百度 (BIDU) 2026 年第二季法說會：GPU 雲端營收激增 283%

重點總覽

• 百度公布 2026 年第二季營收為 313 億元人民幣，年減 4%、季減 2%。歸屬於百度的淨利潤為 23 億元人民幣，稀釋後每股 ADS 盈餘為 5.74 元人民幣。
• 百度核心 AI 驅動業務貢獻營收 125 億元人民幣，占百度一般業務營收的一半。AI 雲端基礎設施營收年增約 50%。
• GPU 雲端為主要成長動力，繼第一季成長 184% 後，營收年增達 283%。管理層表示，AI 雲端基礎設施下半年可望維持強勁成長，並有進一步加速的潛力。
• 非 GAAP 營業利益為 38 億元人民幣，營業利益率為 12%。在 GPU 雲端營收占比提升的支撐下，AI 雲端基礎設施的利潤與利潤率均較去年同期改善。
• 隨著用戶注意力競爭加劇，加上百度有意限制 AI 搜尋的變現速度，線上行銷業務持續承壓。管理層預計此壓力將在下半年持續。
• Apollo Go（蘿蔔快跑）在本季完成約 100 萬次全無人駕駛營運訂單。截至 2026 年 6 月，累計向公眾提供服務訂單量突破 2,300 萬次，同時在杜拜、香港與倫敦推進國際商業化進程。

關鍵財務數據

營業費用下降主因是預期信用減損降低、管道支出減少以及研發人員相關費用減少。其他收益總額由去年同期的 49 億元人民幣降至 1.84 億元人民幣，反映出長期投資的公允價值收益減少以及淨匯兌損失增加。

業務與營運表現

AI 雲端與昆侖芯

在訓練與推理工作負載需求的支撐下，AI 雲端基礎設施營收年增 50%。線上遊戲、電商與生活內容領域的既有客戶增加了使用量與支出，同時百度也在網路、具身 AI、自動駕駛、智慧型手機及金融服務等領域拓展了新客戶。

嵌入式 AI 營收年增約六倍。GPU 雲端營收成長 283%，連續第四個季度實現三位數成長。管理層表示，相較於傳統 CPU 雲端，GPU 雲端具備更具吸引力的利潤空間，且在營收結構中的占比持續擴大。

千帆 MaaS 平台上外部客戶的 Token 使用營收年增超過九倍。百度將此成長主要歸因於日均 Token 消費量的提升。

昆侖芯的需求依然廣泛。百度擴大了對最新版本中國基礎大模型的相容性，包括 Kimi K3、GLM5.2、Minimax M3 和混元 3。公司正在推進其產品路線圖，包括用於大規模推理的 M100 以及即將推出的 M300s。管理層表示，昆侖芯擬議的上市進程仍在持續進行中。

AI 應用與搜尋

在推出允許用戶透過自然語言指令創建獨立 Android 與 iOS 應用程式的秒噠 3.0 後，秒噠 6 月的月活躍用戶數較 3 月成長了 67%。

6 月百度文庫與百度網盤的 AI 日活躍用戶滲透率年增 27.4%。文心助手日活躍用戶成長 83%，而日均對話輪數增加超過兩倍。

百度表示，AI 搜尋答案的可靠性與結構化表現均有提升，同時幻覺率保持在低位。然而，公司將產品品質置於短期變現之上。來自 AI 聊天機器人的競爭以及資訊消費習慣的改變，繼續對廣告業務構成壓力。

Apollo Go 自動駕駛計程車

Apollo Go（蘿蔔快跑）在第二季提供約 100 萬次全無人駕駛營運訂單。截至 2026 年 6 月，累計向公眾提供的服務訂單已超過 2,300 萬次。

季度搭乘量暫時受到部分中國國內城市監管相關營運調整的影響。據管理層稱，受影響地區的營運已於 8 月開始恢復。截至 6 月底，Apollo Go 全無人駕駛車隊平均每 1,440 萬公里僅發生一次安全氣囊彈出。

在杜拜，Apollo Go 於 7 月開展全無人駕駛商業化營運，用戶可透過 Apollo Go 與 Uber 預約行程。6 月獲得香港首個全無人駕駛測試牌照，並於 7 月在機場島啟動測試。7 月亦與 Uber 及 Lyft 合作，在倫敦展開開放道路測試。

管理層展望

管理層預計 AI 雲端基礎設施將在 2026 年下半年保持強勁成長，且有進一步加速的潛力。該展望基於強勁的算力需求、持續擴大的客戶儲備以及推理應用的普及。

管理層亦認為 AI 雲端基礎設施利潤率具備長期提升空間。驅動因素包括 GPU 雲端占比提高、資源利用率提升、單位推理成本降低、營運槓桿效應，以及昆侖芯與百度全棧 AI 架構帶來的成本優勢。

在線上行銷方面，由於用戶注意力競爭激烈，加上百度有意克制 AI 搜尋的變現，管理層預計下半年壓力仍將持續。

隨著中國國內營運恢復及向新市場擴展，Apollo Go 預計未來數季搭乘量將重拾成長動能。其優先事項包括提高安全與營運標準、擴大車隊與搭乘量規模、推進國際拓展，以及使更多城市達到單車經濟學 (Unit Economics) 收支平衡。

在取得股東批准、香港聯交所核准及滿足其他適用條件的前提下，百度預計其香港上市地位轉換為雙重主要上市將於 2026 年內生效。公司亦在為未來納入港股通做準備，這仍取決於資格要求及交易所決定。

風險與觀察重點

• 總營收年減 4%，百度一般業務營收亦下滑 4%。
• 線上行銷面臨來自用戶行為轉變、AI 聊天機器人競爭以及 AI 搜尋變現延後帶來的持續壓力。
• 儘管訓練與推理需求持續上升，整個市場的 AI 算力供應依然緊張。
• Apollo Go 搭乘量暫時受到部分中國城市監管相關營運調整的影響。
• 百度仍處於龐大的 AI 投資週期中。管理層坦承，不同投資的回收期有所差異，部分投資需要更長時間才能展現完整價值。
• 香港雙重主要上市轉換及潛在的納入港股通，仍取決於監管機構、交易所及股東的批准。

分析師問答亮點

文心大模型定位：管理層表示，百度將繼續投資文心大模型（ERNIE），並採用應用驅動的開發方法。優先發展的能力將支援 AI 搜尋、數位人、秒噠、Famou Agent 與 DuMate。公司的目標是加速模型迭代，讓文心大模型重返基礎大模型的第一梯隊。

雲端盈利能力：管理層預計 GPU 雲端貢獻的提升將支撐利潤率，因為其利潤率空間優於 CPU 雲端。隨著使用規模擴大，千帆 MaaS 與昆侖芯將可提供額外的長期利潤與成本效益。

平衡投資與盈利能力：百度表示將繼續果斷投資 AI，同時在投資資本報酬率、營運效率及現金流方面保持紀律。管理層預期，隨著規模擴大與變現日趨成熟，這些投資將隨著時間推移轉化為更具永續性的利潤成長。

自動駕駛計程車擴展：百度並不認為國內與國際市場是相互排斥的。進入市場將取決於法規、出行需求、搭乘定價、路況及商業可行性。管理層相信，海外市場較高的搭乘價格，結合較低成本的車輛與經驗證的營運模式，能支撐更強勁的單車經濟效益。

法說會完整逐字稿

完整財報電話會議逐字稿

管理層陳述

Operator

Hello and thank you for standing by for Baidu's Second Quarter 2026 Earnings Conference Call. [Operator Instructions] Today's conference is being recorded.

[Operator Instructions] I would now like to turn the meeting over to your host for today's conference, Juan Lin, Baidu's Director of Investor Relations.

Juan Lin

Hello, everyone, and welcome to Baidu's Second 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, MEG; Dou Shen, our EVP in charge of Baidu AI Cloud Group, ACG; and Henry Haijian He, our CFO. After our prepared remarks, we will 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. Private Securities 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 the SEC and the 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 unaudited non-GAAP financial measures. Our press release contains a reconciliation of the unaudited non-GAAP measures to the unaudited 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.

Yanhong Li

Hello, everyone. In Q2, Baidu General Business generated total revenue of RMB 25.2 billion, with Baidu Core AI-powered Business continuing to represent half of the total, reinforcing AI's position at the core of our business. AI Cloud Infra delivered another quarter of strong growth with overall revenue increasing 50% year-over-year, once again outpacing the broader market. Within AI Cloud Infra, GPU Cloud revenue nearly quadrupled year-over-year, growing 283% and accelerating significantly from an already strong 184% growth rate last quarter. With AI-powered business now at the core of our revenue mix, we are focused on building a stronger foundation for its next phase of growth across our full AI stack from chips and cloud infrastructure to models and applications. We are continuing to strengthen the capabilities that will support sustained innovation, power our future growth and reinforce our long-term competitiveness.

Let me now turn to the key business highlights of this quarter, starting with our proprietary AI chips, Kunlunxin. In Q2, Kunlunxin continued to demonstrate strong business momentum with demand remaining robust and broadening across industries. A growing number of customers are adopting its chips for an expanding range of AI workloads, reflecting increasing market recognition of Kunlunxin's stability, efficiency and versatility at scale. Kunlunxin continued to strengthen its software ecosystem, broadening compatibility with leading models and frameworks and improving ease of deployment across enterprise environments. Building on its support for ERNIE and other leading foundation models in China, Kunlunxin further extended its coverage in Q2 to include newer versions of major Chinese foundation models such as Kimi K3, GLM5.2, Minimax M3 and Hunyuan 3. It also improved inference throughput and overall compute efficiency, strengthening its ability to support diverse and demanding AI workloads at scale.

Over more than a decade, Kunlunxin has successfully developed and commercialized three generations of AI chips. Building on this track record, it continued to advance a clearly defined product road map, including the latest M100 optimized for large-scale inference and the upcoming M300s. This road map reflects Kunlunxin's deep understanding of evolving AI technology, workloads and their compute requirements, positioning it to support the next wave of AI innovation. As we continue to advance our AI infrastructure capabilities, we believe Kunlunxin will play an increasingly important role within our full stack AI architecture and enhance our ability to deliver high-performance, reliable and cost-efficient AI computing at scale. As demand for AI computing in China continues to grow, we believe our proprietary AI chips and full stack capabilities will become increasingly valuable. Supporting the future growth of our AI businesses and reinforcing our long-term competitiveness in AI.

Building on our strength at the infrastructure layer, AI Cloud Infra delivered another quarter of strong growth. In Q2, AI Cloud Infra revenue increased by 50% year-over-year, continuing to outpace the broader industry. Several factors combined to drive this sustained growth momentum. First, AI Cloud Infra continued to benefit from strong demand for AI computing. Demand remained robust across both training and inference workloads, while computing supply remains constrained across the market. Second, our existing key clients, including leading companies in online gaming, e-commerce and lifestyle content continue to increase both their usage and spending with us. Meanwhile, our overall customer count grew rapidly with new clients spanning companies at varying sizes.

Third, demand remained broad-based across industry verticals, including Internet, embedded AI, autonomous driving, smartphones, financial services and more. Within this mix, Internet and autonomous driving sustained strong growth, while embedded AI revenue grew approximately sixfold year-over-year in Q2. Based on these trends, we believe AI Cloud Infra revenue growth will remain strong in the second half with the potential for further acceleration. Importantly, the growth in AI Cloud Infra was accompanied by rapid profit growth and expanding margins on a year-over-year basis, reflecting continued improvement in the overall health and quality of the business. Within AI Cloud Infra, GPU Cloud revenue growth accelerated sharply to 283% year-over-year, building on an already high base of 184% growth last quarter. This momentum reflects strong underlying demand for scalable AI compute in the public cloud. The mix of our business continued to shift towards higher-quality revenue streams with GPU Cloud accounting for a growing share of AI Cloud Infra revenue. Given its more attractive margin profile, this shift is contributing to a healthier revenue mix and strengthening the long-term profitability of our cloud business.

On MaaS, our Qianfan MaaS platform offers one of the most comprehensive model libraries covering Baidu's ERNIE family as well as virtually all of China's leading models. A key priority for Qianfan is to make model inference at scale more reliable and cost efficient for customers. Leveraging our deep expertise in AI infrastructure and engineering, we further enhanced model serving through continued inference optimization, delivering higher throughput and greater service stability while reducing latency and inference costs. In Q2, revenue from external customers' token usage on Qianfan grew more than ninefold year-over-year, primarily driven by rapid growth in daily average token consumption among these customers.

Turning to foundation models. Advancing ERNIE and our overall model capabilities remains important to our next phase of AI-driven growth. Our commitment to foundation model innovation remains unwavering. As discussed in prior quarters, we reorganized our model teams into two groups with clearer mandates and greater focus across foundation models and applications. More recently, we welcomed a new generation of top AI talent to work on foundation models, further demonstrating our determination to compete and innovate at the forefront of AI. We believe these efforts will support the continued evolution of ERNIE and strengthen the foundation for future innovation across both models and AI applications.

Moving next to AI applications, where we continue to enhance product capabilities and expand real-world use cases. Let me begin with digital humans. As our digital human technology continues to advance, it is delivering stronger performance at lower cost and enabling an expanding range of use cases from e-commerce live streaming and digital human videos to real-time interactive digital humans and our newly introduced video podcast. These advances are opening up far broader possibilities for how digital humans can be used across industries. Our digital human capabilities are gaining increasing recognition from clients. In Q2, we continue to win new clients, including leading companies across industries, while existing clients also meaningfully scaled their usage. Some of our clients started with a pilot and after seeing what our digital human technology could deliver, expanded their usage.

A well-known Chinese Internet company, for example, expanded its digital human live streaming deployment to approximately 2.5x the previous level after just one quarter of use. Meanwhile, we continue to advance the global expansion of our digital human capabilities. Since launching our overseas digital human platform last quarter, we've seen encouraging momentum with its differentiated capabilities, delivering compelling results for merchants and creators overseas. As demand continues to unfold across more industries and regions, we believe the long-term growth potential for digital humans remains substantial.

Turning next to Miaoda, our vibe coding platform. With the launch of Miaoda 3.0 last quarter, users can now generate stand-alone mobile apps for both Android and iOS using natural language. Applications that once required a professional development team, a lengthy development cycle and significant investment can now be completed far more easily through Miaoda, even directly from a phone. We are seeing users engage with Miaoda more deeply. An increasing number of users are moving beyond one-off experimentation and returning to Miaoda to continue developing, iterating on, and refining their applications over time, reflecting stronger user stickiness. In June, Miaoda's monthly active users increased by 67% compared with March. Adoption is also expanding across industries, ranging from technology and education to health care, manufacturing, financial services and logistics, demonstrating Miaoda's applicability across diverse business scenarios and its broader commercialization potential.

We are also applying AI to help enterprises solve complex operational problems. A good example is Famou Agent, which can autonomously explore possible solutions to identify the best ones. Following the launch of Famou Agent 2.0 last quarter, we have continued to improve its usability and expand the scenarios it can address. Famou Agent has attracted growing interest from leading enterprises and begun to gain early commercial traction this quarter. We are pleased to see Famou Agent moving beyond efficiency gains to help enterprises optimize their operations and deliver real tangible business value. As its capabilities continue to advance, we believe its potential will continue to grow.

Another key direction for our AI applications is general-purpose agents. Earlier this year, we launched DuMate, our general-purpose agent for everyday productivity with seamless access across PC and mobile. In Q2, we introduced an enterprise version and continued to expand DuMate's proprietary Baidu skills and specialized toolkits, broadening the range and sophistication of tasks it can support. Meanwhile, our flagship consumer-facing AI applications, Baidu Wenku and Baidu Drive continue to embrace AI across the board, introducing new AI capabilities, sharpening existing ones. And this quarter, rolling out an upgrade to GenFlow that brings AI more deeply into users' everyday workflows. In June, AI DAU penetration across Baidu Wenku and Baidu Drive increased by 27.4% year-over-year, reflecting broader adoption of their AI-powered features.

Turning to AI search. We continue to improve both the quality of AI-generated answers and the overall user experience. Users are increasingly receiving answers that are more reliable, better structured and more effectively presented. At the same time, hallucination rates remained low, while our models became more effective at assessing content quality, helping reduce the incidence of low-quality answers. Together, these improvements drove better user experience and higher user satisfaction. We also further integrated AI search with ERNIE Assistant, extending the search experience beyond onetime answers into more seamless and interactive conversations that can better address users' follow-up questions and broader needs. In June, ERNIE Assistant's daily active users grew 83% year-over-year, while daily average conversation rounds more than tripled, reflecting growing user adoption and deeper engagement with this evolving search experience.

Turning now to AI in the physical world. Let me discuss Apollo Go, our autonomous ride-hailing service. This quarter, we continued to advance global expansion while further enhancing safety, operational performance and the rider experience. Hong Kong marked an important milestone for Apollo Go this quarter. In June, we received Hong Kong's first permits for fully driverless testing and began testing on Airport Island in July. This made Apollo Go the first autonomous ride-hailing service provider globally to conduct fully driverless testing in a right-hand drive, left-hand traffic robotaxi market. Hong Kong is one of the world's most sophisticated urban mobility markets with a complex operating environment and rigorous standards for both technology and operations. Reaching this milestone in Hong Kong provides strong validation of the maturity and adaptability of our technology and operational capabilities.

The experience we have gained in Hong Kong is already helping us advance more efficiently in London. In July, Apollo Go began open road testing there in partnership with Uber and Lyft. Together, our progress in these two markets demonstrates our technology's ability to generalize across different operating environments, giving us greater confidence in expanding into more and more high-value right-hand drive, left-hand traffic robotaxi markets over time. We also made progress across several other international markets. In Dubai, we launched a fully driverless commercial operations in July and now operate at the largest scale among fully driverless autonomous ride-hailing services in the city with rides available through both the Apollo Go and Uber apps. In Switzerland, we began open-road testing in partnership with PostBus. We also signed a memorandum of understanding with Kazakhstan's Turlov Private Holding Limited to jointly explore autonomous ride-having services in the country.

Overall, Apollo Go delivered around 1 million fully driverless operational rides in Q2. As of June 2026, cumulative rides provided to the public by Apollo Go exceeded 23 million. Ride volume during the quarter was temporarily affected by operational adjustments in certain domestic cities due to regulatory considerations. Over this period, we conducted a systematic review to further strengthen the robustness of our autonomous driving systems and the rigor of our operational processes. As of August, operations in the affected cities have begun to resume on a stronger footing. Meanwhile, we continue to expand our operations across other domestic markets. We are confident that ride volume will regain momentum over the coming quarters as we steadily ramp up operations and pursue further expansion.

In Q2, we continue to raise the bar on safety and the rider experience. As of the end of June, our fully driverless vehicles recorded an average of approximately one airbag deployment, every 14.4 million kilometers, underscoring our industry-leading safety performance. We also enhanced pickup and drop-off point recommendations to reduce walking distances and avoid unsuitable stopping locations while further improving perception and motion planning capabilities to deliver smoother and more consistent rides. These improvements represent an even higher operating standard, one we intend to build on as we continue to integrate Apollo Go more seamlessly into urban transportation systems, making it a more convenient and trusted part of everyday mobility.

Looking ahead to the second half, our priorities for Apollo Go are clear: further enhance our safety standards and operational capabilities, advance our global expansion, scale our fleet and ride volumes, and bring more cities to unit economics breakeven. We believe progress across these priorities will further strengthen Apollo Go's leadership in autonomous ride hailing and lay a stronger foundation for scaling its operations safely and sustainably over the long term.

To summarize, the progress we made across our full AI stack this quarter reaffirms Baidu's transition into an AI-first company and further strengthened the foundation for our next phase of growth. We are also actively expanding our AI businesses into global markets and are encouraged by the progress we are already seeing, including in AI applications and robotaxi. With this stronger foundation, we believe we are well positioned to capture a broader range of opportunities across markets over time.

With that, let me turn the call over to Henry to go through the financial results.

Haijian He

Thank you, Robin, and hello, everyone. We were pleased with the continued momentum of Baidu Core AI-powered Business this quarter. Revenue from Baidu Core AI-powered Business reached RMB 12.5 billion and continue to account for half of Baidu General Business revenue. AI Cloud Infra revenue grew approximately 50% year-over-year, with GPU Cloud revenue growing 283% year-over-year in Q2. This performance reinforces AI-powered business as a key driver of Baidu's long-term growth. In addition, we advanced an important capital market initiative to broaden access to capital markets and create long-term shareholder value.

In July, our Board approved a motion to pursue the voluntary conversion of our Hong Kong listing to dual primary status. Since then, we have submitted our application and received Hong Kong Stock Exchange's acknowledgment. We also plan to convene an Extraordinary General Meeting on August 26 to seek shareholder approval for certain related matters. We currently expect the conversion to become effective within this year, subject to the approval of the shareholders and the Hong Kong Stock Exchange. Once effective, we believe the dual primary listing will broaden our investor base, enhance the liquidity of our shares and provide greater flexibility in accessing capital across both markets. As we move forward, AI remains central to Baidu's long-term growth and competitiveness. We will continue to invest with conviction and discipline in the capabilities and businesses where we see the strongest long-term opportunities.

Now let me walk through the details of our second quarter 2026 financial results. Total revenue of Baidu was RMB 31.3 billion, decreasing 2% quarter-over-quarter and 4% year-over-year. Revenue from Baidu General Business was RMB 25.2 billion, decreasing 3% quarter-over-quarter and 4% year-over-year. Revenue from iQIYI was RMB 6.3 billion, increasing 1% quarter-over-quarter and decreasing 5% year-over-year. Cost of revenues was RMB 19.1 billion, decreasing 3% quarter-over-quarter, primarily due to a decrease in costs related to AI Cloud business, partially offset by an increase in traffic acquisition costs and increasing 4% year-over-year, primarily due to increases in costs related to AI Cloud business.

Operating expenses were RMB 9.2 billion, decreasing 1% quarter-over-quarter and decreasing 17% year-over-year. The year-over-year decrease was primarily due to decreases in expected credit losses, channel spending expenses and R&D personnel-related expenses. Operating income was RMB 3.0 billion and operating margin was 10%. Non-GAAP operating income was RMB 3.8 billion and non-GAAP operating margin was 12%. Total other income net was RMB 184 million compared to RMB 626 million last quarter and RMB 4.9 billion for the same period last year. The year-over-year decrease was primarily due to a decrease in fair value gain from long-term investments and an increase in net foreign exchange loss arising from exchange rate fluctuation between Renminbi and U.S. dollar.

Income tax expense was RMB 1.0 billion compared to RMB 528 million last quarter and RMB 881 million for the same period last year. Net income attributable to Baidu was RMB 2.3 billion. Net margin for Baidu was 7% and diluted earnings per ADS was RMB 5.74. Non-GAAP net income attributable to Baidu was RMB 2.6 billion. Non-GAAP net margin for Baidu was 8% and non-GAAP diluted earnings per ADS was RMB 7.22. We define 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 June 30, 2026, total cash and investments were RMB 283.1 billion. Operating cash flow was RMB 3.4 billion. Baidu General Business had approximately 27,000 employees as of June 30, 2026.

With that, operator, let's now open the call to questions.

Operator

[Operator Instructions] The first question today comes from Alex Yao with JPMorgan.

分析師問答

Alex Yao

So with multitrillion parameter models emerging rapidly and pushing the frontier on benchmark performance, how does Baidu think about ERNIE's competitive positioning from here? Following the recent addition of a senior foundation model talent, what are the key technical and product priorities for ERNIE? And what should investors expect from its next stage of development?

Yanhong Li

This is Robin. First, from an industry perspective, foundation models are still evolving rapidly, roughly every few months, different model takes the lead in some capability. This shows the field remains highly dynamic and the competitive landscape is far from settled. In a market like this, we believe long-term competitiveness ultimately comes down to sustained technology investment, application-driven approach and patience. Baidu has always been a company that believes in technology and is willing to commit to it for the long term. Our experience has repeatedly shown that meaningful technological innovation takes patience and persistence. Today, many of Baidu's important AI assets, including Kunlunxin and Apollo Go, are the result of more than a decade of sustained investment. They become a key source of our differentiated competitiveness. And their performance and commercial value are gaining increasing broad recognition. So we are very proud of that.

ERNIE has likewise always been an important part of Baidu's AI strategy and full stack AI capabilities. We were among the first companies in China to invest in foundation models. There were trials and errors along the way, but our commitment to make ERNIE competitive remains unwavering. Going forward, we will continue to invest resources needed to drive ERNIE's ongoing development. As part of this effort, we have further optimized our organization and recently brought in top AI talent. We are confident in accelerating AI's iteration and bring ERNIE back into the top tier of foundation models.

Looking ahead, we will continue to take an application-driven approach. Foundation models span a very broad range of capabilities and no single model can lead in every dimension at all times. We will, therefore, focus on capabilities that matter most to Baidu's applications and make ERNIE strongest in these areas. Spanning AI search, digital human, Miaoda, Famou and general-purpose agents like DuMate. These applications are vital parts of ERNIE's continuous improvement. I take AI search as an example, when we improve ERNIE's ability to understand user intent and assess content quality, we apply those improvements directly to search and feed. This lets us quickly see the results, identify what still needs work and feedback the relevant data into model training, which makes our model better at user intent understanding and content quality assessment. And we see this loop as an important path for ERNIE's development, one that translates technological progress more directly into better product experiences and real user and commercial value, and then ultimately benefiting a broader range of users and businesses.

Operator

The next question comes from Alicia Yap with Citigroup.

Alicis a Yap

My question is on cloud. So Baidu AI Cloud Infra revenue has maintained strong growth. Could management discuss the key growth drivers and also your outlook for the revenue growth over the next few quarters? And also, how should we think about the long-term margin potential as the business scale?

Dou Shen

Alicia, this is Dou. AI Cloud Infra revenue grew 50% year-over-year in Q2. This remains a robust growth rate and above the industry average, I believe. So over a longer horizon, our AI Cloud Infra has sustained rapid growth for several consecutive quarters, consistently outpacing the industry. The standout of this quarter was the GPU Cloud, whose revenue grew 283% year-over-year, marking its fourth straight quarter of triple-digit growth and accelerating further from 184% in Q1. Looking ahead, we see several drivers supporting continued growth.

Currently, demand for AI computing in China remains very strong. And as AI becomes more deeply embedded in real-world applications and business workflows, particularly as inference continues to scale rapidly, we expect demand to grow further. Meanwhile, our customer base is also expanding rapidly with new customers of different sizes adopting our AI Cloud Infra, while existing key customers keep increasing both usage and spending. Demand is also broadening across industries and use cases, including internet, gaming, embodied AI, autonomous driving, smartphones, financial services, and others. Actually, most importantly, we have built and continued to strengthen differentiated full stack AI architecture, spanning chips, cloud infrastructure, models, and applications with competitive offerings at every layer.

At the application layer, in particular, we moved early to build a portfolio of agents and AI applications with products such as Famou Agent, DuMate, Miaoda, and Baidu Yijing gaining traction and strengthening our ability to capture an increasingly diverse range of AI opportunities. So based on current demand trends, our customer pipelines and these differentiated advantages, so we feel confident that AI Cloud Infra can maintain strong growth in the second half with the potential for further acceleration.

On the profitability side, you just mentioned, we are pleased with the continued improvement alongside rapid revenue growth. In Q2, AI Cloud Infra profit and margins both increased year-over-year. Going forward, we think several factors should support further margin expansion. First, GPU Cloud is growing significantly faster than the AI Cloud Infra in general and continues to represent a large share of the mix. It also carries a more attractive margin profile than traditional CPU Cloud, with further room for margin improvement as it scales, supported by continued optimization of its product and customer mix, better resource utilization and greater operating efficiency. So as GPU Clouds contribution increases, the mix shift should continue to lift overall margins.

Second, on MaaS. Revenue from external customers token calls on Qianfan is growing very fast. While MaaS still represents a relatively small share of our AI Cloud Infra revenue today, so the early momentum we are seeing is very encouraging. As usage skills and unit inference costs keep coming down, we believe over the longer term, MaaS-related businesses will be able to unlock more profit potential and become an increasingly meaningful contributor to margins. Finally, our full stack AI capabilities and self-developed chips also provide end-to-end cost advantages that should support margin expansion. So taken together, we think there's still a lot of room for AI Cloud Infra margins to improve over the long term. Thank you.

Operator

The next question comes from Miranda Zhuang with Bank of America Securities.

Xiaomeng Zhuang

My question is about margin. So with AI-powered business now accounting for half of the revenues and also CapEx continuing to ramp, so how do you think about Baidu's operating margin trajectory? And how will management balance the continued AI investments with profitability?

Haijian He

Thank you, Miranda. This is Henry. This quarter, AI-powered business continued to account for half of the Baidu General Business revenue, further underscoring AI's positioning at the center of our business. Within AI-powered business, AI Cloud Infrastructure sustained rapid revenue growth with profit also growing quickly and margins improving year-over-year. Within AI Cloud Infra, our GPU Cloud business, which typically carries a better margin profile, continue to increase as a percentage of revenue. As this favorable mix shift continues, together with the strong market demand and the cost advantage we get from our self-developed chips and full stack AI capabilities, we believe there is still meaningful room for AI Cloud Infrastructure margins to expand over the long term.

As the business scales, we also expect better resource utilization and greater operational leverage to provide further support for the margin expansion. We also see attractive long-term profitability potential in our AI applications. Many of these applications are sticky and subscription-based by nature with the potential to deliver increasingly attractive margins over time as they scale. As adoption growth and monetization progresses, we expect them to become a more meaningful contributor to overall profitability.

Meanwhile, I think we are still in an AI investment cycle and our commitment to that investment is unwavering. We invest with conviction, but just as importantly, we spend wisely and stay closely focused on the ROI. Our investments are driven by a clear demand from both customers and our internal business, allowing much of where we will invest in to be able to work quickly and begin contributing to revenue relatively soon. Meanwhile, we are continuously strengthening our supply chain management capabilities, which we believe will increasingly help us improve capital efficiency as we scale. Together, these strengths give us a good visibility into returns and are confident in our ability to improve investment efficiency over time.

That said, different AI investments play out on different time lines, and we will take longer to fully deliver their value. Some of them will take longer. We are now in a critical phase of investment, and we intend to keep investing decisively in the areas that matters most to our long-term competitive position while maintaining the same discipline around ROIC, operating efficiency and cash flow. As our AI business scale further and monetization matures, we believe these investments will increasingly translate into some and more sustainable profit growth. Thank you...

Operator

The next question comes from Lincoln Kong with GS.

Lincoln Kong

Could you update us on the progress of Kunlunxin's proposed listing and the key milestones ahead? I'm wondering what will drive its future growth? And how does management view its long-term commercial potential and strategic role within Baidu AI ecosystem?

Dou Shen

Okay. I'll take it. This is Dou. The listing process for Kunlunxin is still ongoing, and we will update the market as soon as we have more to share. From a business perspective, we remain very confident in Kunlunxin's long-term growth and commercial potential for a few reasons. First, across the industry, demand for AI compute continues to grow across both training and inference. As model capabilities keep improving and more applications moving into real-world use, especially as agents advance and expand into a wider range of use cases. So we are seeing inference pick up pace in particular. So we believe this trend will continue, creating a long-term structural growth opportunity for the AI chip industry.

Secondly, the domestic market carries significant growth potential, while with supply likely to remain constrained for some time. So against this backdrop, customers are increasingly seeking high-performance, reliable and cost-efficient domestic AI chips. So we believe this creates substantial opportunities for chip providers with strong technical capabilities and the ability to deliver at scale. Following more than a decade of investment, Kunlunxin has built solid capabilities in the chip performance, hardware software integration, compatibility with the mainstream models and frameworks and large-scale deployment, earning growing recognition from customers. Those are the things that put Kunlunxin in a good position in this market and capture the commercial opportunities arising from China's growing AI computing needs.

Thirdly, within Baidu's AI ecosystem, Kunlunxin is an important part of the infrastructure layer in our full stack AI architecture, spanning chips cloud infrastructure models and applications. The close coordination across these layers enables end-to-end optimization, allowing us to deliver greater performance, reliability and cost efficiency. So this supports the long-term deployment of AI Cloud Infra and our other AI businesses while further strengthening the competitiveness of Baidu's full stack AI capabilities. Looking ahead, we expect Kunlunxin to keep playing a meaningful role in our AI infrastructure, capturing a broader range of commercial opportunities and serving a wider range of market needs. Thank you.

Operator

The next question comes from Wei Xiong with UBS.

Wei Xiong

Could you walk us through the expected time line for the Hong Kong dual primary listing conversion and potential Stock Connect inclusion? Also, what's the strategic rationale? And how could it affect Baidu's investor base, share liquidity, and valuation over time?

Haijian He

Thank you. This is Henry. Let me start with the time line. Our Board has approved the conversion to a dual primary listing back in July. We also filed our application with the Hong Kong Stock Exchange and received its acknowledgment. The next step is our extraordinary general meeting scheduled on August 26. During that meeting, we will seek shareholder approval for certain matters required in preparation for conversion. And from there, we expect the conversion to take effect within this year, subject to approval of the Hong Kong Stock Exchange and other applicable conditions.

On Southbound Stock Connect, we are also actively preparing for potential inclusion following the conversion and hope our shares can be included at the earliest opportunity. And of course, this will remain subject to the applicable eligibility requirements and review of procedures and the decisions for the relevant exchange. As for the rationale, due primary listing is really about broadening our investor base, enhancing the liquidity of our shares and giving us greater flexibility in accessing both the Hong Kong and the U.S. capital markets. It also allows more investors, particularly in Asia, to better understand and participate in Baidu's value as an AI-first company.

Looking further out, if we achieve Stock Connect inclusion down the road, we would expect that to meaningfully expand participation from Mainland China investors specifically, which should support an even more diversified shareholder base over time. We will be happy to keep you updated as we make further progress.

Operator

The next question comes from Thomas Chong with Jefferies.

Thomas Chong

Could management update us about AI search progress across product capabilities, user experience, and monetization. We are seeing online marketing revenue remained under pressure in Q2. What were the main factors? And how does management expect the business to trend in the second half?

Rong Luo

Thomas, this is Julius. Let me take your question. I think over the past few quarters, our focus on the AI search transformations have been improving the quality of our AI answers through enhancing the user experiences a lot. Accuracy and authority has always been our core strength, and we have kept reinforcing that as the AI transformation moves forward. And now our AI search can better understand what users are looking for. The answers are more reliable, better structured and presenting in richer formats. And meanwhile, hallucination rates remains low, and our models now are getting better at telling good content from bad. So we are facing more high-quality answers and fewer weak ones. Users have responded quite well on these changes, and we are seeing steady improvements in user satisfaction, the willingness to search and retention.

In this quarter, we further integrated AI search with the ERNIE assistant, turning one of the search answers into more coherent interactive multi-round conversations that better address the follow-up questions and broader user needs. And we are also continuing to strengthen the to-use multistep planning and complex task executions to help users to get more done. Recently, the ERNIE Assistant's Task Agent topped two influential third-party agent benchmarks, the Pinch Bench V2, which is a global benchmark focused on the real-world complex tax complexion and the SuperCLUE XCloud evaluation of the leading domestic agent products. I think these results can help to reinforce the ERNIE Assistant's leading capabilities in the to-use multistep planning and task executions.

That said, the competition in this industry remains very intense and as new product forms like AI chatbots continue to get traction, the ways users discover and consume information is keep evolving and competition for users' time and attention has intensified further. Meanwhile, we have continued to push forward with the AI search transformations while deliberately holding back on monetizing the AI search, both of which we have weighted on our advertising businesses in the near term. Given these dynamics are likely to persist, we expect our advertising business to remain under pressure in the second half. On monetization, our priority right now is still to getting the products and the user experience right. As model capabilities, the user experiences and the task complexions continue to improve, we believe that more monetization opportunities, which fit naturally into the AI experiences will emerge in the future. Thank you, Thomas.

Operator

The next question comes from Ellie Jiang with Macquarie.

Ellie Jiang

I have a question on robotaxi, please. So with China's recent introduction of the new robotaxi policies, how does management view the evolving regulatory environment? How should we think about Apollo Go's relative focus and also the pace of expansion across domestic overseas market? And it would be great if management can talk about the progress that Apollo Go has made in the overseas commercialization side.

Yanhong Li

This is Robin. Let me answer this question. The global robotaxi industry is evolving very quickly. In the past, the industry's focus was on whether robotaxis could deliver a safe, comfortable riding experience. Today, that focus has expanded to whether robotaxis can operate reliably at scale and fit into the broader transportation system. In line with this trend, major markets around the world are also iterating on and refining their regulatory frameworks for robotaxis.

In China, for example, the country's first mandatory national standard on safety requirements for Level 3 and Level 4 automated driving systems was recently issued. And Apollo Go contributed its extensive technical and operating experience to the L4 requirements under this standard. Safety has always been our top priority, and we maintain an industry-leading safety record globally. We will continue to uphold high standards on safety and operations. More broadly, a clearer, more systematic regulatory frameworks will help raise operational standards across the industry, build public trust and lay a stronger foundation for the long-term orderly growth of robotaxis. Against this backdrop, we remain positive on Apollo Go's global expansion.

We do not view domestic and international markets as an either/or choice. We are highly open and adaptive. We assess each city based on its regulatory framework, mobility demand, right pricing, road conditions, and commercial viability and setting our pace of entry and expansion accordingly. Backed by proven technology and operating experience, we are ready to move quickly and scale efficiently in any city where regulations and market conditions allow. Our goal is to go deep and build a solid presence in every city we enter regardless of country boundaries.

This is reflected in our progress across different cities. In Dubai, Apollo Go has entered fully driverless commercial operations and is scaling up, and we now operate at the largest scale among robotaxi services providers over there. In London, we are advancing testing and development with partners, including Uber and Lyft. In Hong Kong, we became the first robotaxi service globally to conduct fully driverless testing in our right-hand drive, left-hand traffic market. In Shenzhen, the number of rides are picking up very quickly, making it one of our largest markets.

As our fleet expands and our operating model matures, we expect vehicle and operating costs to keep coming down, while scale brings additional efficiency gains. In the past, Apollo Go achieved unit economics breakeven in a market with relatively low taxi fares. In the future, in overseas markets with higher ride prices, our low-cost vehicles and proven operating model have the potential to deliver even stronger unit economics. The international market outside of U.S. and China is also larger than the domestic China market. So the addressable opportunity is quite substantial. Looking ahead, supported by our advantages in technology, cost and operations, we are confident in bringing more cities to unit economic breakeven. Thank you.

Operator

Ladies and gentlemen, that does conclude our conference for today. Thank you for participating, and you may all disconnect.