
Baidu (Trans): AI cloud infra +94%, ads still under pressure
Below is Dolphin Research's $Baidu(BIDU.US) $BIDU-W(09888.HK) FY26 Q3 earnings call Trans.
I. Key takeaways
1. Shareholder returns: Plans to announce its first cash dividend by year-end (subject to BOD approval and market conditions). YTD share repurchase totals Approx. RMB 373 mn, with buybacks to continue under the existing authorization at a measured pace, balancing growth investment and returns.
2. Outlook: No quantified guidance. Management expects AI Cloud Infra to sustain rapid growth over the next few quarters and gradually offset ad headwinds; ads remain pressured near term by macro, competition and the AI search transition, keeping profit and margins under pressure. Revenue trajectory is expected to improve over the medium to long term.
3. Capex and cash flow
a. Capex: Q3 capex was RMB 4.0 bn lower vs. Q2, mainly as earlier compute reserves are ample and utilization improved. In Q4 and next year, growth will be prioritized with existing compute where possible, and new investments will be paced to match customer demand and collections.
b. Cash flow: Operating cash flow was RMB 2.8 bn, positive for the fifth straight quarter. Increased customer prepayments accelerated collections.
c. Liquidity: Net cash was Approx. RMB 133 bn at quarter-end, with long-term investments of Approx. RMB 43 bn. Cash plus investments totaled Approx. RMB 276.6 bn.
4. Key financials this quarter
a. Revenue: Total revenue was RMB 30.8 bn (-1% YoY). Core BIDU revenue was RMB 25.0 bn (+1% YoY), returning to growth; iQIYI revenue was RMB 6.0 bn (-10% YoY).
b. Mix: AI-driven revenue was RMB 13.2 bn, accounting for 53% of core revenue. AI Cloud Infra revenue grew +94% YoY (vs. +50% in Q2); GPU Cloud rose +191% YoY and Approx. +43% QoQ.
c. Profit: OP was RMB 626 mn (OPM 2%). Non-GAAP OP was RMB 1.7 bn (6%), and Non-GAAP attributable net profit was RMB 2.4 bn (8%).
d. Cost and opex drivers: COGS was RMB 19.9 bn (+9% YoY), driven by higher cloud costs and TAC, partly offset by lower content costs. Opex was RMB 10.3 bn (-13% YoY on lower expected credit losses and channel spend) and +12% QoQ (on higher expected credit losses).
II. Earnings call details
2.1 Management remarks
1. AI Cloud Infra: GPU Cloud nearly doubled revenue, with profits and margins up both YoY and QoQ. Demand and longer-term contracts underpin continued growth.
a. Revenue mix: Higher-margin lines such as GPU Cloud, Kunlun Chip and MaaS now represent the majority of AI Cloud Infra revenue, and better utilization is expanding margins and shortening payback. Profitability across these lines is also improving.
b. Demand: Training and inference demand are both strong, with inference growing faster and still early in scaling. E-comm, internet, autonomous driving and embodied AI are key growth verticals.
c. Long-term deals and prepayments: The top three GPU Cloud customers by TCV renewed into multi-year terms this year, with one extending to 6x the prior term. Prepayments are rising, improving visibility and cash profile.
d. MaaS: The Qianfan platform offers ERNIE alongside mainstream domestic open-source models. External customer token-based revenue rose over 12x YoY, with margin continuing to improve.
2. Model layer: Launched ERNIE 5.5 Flash Preview, with R&D anchored in application-driven iteration, and rolled out a model trading marketplace.
a. ERNIE 5.5 Flash Preview: 280 bn total params with 14 bn activated per token, featuring strong coding and Agent capabilities. The design targets speed and cost efficiency.
b. Token Exchange: A neutral and transparent marketplace for model services, enabling customers to compare price and performance side-by-side. This aims to streamline model selection.
3. AI apps: Building both general-purpose and vertical Agents, with monetization still early.
a. DuMate: Sep. MAU rose over 30x MoM, achieved without heavy marketing or super-app funnels. An enterprise version launched in Jun., and this quarter integrated capabilities such as Comate.
b. Digital humans: Since Q4 2025, ultra-realistic digital human production costs fell Approx. 80%, with production time down by over an order of magnitude. New intl clients include those in financials, autos, apparel, and Lazada.
c. Famou 3.0: An industrial laser equipment client used it to plan four robotic arms for coordinated cutting, reducing solution design time from 4–8 hours to Approx. 20 minutes.
4. AI search: User experience remains prioritized over near-term monetization, which is weighing on online marketing and overall profit in the short run.
a. ERNIE assistant: Deeper integration with AI search now supports follow-up queries and Agent-based tasks. In Sep., Avg. daily dialogue turns rose +126% YoY.
5. Robotaxi (Apollo Go): Entered a new phase focusing on high-demand cities and deeper penetration, now covering 30 cities globally.
a. Shenzhen: Fully driverless orders rose Approx. 170% QoQ in Q3. Penetration continues to improve.
b. Intl: Dubai fleet doubled, making it the largest fully driverless ride-hailing operator locally and integrated with Uber. In London, expanded public-road testing with Uber and Lyft; fully driverless testing at Hong Kong Airport Island is progressing steadily.
2.2 Q&A
Q: What is the growth trajectory, drivers and revenue mix for AI Cloud Infra over the next few quarters, and can growth be sustained?
A: We are confident AI Cloud Infra will sustain its momentum in the coming quarters, with GPU Cloud as the primary driver. Confidence stems from three areas: demand for both training and inference is strong and accelerating, as enterprises embed AI in core workflows making compute a recurring need, and Agent adoption is still early yet will lift demand as capabilities improve and deployment eases.
Second, the customer base is broadening across industries and sizes, with existing customers deepening usage. For example, a leading autonomous driving chip customer expanded from GPU Cloud into storage and networking, and BIDU has become its largest infra provider; as customers expand AI workloads, they require broader cloud services, and our full-stack capabilities help capture more cloud spend.
Third, the revenue mix is improving: GPU Cloud plus MaaS now account for over half of AI Cloud Infra revenue, with GPU Cloud mix still rising and MaaS small but growing fast; a sizable portion is recurring. More customers, especially in GPU Cloud, are signing longer contracts with prepayments, with some renewals extending multiples of prior terms, improving visibility.
Q: How do you position ERNIE domestically and on the roadmap, and how do you view returns on continued investment?
A: The new model ranks No.2 globally among open-source peers, with only ~1/10 the params of the leader. ERNIE 5.5 Flash Preview holds this rank on the Code Arena WebDev leaderboard. Foundational models are a core layer of our full-stack AI, and we will keep investing to accelerate iteration and will share more updates soon.
Our R&D is application-driven, aiming to lead in areas that matter most for apps, such as natural expressions and motion for digital human Agents, and deep-thinking models for AI search across coding, reasoning and Agent tasks, while balancing speed and run-cost. Discipline remains key: we focus resources and keep improving training efficiency.
Returns come from two areas: first, stronger in-house models improve app experience and lower cost, supporting monetization and profitability; second, via Qianfan MaaS we deliver ERNIE externally, driving demand for model services and enriching AI Cloud Infra revenue streams.
Q: As AI Cloud Infra scales, what is the margin outlook and the key drivers?
A: AI Cloud Infra profits and margins rose YoY and QoQ in Q3, with further room to expand long term. Four drivers: mix shift toward higher-margin GPU Cloud, Kunlun Chip and MaaS with improving profitability; scale driving higher resource utilization through cross-workload scheduling and peak balancing, reducing unit costs.
Third, full-stack, end-to-end optimization, with MaaS-side inference optimizations enabling more tokens at lower latency on the same compute. Fourth, products like Famou let us deliver higher-value services at each layer, lifting the contribution from premium services.
Q: With continued investment in compute infra, how do you assess returns?
A: Beyond revenue and margin, we focus on payment terms and expected payback, aiming to recover compute investments faster. We seek to fund the next cycle with customer cash collections, and the share of new contracts with prepayments has risen rapidly this year. As demand grows, scale expands and margins improve, cloud infra payback periods are shortening; with more prepay and faster recovery, we can reinvest earlier and scale faster.
Q: How is competition evolving in AI apps and Agents, and how are DuMate and others performing and monetizing? What is the long-term revenue potential?
A: Native AI app monetization is still early; the focus now is building products users repeatedly choose. We believe AI’s value must be realized in applications; the app market is evolving quickly, with new Agents emerging rapidly, and long-term competitiveness depends on reliably completing real-world tasks and delivering measurable productivity gains.
We have a portfolio spanning DuMate plus digital humans and vertical Agents like Famou, serving consumers and enterprises; DuMate is scaling fast primarily via organic growth, without heavy marketing or cross-traffic from other products. Monetization models will differ across products, but users pay for value.
As apps take on more complex, higher-value work, willingness to pay should rise, and we expect AI apps to become an increasingly important growth engine. We will iterate product value and align pricing with outcomes.
Q: What are the capex plans and investment priorities for this year and next?
A: Q3 capex was RMB 4.0 bn lower than Q2, while cloud growth remained strong. First, prior compute reserves are sufficient, allowing some incremental demand to be met with existing capacity; second, ongoing utilization improvements are freeing more capacity for external clients. In Q4 and into next year, we will further leverage existing compute, and pace new investments to match demand and collection cycles.
To enhance cloud returns, we will focus on three areas: further improve utilization; expand the mix of higher-margin GPU Cloud, Kunlun Chip and MaaS under a full-stack approach; and accelerate collections, with more GPU Cloud clients prepaying and signing long-term deals to reduce capital tied up in procurement and delivery. Foundational model work remains application-driven rather than leaderboard-chasing.
Q: What are the latest plans for dividends and buybacks?
A: We expect to announce the first dividend by year-end; YTD buybacks are Approx. RMB 373 mn. Any dividend is subject to BOD approval and market conditions; buybacks will continue under the current authorization at an appropriate scale, balancing long-term growth and shareholder returns.
Q: What is the status of the Kunlun Chip IPO?
A: No listing timetable was provided; the company continues to pursue an HK listing process and will update the market in due course. Kunlun Chip confidentially filed with HKEX earlier this year.
Q: What role do in-house chips play in AI Cloud Infra, and how much can they lift margins long term?
A: Kunlun Chip M100 is expected to enter mass production soon, with M300 next year for ultra-large multimodal models. Kunlun Chip reflects 15 years of sustained investment and has been deployed at scale in production at leading enterprises across financials, telecom, energy and internet.
Enterprises assess compute beyond peak performance, including compatibility with mainstream models and frameworks, ease of migration and deployment, and stability at large cluster scale. Kunlun performs well in these areas, with growing recognition of its efficiency, especially for inference; the product line builds on the shipped P800, with M100 optimized for large-scale inference.
Long term, as the domestic semiconductor supply chain matures and chips iterate and scale capacity, Kunlun will become increasingly important within AI Cloud Infra. In-house chips enable end-to-end optimization across chip, cloud, model and application, offering further cost-down as deployments scale and supporting margin expansion; no quantified margin impact was provided.
Q: How will core revenue trend over the next few quarters and long term?
A: Ads remain pressured near term, while AI Cloud Infra is expected to grow rapidly and gradually offset ad headwinds. This quarter’s return to growth in core revenue was mainly driven by AI Cloud Infra, with AI-driven lines at Approx. 53% of core revenue. Ads are weighed by macro, intense competition and the AI search transition.
As one of the earliest to rebuild search with AI globally, we are taking an aggressive approach; user engagement has deepened, but monetization will take time. We prioritize user experience over early monetization, which will continue to pressure near-term revenue, while we remain confident in the medium- to long-term trajectory.
Q: What is the medium- to long-term outlook for profit and margins?
A: Profit and margins will remain pressured near term, with AI Cloud Infra the key profit driver over time. Ads are weak, and we will continue to invest in cloud infra, in-house chips and foundational models, which will cap near-term margin improvement.
AI Cloud Infra saw YoY margin expansion this quarter and a larger contribution to total profit; as higher-margin lines scale, it should become an even more important profit source. We will maintain necessary investment while focusing on efficiency and returns.
<End of text>
Risk disclosure and statements:Dolphin Research disclaimer and general disclosure
Login to unlock13,816characters for free
This content is only available to signed-in users. Sign in to your Longbridge account to read the full post.
