I'll take this question.
Robin Li
Co-Founder and Chief Executive Officer, Baidu
Earlier this year, Kunlunxin confidentially submitted its listing application with the Hong Kong Stock Exchange. We continue to move forward with Kunlunxin's listing process and will update to the market as appropriate. Kunlunxin reflects 15 years of sustained investment in our own AI chips, and it is already deployed at scale in production environments at leading enterprises in sectors like financial services, telecommunications, energy, internet, etc., etc. When choosing AI compute, enterprises look beyond peak chip performance to whether the chips run reliably and efficiently on real workloads. That means compatibility with mainstream models and frameworks, easy migration and deployment, and stable operation in large clusters. Kunlunxin has built strong capabilities in all of these areas, and customers increasingly recognize its efficiency, especially in inference workloads. We're also expanding Kunlunxin's product portfolio Building on the P800, which is already on the market, we expect the M100, optimized for large-scale inference, to be ready for volume production soon. And next year, we plan to introduce the M300, designed for ultra-large multimodal models. Together, these products will cover a broader range of compute needs. As agents move into real-world applications and inference demand accelerates, we see strong growth opportunities for Kunlunxin. Over the longer term, as the domestic semiconductor supply chain matures and we continue to improve our chips, and scale-up production, we are confident that quantum sensing will play an increasingly important role in AI-accelerated infra. Quantum sensing is also a key component of our infrastructure layer. With our own chips, we can better integrate and optimize end-to-end across chips, cloud infrastructure, models and applications. steadily lowering costs. As quarantine deployment scale and this optimization depends, we see further room for cost production. Over time, we believe these gains will strengthen AI Cloud infrastructure competitiveness and support long-term margin implement. Thank you.
Conference Call Operator
Operator
Once again, if you wish to ask a question, please press star one on your telephone and wait for your name to be announced. Your next question comes from Huizhong with UBS.
Hui Zhong
Analyst, UBS
Good evening, management. Thank you for taking my question. It's encouraging to see Baidu General Business return to positive revenue growth in third quarter. So how does management view the revenue trajectory over the next few quarters and over the longer term? Also, could Majima share your outlook for profit and margin trend going forward and over the longer term as well? Thank you.
Henry Haijian He
Chief Financial Officer, Baidu
Thank you. This is Henry. For this quarter, Baidu General Business returned to a year-over-year growth on the revenue side. The main driver was AI Cloud Infra, as you noticed. which has become a new growth engine for Baidu and is building an increasing share of our revenue and profit. Baidu's core AI-powered business now makes up 53% of Baidu's general business revenue, making a clear shift in our revenue mix. In the near term, our advertising business will remain under considerable pressure from micro-conditions, intense competition, and our ongoing AI search transformation. As one of the first companies globally to rebuild the search with AI, we have taken an aggressive approach and we are already seeing deeper user engagement. However, translating these gains into revenue will take time and we are prioritizing user experience rather than rushing to monetization. This choice will continue to weigh on near-term revenue But we believe it will strengthen our competitive positions and support long-term growth. Meanwhile, AI Cloud Infra continues to grow strongly, and we expect it to sustain rapid growth in the coming quarters. As its share of the revenue rise, we also expect it to increasingly offset pressures in advertising and support overall growth. Despite near-term challenges, we remain confident our revenue trajectory will improve over The median to long term. Lastly, on a profit, our profit and margin will remain under pressure in the near term in the midst of the continued softness in advertising. We will keep investing in AI, including cloud infrastructures, in-house chips, and foundation models. These investments are critical to our long term competitiveness, but they will limit our near term profit improvement. That said, AI Cloud Infra improved both profit and margin year-over-year this quarter and contributed more to our overall profit, showing that its strong growth is translating into real profit. Looking ahead, as higher margin business made up a larger share of revenue, we expect AI Cloud Infra to become a more important profit contributor and support media to long-term profit growth. Overall, we will maintain the necessary investment while focusing on investment efficiency and returns so that growth gradually translating into sustainable profit margins going forward. Thank you.
Conference Call Operator
Operator
Ladies and gentlemen, that does conclude our conference for today. Thank you for participating. You may all disconnect.