BAIDU INC

Business Model and Scope

Baidu is no longer adequately described as a search company, although search advertising remains the financial base that funds its reinvention. Beginning in the fourth quarter of 2025, the company renamed Baidu Core as Baidu General Business. That group contributed more than 70% of 2025 consolidated revenue and contains three unlike economic systems: the Mobile Ecosystem, including Baidu App, ERNIE Bot, Haokan and Baidu Post; AI Cloud, spanning infrastructure and applications such as Wenku and Drive; and Intelligent Driving & Other, including Apollo Go and DuerOS. iQIYI remains a separately reported online-entertainment business with subscription and advertising economics.

The Mobile Ecosystem monetizes attention and commercial intent, principally through online marketing. AI Cloud sells computing capacity, model services and applications to enterprises and public-sector customers. Apollo Go deploys autonomous ride-hailing rather than merely licensing research, so its economics include vehicle utilization, fleet operations and regulatory access. iQIYI pays for content and distribution to attract subscribers and advertisers. These activities share technology and traffic, but they do not share the same unit economics, capital intensity or competitive position.

In 2025 Baidu reported revenue of RMB129.1 billion, down 3% from 2024. Online-marketing revenue fell to RMB67.8 billion from RMB78.6 billion in 2024 and RMB81.2 billion in 2023; other revenue rose sufficiently to offset only part of that erosion. Baidu General Business revenue was RMB102.5 billion, down 2%, while iQIYI revenue declined 7% to RMB27.3 billion. Baidu App had 679 million monthly active users in December 2025, demonstrating reach but not, by itself, profitable engagement.

Customers and Purchasing Decisions

The paying customer depends on the product. Search and feed users usually pay nothing; advertisers pay for measurable access to those users, often through performance-based auctions. The advertiser base includes small and medium-sized businesses across healthcare, retail, e-commerce, media, games, business and life services, and transportation. Their willingness to pay depends on conversion quality, regulatory permission to advertise, and whether commercial discovery occurs on Baidu rather than inside competing apps.

Distribution partners in Baidu Union supply traffic and placements and receive revenue shares or fees. They are suppliers of audience as much as customers of technology. If publishers or handset ecosystems command scarce distribution, they can retain more of the advertising economics. Conversely, Baidu's scale and monetization tools can make participation attractive to fragmented partners.

Cloud customers buy computing, storage, model training or inference, and applications. Large enterprises and public-sector bodies can negotiate price, demand data localization and reliability, and spread workloads across vendors. Their switching costs rise when Baidu's models, data tools and applications become embedded in workflows, but standardized infrastructure and open models keep switching possible. Apollo Go serves riders, while municipalities, transport authorities, vehicle makers and fleet partners determine whether the service can operate and scale. iQIYI serves two paying groups—members and advertisers—whose interests can conflict when advertising interrupts the viewing experience.

The central customer question is not whether Baidu has many users. It is whether it can preserve high-value commercial intent as users migrate toward short video, closed e-commerce ecosystems and AI-generated answers. A large audience becomes economically valuable only when Baidu can match demand with an advertiser, cloud workload or paid service at a contribution margin exceeding traffic, content, compute and sales costs.

Profit Creation and Value Capture

Baidu's established profit engine converts user intent into auction-priced advertising. A query or feed interaction reveals potential demand; Baidu ranks paid and unpaid results, charges advertisers for qualified exposure or clicks, and pays traffic-acquisition costs where partners supplied the audience. Incremental revenue can carry attractive margins after the search index, product engineering and sales infrastructure are in place. The economic surplus is divided among Baidu, distribution partners, advertisers and users: Baidu retains the auction and data advantage, partners capture traffic payments, advertisers keep conversion value above acquisition cost, and users receive a free service.

That engine is weakening in reported revenue. Online marketing contracted by RMB13.4 billion between 2023 and 2025. AI products may deepen engagement, but answer-style interfaces can also reduce page inventory and outbound clicks. If generative answers satisfy users without producing monetizable commercial actions, better utility can coexist with lower advertising profit. The relevant bridge is therefore from query to paid conversion, not from model usage to revenue.

AI Cloud creates gross profit when contract revenue exceeds servers, depreciation, bandwidth, data-center custody, power, model-development support and customer-acquisition costs. Scale can improve hardware utilization and spread software costs, while proprietary applications can carry better margins than raw infrastructure. Yet 2025 cloud growth came with higher bandwidth, depreciation and custody costs, and the filings do not disclose a stand-alone cloud margin. Assertions that cloud growth already replaces search profit would exceed the evidence.

iQIYI earns membership fees and advertising revenue but must continuously fund content, bandwidth and marketing. Hit content can spread fixed production cost across many viewers; a weak slate reduces subscriptions and advertising before costs can fully adjust. Apollo Go's prospective profit is more capital-intensive: fare or platform revenue must cover vehicles, remote operations, mapping, maintenance, insurance, charging and fleet depreciation. Regulators, fleet partners and vehicle suppliers will capture part of any surplus.

The consolidated result shows why accounting labels must be separated. In 2025 revenue was RMB129.1 billion against cost of revenue of RMB72.4 billion, selling and administrative expense of RMB25.8 billion, R&D of RMB20.4 billion and a RMB16.2 billion long-lived-asset impairment. The company reported a RMB5.8 billion operating loss; excluding that impairment, operating income would have been about RMB10.4 billion. Net income attributable to Baidu was RMB5.6 billion partly because RMB12.5 billion of other income sat below operating profit. Neither the impairment nor non-operating income should be mistaken for current cash unit economics.

Industry Structure and Capital Cycle

Baidu competes across several converging industries. In discovery and advertising, it faces other search services, short-video platforms, social and commerce apps, and vertical services that keep user activity inside closed ecosystems. In generative AI, foundation-model developers compete on capability, latency, price, distribution and developer adoption. In cloud, large providers offer overlapping compute and model services. iQIYI competes with long-form platforms, short video, games and other leisure. Apollo Go competes with conventional ride-hailing as well as other autonomous-driving developers.

Buyer power is meaningful. Advertisers can shift budgets when measured returns deteriorate; cloud customers can run tenders and use multiple vendors; consumers can change apps at little monetary cost. Supplier power is also rising. Advanced chips, data-center capacity, electricity and network connectivity are concentrated or constrained, while export controls can limit access to leading hardware. Content creators and studios bargain with iQIYI, and vehicle, component and municipal partners matter to Apollo Go.

Substitutes attack the profit pool from both sides. A user may begin a product search in an e-commerce app, obtain an answer from an AI assistant, or consume short video instead of using Baidu or iQIYI. An enterprise may deploy an open-source model on another cloud. A city can continue using human-driven transport rather than authorize autonomous fleets. These substitutes cap price and increase the spending needed to retain attention.

Entry conditions differ by layer. A credible Chinese search index, advertiser marketplace, safety system and national infrastructure require data, engineering, capital and regulatory approval. But model software is diffusing rapidly, and a well-funded platform with existing distribution can enter AI services without rebuilding the old web-search ecosystem. The capital cycle is therefore unfavorable when many large firms purchase similar accelerators and discount inference to gain users: capacity expands before durable willingness to pay is known. Baidu's capital expenditure rose to RMB12.1 billion in 2025, from RMB8.1 billion in 2024, primarily for generative-AI servers and network hardware. Returns depend on utilization and pricing after competitors finish comparable build-outs.

For iQIYI, content investment follows a separate cycle: successful programming attracts subscribers, prompting competitors to spend, which raises talent and production costs until weaker returns force retrenchment. Autonomous driving has an even longer cycle because fleets and permits must precede dense utilization. Baidu is consequently allocating cash from a declining high-margin activity into markets where both capacity and price are unsettled.

Sources and Durability of Competitive Advantage

Baidu's strongest established advantages are Chinese-language search data, a large user base, an advertiser-and-distribution network, a mature ranking and monetization system, and brand familiarity. Baidu App's 679 million monthly users provide repeated interactions and distribution for new services. Historical query and conversion data can improve relevance; advertisers follow effective demand, and their participation increases auction density. These mechanisms reinforce one another when users continue to bring commercially valuable queries.

The durability is conditional. Closed platforms restrict data and divert user intent, while AI assistants change the interface from ranked links to synthesized answers. Training data and model techniques can diffuse, and competitors can subsidize inference. Baidu's installed distribution lowers the cost of introducing ERNIE products, but distribution does not guarantee that new usage will monetize at search-like margins. The decline in online-marketing revenue is contradictory evidence against an unqualified network-effect claim.

Cloud switching costs and application integration may become advantages when a customer builds processes around Baidu's tools, but the filings do not demonstrate superior segment returns. Apollo's accumulated autonomous-driving data, safety processes and permits could be difficult to replicate, yet commercial density and full cost remain unproven in the disclosed accounts. iQIYI's library and brand can attract viewers, though hit-driven content is perishable and competitors can bid for similar talent.

The 2025 RMB16.2 billion impairment of Baidu General Business long-lived assets is especially important. It reflects a reduction in expected discounted cash flows for the tested asset group, even though the filing does not identify product-level causes. It is evidence that prior economic expectations did not fully hold. A defensible conclusion is that Baidu owns valuable distribution and technical assets, but the ability to convert them into durable returns through the AI transition remains under test.

Operating System and Strategic Trade-offs

Baidu's operating system links large-scale product experimentation with centralized infrastructure. Search and feed collect user signals; the advertising platform translates them into bids; cloud infrastructure supports internal and external AI workloads; and ERNIE technology can be distributed through Baidu's consumer and enterprise products. This shared stack can reduce duplication and accelerate deployment. It also makes the allocation of compute and R&D costs across mature and emerging products opaque.

R&D expense declined from RMB24.2 billion in 2023 to RMB22.1 billion in 2024 and RMB20.4 billion in 2025, while AI-related capital expenditure increased. That shift does not necessarily mean innovation weakened: expenditure may have moved from personnel and development into hardware. It does mean that investors should not equate server purchases with proprietary technical progress. Capitalized infrastructure creates depreciation and utilization risk; research expense is consumed immediately.

Management must run three different feedback loops. Advertising requires relevance, conversion measurement and policy enforcement. Cloud requires service reliability, workload utilization and disciplined contract pricing. Autonomous driving requires safety incidents, regulatory milestones and fleet economics to be measured at the city level. iQIYI must balance content acquisition with subscriber retention. Consolidated growth can hide deterioration in one loop behind spending in another, so operating discipline should be judged by disclosed cash conversion and segment evidence, not product announcements.

Governance is part of the system. The Cayman holding-company and VIE structure separates public shareholders from direct ownership of certain restricted PRC operations. Contractual control depends on enforceability under PRC law. Concentrated voting power further limits outside shareholders' influence. These arrangements can support strategic continuity, but they weaken the ordinary mechanism by which minority owners discipline capital allocation.

Financial Resilience

Baidu entered the AI investment cycle with substantial liquidity. At year-end 2025 it held RMB24.6 billion of cash and cash equivalents, RMB90.7 billion of short-term investments, RMB123.9 billion of long-term time deposits and held-to-maturity investments, plus RMB44.9 billion of long-term investments. These categories differ in liquidity and valuation, but together they provide material resources.

They must be considered against obligations. Reported balances included RMB7.6 billion of short-term loans, RMB14.8 billion of current long-term loans, RMB3.4 billion of non-current long-term loans, RMB4.6 billion of current notes, RMB51.0 billion of non-current notes, and RMB8.2 billion of current and non-current convertible notes. The long-term debt maturity schedule, excluding convertibles, was weighted beyond 2026, which reduces immediate refinancing pressure but does not eliminate interest and principal claims.

The main warning is cash conversion. Operating activities used RMB3.0 billion in 2025 after generating RMB21.2 billion in 2024 and RMB36.6 billion in 2023, primarily because working capital deteriorated. Capital expenditure was another RMB12.1 billion. One year does not establish a permanent reversal, but negative operating cash flow during an investment surge is incompatible with treating reported net income as freely distributable cash.

Resilience therefore rests more on accumulated liquidity than on current operating momentum. A prolonged combination of advertising decline, price competition in AI, content commitments and compute spending could consume that buffer. Currency controls and subsidiary or VIE restrictions may also prevent every balance from being moved freely to the entity that owes a particular obligation. The balance sheet can fund experimentation, but it does not prove that experimentation earns its cost of capital.

Capital Allocation and Shareholder Outcomes

Baidu's allocation problem is to harvest search without starving it, while funding businesses that may replace its economics. The company spent heavily on R&D over the five-year record and accelerated AI infrastructure in 2025. This is strategically coherent if cloud, applications or autonomous services generate durable cash flows; it destroys value if industry-wide capacity makes compute a commodity or if new interfaces erode advertising faster than new revenue scales.

The RMB16.2 billion impairment is a retrospective capital-allocation signal. Although non-cash in 2025, it indicates that the expected value of prior investments or assets fell. Excluding it clarifies current operating performance but does not reverse the economic loss. Similarly, share-based compensation of RMB3.6 billion transferred value to employees even though it declined from RMB6.3 billion in 2023.

The board authorized up to US$5 billion of additional share repurchases in the first quarter of 2026 through December 2028 and stated an intention to establish a dividend policy, with an initial payment expected in 2026. These actions fall within the filing cutoff, but an authorization is not a return until executed. Repurchases benefit continuing holders only when they retire shares at a value below conservative intrinsic value and do not displace higher-return investment or needed liquidity. A dividend can impose discipline, yet it competes with an unusually uncertain infrastructure program.

Public shareholders receive residual economics only after employees, lenders, content and hardware suppliers, distribution partners, tax authorities, VIE counterparties and iQIYI's noncontrolling holders. The appropriate measure of shareholder outcome is sustained per-share cash generation after these claims, not consolidated revenue growth or gross AI usage.

Legal and Regulatory Exposure

Baidu operates in areas central to PRC regulation: internet content and censorship, advertising, algorithms and generative AI, cybersecurity, personal information, cross-border data transfer, cloud services and autonomous driving. A rule change can reduce monetizable inventory, require model or content remediation, constrain data use, delay a city launch or raise compliance cost. Advertising restrictions are particularly relevant in regulated industries such as healthcare, where customer demand and legal exposure can coexist.

The VIE structure is a fundamental legal exposure rather than a drafting technicality. The listed Cayman company relies on contracts to control and receive benefits from businesses it may not own directly. If those arrangements were challenged, public shareholders could have weaker remedies than direct owners. Cash transfers among the holding company, subsidiaries and VIEs can also be subject to approvals, taxes and solvency constraints.

Advanced-computing export controls can restrict chip supply or raise infrastructure cost. Privacy and cybersecurity rules affect the data needed for model training and personalization. Apollo Go depends on local permits, safety standards and liability rules that can vary by jurisdiction. iQIYI faces content review, intellectual-property and consumer-protection obligations. Cross-border audit and listing regimes add another channel through which political decisions can affect shareholder access independent of operating performance.

The relevant legal test is not merely whether Baidu is currently compliant. It is whether the business can preserve product quality and returns after complying with tighter rules. Regulation can reinforce incumbency by raising entry costs, but it can simultaneously remove profitable activity or trap capital.

Conclusion, Uncertainties and Disconfirming Evidence

Baidu has a real but narrowing economic foundation: a scaled Chinese-language discovery product, a large audience and an advertising system that historically monetized intent efficiently. It also has liquidity and technical infrastructure sufficient to attempt a transition toward AI cloud, applications and autonomous services. Those assets explain why the company can invest at a scale unavailable to most entrants.

The contrary evidence is substantial. Online-marketing revenue declined for two consecutive years, 2025 operating cash flow turned negative, and a RMB16.2 billion impairment showed that prior cash-flow expectations for General Business assets fell. Meanwhile, the replacement businesses require more hardware, face strong buyers and suppliers, and lack disclosed stand-alone margins. iQIYI adds cyclical content economics, and the VIE and concentrated-control structure limits minority shareholders' direct claim and influence.

The thesis that Baidu can convert search-era advantages into durable AI profits would strengthen if online marketing stabilized, operating cash flow recovered after working-capital normalization, and disclosed evidence showed cloud or autonomous operations earning attractive returns after compute, depreciation and fleet costs. It would be invalidated by continued loss of commercial-intent traffic, sustained negative operating cash flow, recurring impairments, AI pricing that fails to cover infrastructure, or regulation that materially restricts data, advertising, VIE benefits or autonomous deployment. The decisive question is not whether Baidu can build capable AI products; it is whether those products can preserve per-share cash economics after customers, suppliers, regulators and capital providers take their share.

Business quality does not by itself establish investment attractiveness; valuation depends on the price paid and the expectations embedded in it.