Baidu, Inc. (NASDAQ: BIDU) — A Liquid Balance Sheet Funding an Illiquid Reinvention

Date: September 1, 2026 · Coverage: Update (prior report: July 2, 2026)
Reporting: U.S. GAAP in Renminbi (RMB) · ADS ratio: 1 ADS = 8 Class A ordinary shares
Structure: Cayman holding company controlling material PRC operations through subsidiaries and contractual VIE arrangements
Listings: Nasdaq ADSs and HKEX 9888 ordinary shares; dual-primary status effective September 1, 2026
Price reference: US$95.28/ADS (August 31 close) · Market capitalization: approximately US$32.35B · 52-week range: US$89.60–165.30

⚡ Claude’s Take

This block is the author’s subjective opinion and general information, not personalized investment advice. Everything after the price-event map is deliberately position-free: it analyzes evidence and scenarios without making a recommendation or setting a price target.

Verdict: SPECULATIVE BUY / accumulate only in the US$90–105 zone. Directional value zone: approximately US$85–125 per ADS. Conviction: medium-low. This is a change from the July 2 stance: the stock has fallen from US$113.30 to US$95.28, and a fresh balance-sheet reconstruction shows that the prior report materially understated fixed-income backing. Cash, short-term investments, and long-term time deposits/held-to-maturity securities, less all debt-like obligations, were approximately US$20.15B at June 30; adding the separate long-term equity and other investment portfolio raises net financial assets to approximately US$26.46B before haircuts. Against a US$32.35B market capitalization, the market is charging little for the operating businesses once those assets are credited conservatively.

That is the attraction—and almost the entire attraction. The operating thesis got worse. Q2 online marketing fell 19%, Baidu App monthly active users declined to 644 million from 679 million six months earlier, group gross margin fell to 39%, and H1 free cash flow was negative RMB11.2B as AI infrastructure spending surged. AI-powered Business grew 25%, but it fell 8% sequentially; growth was concentrated in lower-margin compute, while AI applications grew only 3% and AI-native marketing was flat. Apollo Go’s management-disclosed fully driverless rides fell to roughly 1.0 million from 3.2 million sequentially after regulatory operating adjustments. Baidu still discloses no Apollo fleet, revenue, utilization, or unit economics. This is not evidence of a durable consolidated moat or a clean earnings inflection.

The variant view is therefore asset-backed, not heroic: value Baidu as a shrinking search franchise plus an AI-cloud transition and treat Kunlunxin and Apollo as options, rather than paying today for either option to work. A conservative financial-asset haircut still credits roughly US$21B after redeemable non-controlling interests, leaving only about US$11B of market value assigned to search, cloud, Apollo, Kunlunxin, the iQIYI controlling stake, and corporate costs. The danger is that this “floor” is a funding reservoir, not a locked box: FY2025 and H1 2026 together consumed about RMB26B of post-capex cash, access to mainland balances is regulated, and founder control can keep excess capital inside the ecosystem.

What flips the view bullish: online-marketing decline improves to better than 10%, BGB returns to revenue growth without its operating margin falling below 12%, and trailing BGB free cash flow turns positive while AI-powered Business sustains at least 25% growth. What flips it bearish: another two quarters of at least 15% ad decline, capex remaining above 30% of BGB revenue without segment profitability disclosure, or net financial assets being consumed by weak-return projects and acquisitions. The stock is a quantified falling knife—below both its 50- and 200-day exponential moving averages—so position sizing matters more than false precision. Tag: “The cash is real; the moat is melting.”

📈 Stock Price Action — Five-Year Event Map

BIDU’s five-year tape is a sequence of China-policy and AI-optionality repricings rather than a steady reflection of compounding. From September 1, 2021 through August 31, 2026, the adjusted intraday high was US$182.60 and the low was US$73.58. The current US$95.28 is 47.8% below that five-year high, 42.4% below the 52-week high, and only 6.3% above the 52-week low. Price changes below are Fact from the adjusted daily series; the attributed causes are Interpretation.

Cycle narrative. (1) The March 2022 whipsaw reflected China-ADR delisting and regulatory fear before Beijing’s March 16 market-stability pledge reversed the policy signal; the price move is fact, while the multi-causal attribution is interpretation. (2) The September–October collapse combined zero-COVID demand damage with private-sector and political risk; Baidu had already reported Q2 online marketing down 10% in its August 2022 results.

(3) The reopening rebound followed COVID-rule easing and the PCAOB’s December 2022 determination that it had secured complete China audit access, compressing the ADR risk premium. (4) Confirmation of ERNIE Bot initially brought ChatGPT-style optionality into the stock; disappointment with the prerecorded launch demonstration then exposed how much expectation, rather than revenue, had entered the price.

(5) September 2024 monetary and market support produced a broad China-equity squeeze, but the move faded as the subsequent Q3 report showed revenue down 3% and online marketing down 4%. (6) The 2025 rerating was driven by secondary-market enthusiasm for Kunlunxin and domestic AI chips; that causal attribution is lower-confidence than the price data and should not be mistaken for a contemporaneous earnings inflection.

(7) The January 2026 spike followed Baidu’s primary-source proposal to spin off and list Kunlunxin, an explicit attempt to create separate price discovery. (8) The subsequent collapse tracked the FY2025 impairment, June CMC designation, and the Q2 release, when online marketing fell 19% and the ADS lost 12.7% in one session.

The current technical condition is adverse. At August 31, BIDU was 10.7% below its 50-day EMA of US$106.64 and 17.9% below its 200-day EMA of US$116.09, with the shorter average below the longer one. Raw three- and six-month returns were −30.3% and −19.7%. FactorsToday attributes the stock primarily to China (+1.23 beta) and social-media/internet exposure (+1.86), with negative momentum (−0.26), a high-volatility tilt, and only 54.9% of variance explained. Specific volatility was 36.2%. This is China-internet beta plus considerable company-specific risk, not a low-volatility “cash box.”

Changes Since the July 2, 2026 Report

The new quarter changes both sides of the thesis. Six items matter.

1. The balance-sheet floor was understated, materially. The prior analysis grouped RMB74.3B of long-term time deposits and held-to-maturity investments with illiquid equity investments. Reconstructing the June 30 balance sheet correctly yields RMB240.6B of cash and fixed-income financial assets against RMB103.9B of debt-like obligations, or RMB136.7B/US$20.15B net. A separate RMB42.9B portfolio of long-term equity and other investments sits outside that figure. This is a classification correction, not new cash creation: short-term investments rose while long-term deposits matured or were reclassified. The stronger asset coverage is real, although not all instruments are immediately available at the Cayman parent.

2. Search deterioration remained severe. Online-marketing revenue fell 19% in Q2 after falling 22% in Q1. That narrowly avoids the prior formal test of two consecutive quarters worse than −20%, but the economic conclusion is unchanged: the legacy cash engine is contracting much faster than China’s broad digital-ad market. Baidu App MAU fell to 644 million from 655 million in March and 679 million in December. Management says it is intentionally delaying aggressive monetization of AI-generated search answers to protect user experience; that can be strategically sensible, but it does not change current economics.

3. AI revenue grew, but its quality weakened. AI-powered Business reached RMB12.5B, half of BGB revenue, and grew 25% year over year. It fell 8% sequentially. AI Cloud Infrastructure grew 50% to RMB7.3B and GPU Cloud grew 283%, but AI Applications rose only 3% and AI-native marketing was flat. Management said cloud profit and margin improved year over year, yet it provides no segment revenue reconciliation, gross margin, operating margin, customer retention, or external-versus-internal split. The transition is real; proof that it creates owner earnings is still missing.

4. Cash conversion deteriorated further. Q2 group operating cash flow was RMB3.4B and capex was RMB11.4B, producing negative free cash flow of RMB8.0B. On the Baidu-ex-iQIYI definition, free cash flow was negative RMB8.3B. H1 group free cash flow was negative RMB11.2B. This follows negative RMB15.1B in FY2025. The balance sheet can finance this reinvestment for years, but duration is not return.

5. Apollo’s operating disclosure turned adverse. The earnings release highlighted a 28-city “global footprint” and 350 million autonomous kilometres but omitted quarterly ride count. On the Q2 earnings call, Robin Li said Apollo Go provided approximately 1.0 million fully driverless rides, down from 3.2 million in Q1, after operating adjustments in some domestic cities for regulatory considerations. Some operations had resumed by August. The 28-city figure also mixes commercial operations, tests, and memoranda; it is not 28 scaled commercial markets. Apollo retains valuable data and approvals, but the quarter weakens the utilization narrative.

6. Dual-primary listing became effective. Baidu converted its Hong Kong secondary listing into a primary listing effective September 1, with ordinary shares and Nasdaq ADSs remaining fungible. No new shares or fundraising were involved. The change improves listing redundancy and subjects Baidu to more Hong Kong governance requirements; it does not remove the Cayman/VIE structure, U.S.-China policy risk, or Robin Li’s 59.9% voting control.

The prior bull tests are now mixed. Search stabilization is not confirmed. Cloud revenue growth remains strong, but economics are not confirmed. Annual repurchases remain modest relative to financial assets. Apollo safety and regulatory progress continues internationally, but its Q2 ride decline and missing unit economics move that option further from underwritten value.

1. Executive Summary

Baidu is best understood as three different economic objects sharing one balance sheet: a shrinking but still profitable Chinese search-ad franchise; a rapidly growing, capital-intensive AI infrastructure and applications platform; and a collection of options and adjacencies, principally Apollo, Kunlunxin, and consolidated iQIYI. The accounting presentation obscures the transition because Baidu discloses only two reportable segments—Baidu General Business and iQIYI—and does not provide the margins or capital employed for search, AI cloud, Apollo, or Kunlunxin separately.

The latest operating evidence is poor at the group level. The Q2 filing exhibit shows revenue fell 4% to RMB31.3B. BGB revenue fell 4% to RMB25.2B, online marketing fell 19% to RMB13.1B, and iQIYI fell 5% to RMB6.3B. Group gross margin was 39%, down from 43.9% in FY2025 and 50.3% in FY2024. GAAP operating income was RMB3.0B and BGB operating margin was 12%. AI-powered revenue grew 25% to RMB12.5B, but the faster-growing cloud mix carries higher infrastructure costs and required a step-change in capex. This is revenue substitution with unproven economics, not yet value-accretive growth.

The FY2025 Form 20-F supplies a cleaner warning. Revenue peaked at RMB134.6B in 2023, slipped to RMB133.1B in 2024 and RMB129.1B in 2025. Online marketing fell from RMB81.2B to RMB67.8B over those two years. Reported FY2025 operating income was negative RMB5.8B after a RMB16.2B long-lived-asset impairment; even adding back that noncash charge, operating income was only RMB10.4B, roughly half 2024. The impairment is not a recurring expense, but it is direct evidence that prior search/AI infrastructure investments no longer support their carrying values. FY2025 operating cash flow was negative RMB3.0B and simple free cash flow negative RMB15.1B.

There is nevertheless a genuine operating asset base. Baidu remains the leader in conventional web search in China; StatCounter estimated 58.95% of combined desktop and mobile search-engine page views in August, though the series excludes closed-app discovery and does not measure economic ad share. Baidu’s AI application public-cloud position is credible: IDC placed it first with 30.7% of that specific 2025 market. AI Cloud Infrastructure grew 50% in Q2, token usage rose more than ninefold by management’s measure, and the stack spans Kunlunxin chips, cloud compute, ERNIE models, and enterprise applications. Apollo has accumulated more than 23 million public rides and 350 million autonomous kilometres. These are not imaginary assets. The problem is the absence of disclosed returns.

The balance sheet supplies the counterweight. At June 30, net cash and fixed-income financial assets were about US$20.15B before the long-term equity portfolio, versus a market capitalization of roughly US$32.35B. Shareholders’ equity was about US$40.1B, or approximately US$118 per ADS; a conservative tangible-book calculation after goodwill, acquired intangibles, licensed copyrights and content was roughly US$91 per ADS. The stock therefore trades near tangible book and below stated book. Those figures are not liquidation values: mainland capital controls, VIE claims, minority interests, tax leakage, mark-to-market risk, and reinvestment policy all require discounts. They are, however, too large to ignore.

The strategic conclusion is mixed-to-poor. Search retains a narrow scale and habit advantage, but falling MAU and advertising while competing closed platforms grow indicates that query leadership is not translating into stable economic share. AI cloud sits in a supply boom: Baidu, Alibaba, Tencent, Huawei, ByteDance and state-linked providers are all adding compute, while inference prices fall and open models reduce model-layer pricing power. Apollo owns learning data and approvals but not rider captivity, and peers disclose more commercial detail. iQIYI is a low-margin, hit-driven adjacency. At the consolidated level, neither the share-stability test nor the return-on-capital test supports a durable moat.

The core research question is therefore not whether Baidu has AI capability. It plainly does. It is whether AI revenue can replace search owner earnings before capex and weaker margins consume too much of the financial-asset cushion. The next two to six quarters should be judged on BGB revenue, operating margin, capex, free cash flow, and disclosure quality—not on model benchmarks, token counts, city footprints, or percentage growth from a small GPU-cloud base.

2. Business Overview

Baidu, Inc. is a Cayman Islands holding company. Its ADSs represent Class A ordinary shares; the investor owns a claim on the Cayman parent, not direct equity in all PRC operating entities. Businesses requiring Chinese licenses are conducted partly through variable interest entities controlled by contracts. The structure has operated for years and is consolidated under U.S. GAAP, but contract enforceability, cash movement, and license ownership create a different claim from direct ownership. Each ADS represents eight ordinary shares.

The company reports two segments:

Within BGB, management now separates “AI-powered Business” from “legacy Business” using unaudited internal management records. Q2 AI-powered Business was RMB12.5B, or 50% of BGB, comprising AI Cloud Infrastructure of RMB7.3B, AI Applications of RMB2.5B, and AI-native Marketing Services of RMB2.6B. Legacy Business was RMB10.4B, down 23%. The categories do not perfectly reconcile to BGB and may be revised; they are operating indicators rather than audited segments.

Search and mobile ecosystem. Baidu App, baidu.com, feeds, maps, encyclopedic products, and related distribution connect users with web pages, owned content, merchants, and increasingly direct AI answers. Advertisers historically paid for high-intent clicks and leads. This remains the principal cash engine, but direct answers reduce monetizable outbound clicks, and discovery increasingly occurs inside Weixin, Douyin, Kuaishou, Alibaba, JD, Xiaohongshu, and dedicated AI assistants. Management’s deliberate restraint on AI-search monetization protects experience today at the cost of near-term revenue.

AI cloud and applications. Baidu supplies GPU and Kunlunxin compute, model training and inference, ERNIE foundation models, model-development tooling, and industry applications. The customer proposition is Chinese-language performance plus a domestic, compliant full stack. Applications can become embedded in enterprise workflows and create switching friction. Infrastructure is less differentiated: customers can multi-cloud, models are increasingly open, large peers have much greater procurement scale, and price per token continues to fall.

Kunlunxin. The in-house AI accelerator business is strategically valuable in a market constrained by U.S. export controls and Chinese domestic-substitution policy. Baidu proposed a spin-off and separate Hong Kong listing in January 2026 while expecting to retain control. Offering size, dilution, capital raised, and proceeds to Baidu remain unknown. A listing could create price discovery and external financing; it cannot by itself prove competitive economics.

Apollo and intelligent driving. Apollo supplies autonomous-driving technology, maps, vehicle software, and the Apollo Go robotaxi service. The accumulated ride and kilometre base is a genuine learning asset. The commercial model remains opaque because Baidu does not disclose revenue, active fleet, paid-ride share, rides per vehicle, remote-operator cost, vehicle depreciation, or city contribution margins. Partners and regulators also own important parts of the value chain.

iQIYI. Baidu consolidates the streaming platform but does not own all of it. Revenue declined from RMB31.9B in 2023 to RMB27.3B in 2025, and Q2 returned to an operating loss. Memberships are cancellable, users multi-home, and content must be replenished continuously. The asset adds scale and optionality but little current earnings protection.

Customer value is clear in each business: users receive Chinese-language discovery and answers; advertisers receive intent; enterprises receive compute and applications; riders receive autonomous mobility; viewers receive premium content. Defensibility is uneven. It is highest where enterprise data and workflows become embedded, moderate in search habit and local-scale infrastructure, and lowest in consumer chat, streaming subscriptions, and ride demand.

3. Industry Dynamics

Search and digital advertising

Chinese internet advertising is growing, but its profit pools have migrated toward closed social, video, commerce, and transaction ecosystems. The State Administration for Market Regulation reported RMB1.357T of internet-advertising business revenue in 2025, up 34.6%, with leading platforms growing more than 36%. Definitions are broader than Baidu’s online-marketing line, so this cannot be converted into a defensible market-share estimate. It does establish directional contrast. In the same Q2 2026 period, Tencent’s Marketing Services revenue rose 22% to RMB43.6B while Baidu online marketing fell 19% to RMB13.1B. Tencent cited AI ranking, automated campaign purchasing, and closed-loop Weixin Mini Game and Mini Shop monetization.

Search historically possessed favourable supply-side economics: the fixed cost of crawling, indexing, ranking, moderation, and distribution could be spread across enormous query volume; more advertisers improved coverage; more users and queries improved data. The market is now being redefined. Closed apps prevent general search engines from indexing much useful inventory. Recommendation feeds answer “what should I watch or purchase?” without a query. Generative AI answers “what should I know?” without a click. Commerce platforms observe conversion rather than infer intent. Baidu’s conventional-web leadership therefore covers a narrowing market definition.

Regulation is both barrier and rent cap. A scaled incumbent can spread content moderation, algorithm filing, cybersecurity, personal-data, and generated-content labeling costs. China’s generative-AI rules also constrain training data, answer content, and deployment. Compliance helps incumbents relative to startups but makes product quality and monetization subject to government priorities. It is not a shareholder-owned moat.

AI cloud and the capital cycle

The demand pool is expanding rapidly. IDC estimated China’s 2025 AI Infrastructure-as-a-Service market at RMB48.7B, up 128%, with competition shifting from raw GPU availability toward full-stack compute, scheduling, models, and domestic ecosystems. IDC separately estimated AI-application public cloud at RMB13.7B and model training/inference public cloud at RMB7.9B. These categories overlap and should not be added mechanically.

Supply is expanding at least as dramatically. Baidu’s H1 BGB capex reached roughly RMB17.2B. Alibaba spent nearly US$10B in one quarter and reported AI Cloud & Compute Services revenue of US$7.1B, up 45%, with approximately 12% adjusted EBITA margin. Tencent spent RMB52.8B of capex in Q2, up 176%, and reported negative free cash flow after compute purchases and prepayments. Huawei, ByteDance, and state-affiliated providers add further capacity. Export controls and domestic-chip policy can ration the best processors, but government support also keeps capacity entering when private returns are poor.

This is the central Marathon capital-cycle warning: strong demand does not guarantee strong shareholder returns when all major suppliers expand simultaneously. Compute can become more efficient, token prices can fall faster than utilization rises, and customer workloads can move. Baidu’s full stack and application share are credible advantages; they must be demonstrated in segment margin, utilization, retention, and free cash flow. Revenue growth alone is insufficient.

Autonomous mobility

Robotaxi is also in a capital-deepening expansion phase. Apollo, Pony.ai, WeRide, and automaker- or ride-hailing-affiliated programs are adding vehicles, permits, and cities while group economics remain loss-making or undisclosed. Local government approval creates a barrier to entry, but approvals are jurisdiction-specific and can be tightened. China’s 2023 national trial guide required designated, safety-assessed areas, insurance and operating permission and provided historical context for remote supervision; local pilots may impose different or newer rules.

Pony.ai disclosed Q2 robotaxi revenue of US$12.1M, 1,975 vehicles, and a year-end target above 3,500. WeRide disclosed about 3,400 L4 vehicles, including more than 1,800 robotaxis, and more than 21 daily rides per vehicle. Baidu discloses a larger cumulative ride history but less information about current fleet and economics. Regulation can create incumbent know-how, while municipal promotion can simultaneously sustain uneconomic capacity. The attractive future profit pool may accrue to chips, automakers, or fleet partners rather than the service platform.

Streaming

Long-form streaming is mature, hit-driven, and intensely competitive. iQIYI competes with Tencent Video, Youku, Mango TV, Bilibili, short video, micro-dramas, and games for time and subscription budgets. Original content and recommendation capability can differentiate a slate, but hits expire and subscribers can cancel or multi-home. Content is a recurring reinvestment obligation. Industry structure supports occasional operating profit, not a durable high-return moat.

Industry verdict: structurally mixed-to-poor. Legacy search retains attractive scale economics inside a shrinking boundary. AI cloud and robotaxi have strong demand but simultaneous supply expansion, large capital requirements, and state influence. Streaming has weak captivity. Baidu is migrating from its highest-margin pool into lower-margin pools precisely as competitors also increase supply.

4. Competitive Position

Greenwald’s framework asks whether a company has supply advantages, customer captivity, or both, and whether those advantages show up in stable share and returns. Baidu has pieces of each, but the consolidated evidence fails the test.

Search: narrow scale plus habit, weakening economically. Baidu’s index, query history, language expertise, advertiser base, distribution, maps, and brand lower average cost and improve relevance. Conventional-web leadership remains substantial. Habit is real, but it is not captivity: users can switch engines or begin discovery inside an app at negligible cost; advertisers can allocate budgets across platforms; content owners can favor closed ecosystems. Baidu App MAU fell 5.2% in six months, online marketing declined in both 2024 and 2025 and by 22%/19% in Q1/Q2 2026, and external platform advertising grew. Point-in-time query share does not rescue the economic-share test.

AI cloud: niche application advantage, no broad ecosystem moat. IDC’s 30.7% share estimate in AI-application public cloud supports a meaningful position. Baidu can combine compute, ERNIE, search knowledge, maps, and industry solutions, and embedded workflow applications can create data and integration switching costs. But Alibaba led the separately measured training/inference market with 42.2%; Huawei was second; Baidu was outside IDC’s named top three. Multi-cloud architecture, open models, falling inference prices, and larger competitors prevent a conclusion of broad captivity. No customer retention, cloud margin, or cloud capital-turn disclosure exists.

ERNIE/model layer: capability without demonstrated pricing power. ERNIE reached paid subscriptions in late 2023 but became free in April 2025 as inference prices fell. Baidu has open-sourced parts of its model family, while users can migrate among Doubao, Qwen, Yuanbao, DeepSeek, Kimi, Gemini, and others. Open distribution may stimulate cloud consumption, but the standalone model layer currently looks commoditizing.

Kunlunxin: strategically scarce, economically unproven. Domestic accelerators benefit from export controls and localization mandates. Co-design with Baidu’s workloads can improve utilization and reduce external dependence. Yet chip design has high fixed cost, short product cycles, manufacturing dependence, and competitors backed by larger ecosystems. The proposed listing can expose a market value but is not evidence of sustainable returns.

Apollo: learning curve and approvals, not rider captivity. More than 23 million rides and hundreds of millions of kilometres create a dataset and safety-validation history that new entrants cannot instantly reproduce. City approvals also build institutional know-how. Riders can multi-home, however, and transport partners may own customer access. The Q2 ride reduction shows that regulatory permissions can reduce utilization abruptly. Without intervention frequency, fleet utilization, remote-supervision expense, or contribution margins, a supply-side cost advantage remains a hypothesis.

iQIYI: no durable advantage. Brand, recommendation, and original-content execution matter, but the service lacks structural switching costs and must rebid for attention every season. Q2’s operating loss reinforces that conclusion.

The return evidence is consistent with a narrowing rather than durable moat. One broad-capital proxy, which leaves long-term financial assets in the denominator, fell from about 15.8% in 2023 to 11.6% in 2024 and 3.9% in 2025 after adding back the impairment. A more favourable operating-ROIC proxy excludes all financial assets and yields approximately 8.4% in 2025. The two measures answer different questions; both are group-level estimates and both are below a 15% durable-value-creation threshold. Search may earn much higher returns than AI, but Baidu does not disclose the segment assets or margins needed to prove that.

Competitive-advantage verdict: weak/narrow, not wide. Search has scale and habit; AI applications may have workflow switching costs; Apollo has learning data. Falling MAU, ad revenue, gross margin, cash conversion, and return on capital outweigh those strengths at the consolidated level.

5. Growth History and Forward Opportunities

Growth history separates Baidu’s technical progress from its economic progress.

The growth opportunity with the highest probability is AI cloud. Q2 Infrastructure revenue of RMB7.3B grew 50%; GPU Cloud grew 283%; management said token consumption rose more than ninefold. Domestic demand for compliant compute, inference, and workflow applications is large. Baidu’s search data, PaddlePaddle tools, ERNIE models, Kunlunxin hardware, and industry relationships allow it to offer a full stack. The opportunity becomes high quality only if revenue drives utilization and margin faster than capex, and if applications—not just raw compute—become a larger mix.

AI-native marketing is a second opportunity but currently supplies no growth. Generative creative, automated bidding, better merchant onboarding, and direct-answer commercial formats could improve conversion while reducing advertiser effort. Management is delaying aggressive monetization to protect search experience. That creates a plausible J-curve: near-term ad pressure in exchange for a better product and later formats. It also creates an unfalsifiable story unless engagement stabilizes. The practical tests are Baidu App MAU, online-marketing trend, advertiser return on spend, and BGB margin.

AI applications contributed RMB2.5B in Q2 and grew 3%. Workflow builders, document intelligence, digital humans, and industry software can create higher-quality recurring revenue than infrastructure because they embed in workflows. Current growth does not support that aspiration. Customer count, renewal rate, remaining performance obligations, and application gross margin are not disclosed.

Kunlunxin offers scarcity and external-price-discovery optionality. A listed, separately financed chip company could attract capital without requiring Baidu to bear all expansion. It may also expose a value not recognized inside the group. The offset is dilution, manufacturing concentration, cyclical inventory risk, and the possibility that strategic demand is driven by policy rather than superior economics. Until offering terms and audited segment numbers exist, it remains an option.

Apollo’s opportunity is large in principle: lower driver cost, higher vehicle utilization, autonomous-fleet software, and licensing to automakers or transport platforms. Its 23-million-plus ride history is meaningful, and international testing broadens the data set. The Q2 drop to roughly 1.0 million fully driverless rides shows that deployment is not monotonically scalable. The value-creating path requires paid demand, high rides per vehicle, lower remote-intervention cost, durable permits, and positive contribution margin. City count and autonomous kilometres are intermediate outputs, not owner earnings.

iQIYI could stabilize through better content efficiency, premium pricing, overseas distribution, and micro-drama formats, but its three-year revenue decline and Q2 loss place it outside the main growth thesis. It should be assessed as an exposure that may fund itself, not as a growth engine.

Growth-quality verdict: low today. AI infrastructure is growing quickly but replacing higher-margin search revenue with capital-intensive revenue. Applications and AI-native marketing are near flat; Apollo lacks commercial economics; iQIYI is shrinking. A genuine inflection requires simultaneous BGB revenue growth, stable or improving margin, and positive post-capex cash flow.

6. Financial Quality

The five-year statements show an earnings peak followed by a mix and cash-conversion break.

Revenue is not the main problem; the quality and capital intensity of revenue are. The 7.8-point gross-margin decline from 2023 to 2025 tracks the replacement of high-margin search ads with AI compute, traffic-acquisition costs, and content. Q2’s 39% gross margin extended the trend. BGB’s 12% operating margin may look stable sequentially, but no allocation reveals whether search is subsidizing cloud, Apollo, or Kunlunxin.

The RMB16.2B FY2025 long-lived-asset impairment requires careful treatment. It was noncash and specific to the Baidu Core asset group, not goodwill or the investment portfolio. Adding it back converts a GAAP operating loss into RMB10.4B of adjusted operating profit. But management recorded the charge only after undiscounted expected cash flows fell below carrying value; the discounted fair value was RMB44.7B using approximately a 13% discount rate and management revenue-growth assumptions of 8%–18%. The charge is nonrecurring in the income statement and recurring in its message: prior capital did not earn the expected cash flows.

Net income is a weak guide to core earnings. FY2025 other income was RMB12.5B, including RMB8.6B of interest income, RMB3.2B of equity-method income, interest expense, foreign-exchange losses, and other investment-related items. These financial assets are economically valuable, but their returns and marks can mask operating deterioration. Operating profit and post-capex cash are better measures of the transition.

Cash conversion was poor for reasons beyond impairment. FY2025 operating cash flow was negative despite the noncash charge and RMB3.6B of stock compensation. Receivables, content, other operating assets, and long-term supplier advances used cash; supplier advances more than doubled to RMB18.0B. H1 2026 operating cash flow recovered to RMB6.1B, but capex reached RMB17.3B. Q2 capex alone was roughly 36% of group revenue and 45% of BGB revenue. The earnings release’s positive operating-cash-flow framing therefore describes only the numerator before the largest current reinvestment item.

Research and development expense fell from RMB24.9B in 2021 to RMB20.4B in 2025, or from 20.0% to 15.8% of revenue. Stock compensation fell from RMB6.3B in 2023 to RMB3.6B in 2025, and diluted ordinary shares declined from 2.837B to 2.744B. Lower dilution is positive. The simultaneous decline in expensed R&D and surge in physical capex shows the investment mix moving from people and software toward compute and infrastructure, increasing fixed-cost and obsolescence risk.

Balance-sheet reconstruction

The balance sheet also contains US$3.44B of redeemable and permanent non-controlling interests. These claims should be deducted before treating consolidated assets as belonging to Baidu common shareholders. At year-end 2025, 63% of cash, deposits, and debt investments were held at mainland institutions. Baidu has demonstrated some ability to upstream capital—PRC subsidiaries distributed RMB24.7B to Baidu Hong Kong over 2023–2025—but conversion and dividends remain regulated.

Reported Baidu-attributable equity was approximately US$40.05B, or US$118 per ADS. Subtracting goodwill and acquired intangibles produces tangible equity of approximately US$33.89B, or US$100 per ADS; subtracting licensed copyrights and content as well produces the more conservative approximately US$91 figure used in the opinion block. Neither is liquidation value. Goodwill of RMB36.8B, much of it increased by YY Live, remains exposed to impairment.

Financial-quality verdict: deteriorating. The balance sheet is exceptionally large relative to the capitalization, but revenue mix, gross margin, operating profit, working capital, and free cash flow all weakened. Financial resilience is high; operating earnings quality is low.

7. Capital Allocation

Baidu’s capital-allocation record is mixed to below average. Management has kept the company financially resilient and reduced dilution, but returns on internal reinvestment and acquisition logic do not justify full credit for every balance-sheet dollar.

Reinvestment. The most important current allocation is AI infrastructure. Capex rose from RMB8.1B in 2024 to RMB12.1B in 2025 and RMB17.3B in H1 2026 alone. In a supply-constrained market, early capacity can establish customers and lower unit cost. In a capital boom, the same spending can earn poor returns as competitors add capacity and token prices fall. Baidu discloses no hurdle rate, cloud return on invested capital, capacity utilization, or payback. The burden of proof rises because group-level returns have fallen below a reasonable cost of capital.

Repurchases. Cash spent on repurchases was RMB4.8B in 2023, RMB6.3B in 2024, and RMB5.5B in 2025. Diluted ADS equivalents declined from about 354.6 million in 2023 to 343.0 million in 2025, so the program has produced real per-share benefit. Only about US$2.3B–2.4B was used under the prior US$5B authorization, however. A new US$5B program runs through 2028, and only US$259M had been deployed since Q1 by the Q2 report. That is modest against US$20B of net fixed-income assets and a market capitalization below book.

Dividend. The board adopted Baidu’s first dividend policy in February 2026 and said it expected to declare the first payment during 2026, subject to discretion. No dividend had been declared as of September 1. The policy is directionally favourable because it creates a recurring pathway from mainland-generated value to public shareholders. Its size, sustainability, tax treatment, and funding location remain unknown.

Acquisitions. The reduced YY Live transaction closed in February 2025 for about US$2.1B after years of regulatory delay. It added a mature live-streaming asset as search weakened and AI capex accelerated. Goodwill rose from RMB22.6B to RMB36.8B and acquired intangibles increased. The purchase may supply users or cash flows, but strategic coherence is weak, and the accounting creates future impairment risk. That decision deserves a lower capital-allocation grade than simply preserving cash.

Financing. Baidu issued RMB10B of notes and US$2B of zero-coupon exchangeable bonds in March 2025, followed by RMB4.4B of notes in September. Cheap debt alongside financial assets can be rational when maturities, currencies, and parent-level liquidity differ. It also increases gross leverage and complexity while free cash flow is negative. The correct valuation treatment is to net all debt, not celebrate gross cash.

Kunlunxin separation. A spin-off can match capital needs to the chip subsidiary, expose external valuation, and discipline reporting. It can also dilute Baidu’s ownership and divert an attractive asset into a structure where proceeds do not reach common shareholders. No offering size, ownership percentage, or use of proceeds was final at the report date. No private-market headline value is credited in this report.

Governance and incentives. Robin Li beneficially owned 18.6% of ordinary shares and controlled 59.9% of voting power at January 31, 2026. Class B shares carry ten votes for each Class A vote. Founder control supports long-horizon investment but limits outside-owner influence. As a foreign private issuer, Baidu discloses aggregate executive compensation rather than a U.S.-style named-executive pay-for-performance table and is exempt from standard Section 16 reporting. The 2026 incentive plan can issue up to 10% of shares, with performance conditions discretionary. The observed voluntary insider tape contains awards and sales but no open-market insider purchase.

Capital-allocation verdict: mixed to below average. Reduced share count, falling stock compensation, a dividend policy, and possible Kunlunxin price discovery are positives. Low returns, restrained repurchases, rising capital intensity, new gross debt, YY Live, and founder-controlled governance justify significant asset-value haircuts.

8. Changes and Headwinds — Last Two Years

The last two years mark a structural transition rather than a normal advertising cycle.

Search monetization broke faster than usage leadership. Online marketing fell 3% in 2024, 14% in 2025, 22% in Q1 2026, and 19% in Q2. Baidu App MAU began declining into 2026. Management cites macro pressure and product transition; stronger same-period advertising growth at Tencent and in official market statistics is consistent with platform and format share loss, although public data cannot quantify the causal split. AI answers may improve utility while reducing paid clicks, forcing Baidu to invent new commercial formats without compromising trust.

The income statement absorbed a lower-margin mix. AI-powered Business reached half of BGB, but gross margin moved sharply lower and the filing explicitly warns that generative AI, cloud, and intelligent driving carry margins much lower than online marketing. Cost of revenue rose in Q2 because of AI-cloud costs even as total revenue declined. This is the central headwind: replacing a dollar of advertising with a dollar of GPU infrastructure is not economic neutrality.

The cash-flow regime reversed. Simple free cash flow moved from RMB25.4B in 2023 to RMB13.1B in 2024, negative RMB15.1B in 2025, and negative RMB11.2B in H1 2026. Some spending may be growth capex rather than maintenance, so current free cash flow understates steady-state earnings if projects succeed. It accurately measures what remains for owners today and how quickly the financial cushion can be redeployed.

Asset quality was reset. The RMB16.2B Core long-lived-asset impairment acknowledged weaker future cash flows. Goodwill then increased through YY Live. The sequence—impair operating assets, acquire a mature adjacent asset, and accelerate compute spending—raises the standard for management credibility.

AI growth decelerated sequentially. Q1 AI-powered Business grew 49% to RMB13.6B; Q2 grew 25% to RMB12.5B. Cloud Infrastructure fell 17% sequentially despite 50% year-over-year growth. Seasonality and project timing may explain part of the change, but the trajectory demonstrates that headline growth rates are volatile. Applications at +3% and AI-native marketing at zero growth are not yet absorbing infrastructure economics.

Apollo experienced a regulatory utilization shock. Q4 2025 and Q1 2026 fully driverless rides were 3.4 million and 3.2 million. Q2 fell to about 1.0 million by management commentary. Operational adjustments in domestic cities may be temporary and some operations resumed in August. The episode proves that accumulated approvals are not irrevocable and that footprint expansion can coexist with lower ride volume.

Geopolitical risk became more specific. Baidu disclosed that it had been added to the U.S. Defense Department’s Section 1260H Chinese Military Companies list in June 2026 and said the designation should have no material impact. The designation is not the Treasury sanctions list and does not currently prohibit U.S. persons from trading BIDU. It nonetheless raises escalation, procurement, customer, and capital-access risks and supports a higher required return.

Listing redundancy improved. Dual-primary Hong Kong status widens the potential investor base and offers an additional venue if U.S. access becomes impaired. Fungibility helps price linkage. It does not change the underlying issuer, VIE contracts, or political exposure.

The two-year scorecard is therefore adverse on earnings quality and favourable on financial optionality. Baidu owns more recognized financial backing than the prior analysis credited, but the operating transition has become more expensive and less transparent.

9. Risk Analysis

Search risk is existential to near-term earnings. A rapid shift to AI answers can improve engagement while breaking the click auction. Closed ecosystems own first-party commerce and conversion data. Baidu may preserve traffic yet lose monetization. The key warning is divergence: stable conventional-web share alongside falling ad revenue would mean the measured moat protects the wrong unit of economics.

AI capital-cycle risk is the largest balance-sheet risk. Baidu can fund years of spending, which may encourage supply even if returns remain poor. Hardware becomes obsolete, customer bargaining power rises as alternatives multiply, and inference efficiency lowers revenue per unit of work. A 50% cloud-growth rate can coexist with value destruction if pricing and utilization lag depreciation and financing costs.

Structural ownership risk affects asset value. The Cayman parent relies on contracts for restricted PRC businesses. Regulators could challenge those contracts or limit data, licensing, remittance, and listing arrangements. VIE creditors do not have recourse to the parent in the same way, and common shareholders do not directly own licenses. The structure has precedent and the dual-primary listing adds redundancy, but neither eliminates tail risk.

Geopolitical risk is asymmetric. Current CMC designation does not bar securities trading. Escalation to investment restrictions, tighter export controls, customer bans, or index exclusion would impair both operations and the investor base. Conversely, removal from the list might narrow the discount without improving operations. The valuation therefore embeds this through asset haircuts and a higher capitalisation rate rather than adding a second arbitrary discount.

Apollo risk is nonlinear. A serious safety incident or regulatory suspension can erase utilization and slow permits across cities. Management claims a strong safety record, including one airbag deployment per 14.4 million kilometres; that is a management statistic, not an independently audited loss ratio. Remote supervision, insurance, vehicle depreciation, and idle time may keep unit economics unattractive even without incidents.

Accounting and disclosure risk is moderate. U.S. GAAP statements are audited, but interim disclosure arrives on Form 6-K, operating categories are internally defined and unaudited, segment detail is sparse, and FPI exemptions limit compensation and insider-trading visibility. The most important uncertainties—cloud margin, Apollo economics, and Kunlunxin financials—are exactly where optionality is claimed.

Downside is not synonymous with zero. Even under stress Baidu owns financial assets, search infrastructure, intellectual property, investments, and listed stakes. The relevant bear mechanism is prolonged reinvestment and controlled-capital leakage that converts liquid financial backing into low-return operating assets before shareholders receive it.

10. Valuation Discussion — Assets, EPV and Embedded Expectations

This section asks what the current capitalization requires, using three independent lenses and explicit haircuts.

Current capitalization and accounting multiples

At US$95.28 and approximately 339.5 million ADS equivalents, market capitalization is US$32.35B. Trailing revenue through June was approximately US$18.76B, producing 1.72× price-to-sales. GAAP trailing operating income and net income were negative, so GAAP earnings multiples are not meaningful. Adding back only the disclosed RMB16.2B impairment produces trailing EBIT of RMB8.8B, a 6.9% margin. Baidu trades at approximately 0.81× stated book and 0.95× tangible book before subtracting licensed content.

Peer multiples are weak anchors. Alibaba trades near 1.8× sales and 1.7× book but has different commerce and cloud economics; Tencent trades above 4× sales and nearly 3× book because its engagement, growth, margins, and cash conversion are superior; JD trades near 0.2× sales and 0.9× book because retail revenue has thin margins. BIDU’s proper anchor is its own financial assets plus normalized operating earning power, not a blended China-internet multiple.

Asset reproduction and haircut logic

A conservative adjusted-asset test credits 80% of US$20.15B net cash/fixed-income assets, 50% of US$6.32B long-term equity investments, deducts all US$3.44B of non-controlling interests, and credits 65% of residual parent-attributable tangible operating equity. That produces approximately US$22.9B of adjusted asset coverage, or 71% of current capitalization. The discount reflects mainland transfer friction, maturity, investment opacity, minority claims, and uncertain recovery of operating assets. It does not assume the business disappears.

Using only recognized financial claims, central parent financial-asset credit is approximately US$15.84B. Baidu’s quoted 45.1% economic stake in iQIYI was worth only about US$0.41B at late-August market prices. Subtracting both from current capitalization leaves approximately US$16.1B assigned to all of BGB: search, AI cloud and applications, Apollo, Kunlunxin, maps, and corporate costs. This is the cleanest sum-of-the-parts observation because it does not invent private values for the options.

Earnings-power value

Capitalizing impairment-adjusted trailing operating profit after the FY2025 effective tax rate at 12% produces approximately US$8.8B of accounting earnings-power value for consolidated operations. Adding full parent-claim net financial assets yields approximately US$31.8B of coverage, close to the current capitalization. Applying central asset haircuts lowers coverage to approximately US$24.6B. A maintenance-sensitive method that subtracts FY2025 capex above depreciation lowers operating earnings power toward US$5B. H1 2026 cash conversion would justify less still.

The spread between US$5B and US$8.8B is not modelling noise; it is the thesis. If current capex is mostly temporary growth investment with future high utilization, accounting EBIT understates normalized value. If it is the recurring cost of competing in AI, accounting EBIT overstates owner earnings.

Embedded expectations

Asset-credit choice is the load-bearing variable. Crediting all parent financial assets leaves about US$9.3B for operations. At a 12% capitalisation rate and the FY2025 tax rate, that requires roughly RMB9.3B of zero-growth EBIT, or a 7.3% margin on trailing revenue—close to the impairment-adjusted 6.9%. Under central asset haircuts, the operating stub rises to about US$16.5B and requires RMB16.5B of EBIT, or a 13.0% margin. Under harsher haircuts, the required margin reaches roughly 15.8%.

The market is therefore pricing one of three outcomes: financial assets deserve relatively full credit; operating margin recovers materially; or Apollo/Kunlunxin possess significant option value. It is not pricing both a worthless operating company and inaccessible assets—the common “negative enterprise value” shortcut is incorrect after debt, minority interests, and haircuts.

Scenario coverage of the current capitalization

These are assumption sets, not forecasts. The base case fails to cover current capitalization because its 8% margin recovery does not compensate for asset haircuts and dilution. The bull case requires search stabilization, profitable AI mix, and free-cash-flow recovery; the bear case assumes prolonged ad decline and weak returns on cloud capex. Apollo and Kunlunxin receive no discrete private-market value in any case, so successful external price discovery would add evidence—but not automatically distributable cash.

Valuation verdict: statistically inexpensive with genuine asset backing, but not a simple net-cash arbitrage. Full-credit assets plus normalized accounting earnings roughly reconcile to the market value. Reasonable risk haircuts require a material operating recovery that has not begun. The security’s apparent cheapness and the business’s negative free cash flow are both true.

11. Variant Perception

The common bull narrative is that Baidu is China’s foundational AI company hidden inside a cheap search business: ERNIE and Kunlunxin create a sovereign full stack, AI cloud compounds rapidly, Apollo is the world’s scaled robotaxi leader, and the balance sheet makes downside small. The common bear narrative is that search is being disintermediated, Baidu arrived late to consumer AI, cloud is subscale, Apollo is a cash-burning science project, and China/VIE assets deserve little credit.

The evidence supports neither narrative in full.

Variant 1: the asset cushion is larger than it looks in common data services. Long-term time deposits and held-to-maturity securities are often grouped with long-term investments. Separating them reveals US$20.15B of net cash/fixed-income financial assets before a US$6.32B equity portfolio. The correction matters because many headline enterprise-value calculations also contaminate ADR share counts or mix currencies. Even after minority claims and central haircuts, financial-asset credit remains close to half the market capitalization. That is a fact-based source of asymmetry.

Variant 2: search leadership is not the same as an ad moat. Baidu can remain the largest conventional web search engine while losing economic relevance. Closed apps, commerce platforms, recommendation feeds, and direct-answer AI sit outside many search-share datasets. The decisive evidence is not StatCounter’s 58.95%; it is Baidu’s declining MAU and advertising against stronger peer advertising. Search has a narrowing scale advantage, not stable customer captivity.

Variant 3: the AI mix shift is real and still economically negative. Skeptics who dismiss Baidu’s AI revenue overlook RMB12.5B of quarterly AI-powered Business and 50% growth in Cloud Infrastructure. Enthusiasts who treat it as equivalent to lost ad revenue overlook lower margins, H1 capex of RMB17.3B, and negative free cash flow. The correct view is a technically credible transition whose return on capital is unproven. The next edge will come from economics, not benchmark leadership.

Variant 4: Apollo’s cumulative scale is less informative than current utilization. More than 23 million cumulative rides and 350 million kilometres create learning value, but Q2 rides fell roughly 69% sequentially after regulatory adjustments. Baidu’s “28-city footprint” mixes operations, tests, and exploratory agreements. Commercial value depends on paid rides per active vehicle after remote supervision and depreciation. Those figures are absent. Apollo is a genuine option with a zero in the underwritten sum-of-the-parts, not because it is worthless but because public evidence cannot price it responsibly.

Variant 5: the balance sheet is not a floor unless capital allocation protects it. Financial assets can be upstreamed, repurchased, distributed, acquired away, or converted into compute. Repurchases have reduced shares, and the new dividend policy is constructive. YY Live, modest deployment relative to authorization, founder control, and negative free cash flow argue against full credit. The real downside path is not sudden disappearance of cash; it is a multi-year transfer from financial assets into low-return operating assets.

Positioning variant. BIDU’s market behavior resembles China-internet beta more than a global AI-quality compounder. The factor model gives it negative momentum and near-zero quality exposure; factor-nearest instruments include KWEB, CWEB, BABA, MCHI, and JD. Short interest at 3.92% of float and 4.63 days to cover is elevated from late July but not crowded. The market is skeptical without being positioned for a forced squeeze. Any rerating likely requires a company or policy catalyst rather than passive factor support.

The synthesis is a security with a better balance-sheet setup than the headline narrative and a worse operating-quality setup than the AI narrative. That tension—not a claim that the market has overlooked artificial intelligence—is the variant perception.

12. Fact vs. Interpretation

Two distinctions are especially important. First, “noncash” describes the accounting mechanics of the impairment, not the absence of economic loss. Second, “liquid net assets” should not imply every renminbi is immediately accessible to the Cayman parent. The report uses “net cash/fixed-income financial assets” and explicitly haircuts them for this reason.

13. Open Questions

1. What are AI Cloud’s economics? Baidu should disclose external revenue, gross or operating margin, customer retention, capacity utilization, and capex attributable to cloud. Management’s statement that cloud profit and margin improved year over year is useful but not sufficient to estimate returns.

2. How will AI search monetize without damaging trust? Management is intentionally prioritizing user experience. The unknowns are commercial-query coverage, click-through behavior, answer-ad formats, advertiser conversion, and whether direct answers expand or shrink the monetizable funnel.

3. Is the MAU decline temporary or structural? The fall from 679 million to 644 million occurred over six months. Quarterly users, time spent, query volume, AI-answer engagement, and retention cohorts would distinguish product transition from audience loss.

4. How much of current capex is growth versus maintenance? A useful bridge would separate data-center construction, servers, Kunlunxin, replacement cycles, customer-specific capacity, and prepayments. Without it, normalized owner earnings range widely.

5. Where are the financial assets and what can reach common shareholders? The 20-F gives geographic concentration and historical upstream distributions, but not a current parent-versus-subsidiary maturity ladder. Offshore cash, pledged assets, tax leakage, and remittance approvals determine realizable value.

6. What will the dividend policy actually return? No amount, payout framework, record date, or funding source had been declared. A recurring distribution funded at the parent would materially improve confidence in asset fungibility.

7. What are Kunlunxin’s audited financials and spin economics? Revenue, customers, gross margin, R&D, capex, ownership after offering, proceeds, related-party sales, foundry exposure, and governance are unknown. The listing should not be capitalized before these are available.

8. What happened city by city at Apollo in Q2? Management cited regulatory operating adjustments and later resumptions. Investors need active fleet, rides per vehicle, paid-ride mix, remote-operator ratios, intervention frequency, insurance, depreciation, and contribution profit for at least one mature city.

9. Does YY Live generate cash sufficient to justify its goodwill? Standalone revenue, margin, retention, cash flow, integration synergies, and impairment headroom are not disclosed in decision-useful form.

10. How will dual-primary status change behavior? The conversion adds Hong Kong requirements and redundancy. It remains to be seen whether it expands the shareholder base, increases local disclosure, improves capital returns, or merely changes compliance form.

11. What is the path off the 1260H list? Baidu has said the designation should not have a material impact. Procurement exposure, customer reactions, legal recourse, and the risk of migration to more restrictive U.S. lists require monitoring.

12. Which return-on-capital denominator is operationally correct? Public reporting cannot allocate search, cloud, Apollo, chip, content, and financial assets cleanly. Segment balance sheets and after-tax returns would decide whether a valuable search core is being obscured or consumed.

14. What Must Be True (Bull and Bear, each with a falsification test)

Bull case

Search stabilizes rather than disappears. Baidu App MAU stops declining, online marketing improves to better than −10% by Q4 2026, and returns to growth during 2027 as AI-search formats monetize without impairing engagement. Falsification: MAU falls below 620 million or advertising declines at least 15% for another two quarters while broader platform advertising grows.

AI growth becomes value-creating. AI-powered Business sustains at least 25% growth, BGB revenue turns positive, BGB operating margin holds at or above 12%, and trailing BGB free cash flow turns positive by the end of 2027. Applications and AI-native marketing begin growing faster than infrastructure. Falsification: AI growth falls below 20% while BGB remains negative, or capex stays above 30% of BGB revenue without disclosed cloud profitability and retention.

Financial assets remain available to owners. Net financial assets remain broadly stable after funding the transition, repurchases materially exceed dilution, and the dividend becomes a recurring parent-funded distribution. Falsification: net financial assets decline by more than 20% without a commensurate increase in profitable operating earnings, or large acquisitions replace distributions.

Apollo demonstrates commercial—not only technical—scale. Baidu discloses fleet, paid rides, utilization, and contribution economics; at least one scaled city turns positive during 2027 while safety metrics remain acceptable. Falsification: commercial disclosure remains absent through Q1 2027, another material safety/regulatory suspension occurs, or fleet utilization materially trails Pony.ai and WeRide.

Kunlunxin separation creates transparent value. The offering supplies audited economics, continued control, sensible dilution, and capital directed to high-return growth rather than covering operating deficits. Falsification: related-party dependence, low margins, severe dilution, or manufacturing constraints reveal that the strategic asset lacks standalone economics.

Bear case

Search economics decay faster than AI can replace them. Advertising remains down double digits, MAU contracts, and AI direct answers reduce clicks without new formats. Gross margin stays near or below 40%. Falsification of the bear case: online advertising grows at least in line with the Chinese platform market and BGB margin expands for four consecutive quarters.

AI infrastructure enters a return-compression cycle. Capacity growth, open models, and efficiency improvements push pricing down; Baidu sustains high capex to maintain share, and cloud growth does not produce free cash flow. Falsification of the bear case: Baidu retains more than 25% AI-application cloud share, enters the top three in training/inference, discloses positive segment margins, and earns after-tax ROIC above 15% with cash conversion.

The asset floor leaks. Cash is committed to acquisitions, chips, data centers, or VIE funding while repurchases and dividends remain small. Capital controls and minority claims prevent public owners from receiving full value. Falsification of the bear case: multi-year distributions and repurchases demonstrate that offshore common shareholders receive a material portion of financial assets without weakening operations.

Apollo and Kunlunxin remain expensive options. Neither produces transparent earnings; regulatory, safety, manufacturing, or dilution events reduce their value while Baidu continues financing them. Falsification of the bear case: audited standalone economics and external transactions establish profitable, transferable value without requiring continuing BGB subsidy.

Jurisdictional risk widens. CMC status escalates, export controls tighten, or VIE/capital-market rules impair access. Falsification of the bear case: removal from the list, stable PCAOB access, durable listing fungibility, and continued capital return narrow the structural discount.

The bull and bear cases share one decisive monitor: post-capex cash generation at BGB. If AI revenue grows while that figure stays negative, the transition is building technical capability without yet creating distributable value. If BGB revenue, margin, and free cash flow improve together, the current weak-moat assessment deserves revision.

15. Source Appendix

Primary company evidence takes precedence; third-party sources are identified explicitly.

1. Baidu FY2025 Form 20-F, filed March 17, 2026 — audited statements, VIE structure, risks, ownership, compensation, investments and impairment.
2. Baidu Q2 2026 results, SEC Exhibit 99.1, August 18, 2026 — interim financials, AI metrics, balance sheet and cash flow.
3. Baidu Q1 2026 results, SEC Exhibit 99.1, May 18, 2026 — Q1 operating metrics and cash flow.
4. Baidu FY2025/Q4 results, February 26, 2026 — annual operating metrics and segment cash flow.
5. Baidu Q2 2026 earnings-call event record, August 18, 2026 — management discussion, including Apollo rides and cloud commentary.
6. Repurchase program and dividend policy, February 2026.
7. Kunlunxin proposed spin-off, January 1, 2026.
8. Dual-primary listing effectiveness, September 1, 2026.
9. Baidu response to Section 1260H designation, June 9, 2026.
10. China 2025 advertising-industry release, State Administration for Market Regulation, April 7, 2026.
11. China search-engine page-view share, StatCounter, August 2026 observation.
12. China AI application and model public-cloud analysis, IDC, May 18, 2026.
13. China AI IaaS market overview, IDC, May 29, 2026.
14. Tencent Q2 2026 results, August 12, 2026 — advertising and capex comparison.
15. Alibaba Q2 2026 results, August 20, 2026 — cloud scale, margin and capex comparison.
16. iQIYI Q2 2026 results, August 18, 2026.
17. Pony.ai Q2 2026 results, August 18, 2026, and WeRide Q2 2026 results, August 12, 2026.
18. Hong Kong autonomous-vehicle pilot update, July 23, 2026.
19. CAC generative-AI interim measures and generated-content labeling measures.
20. PCAOB China audit-access fact sheet, December 15, 2022.
21. AZI Trading adjusted BIDU daily history, through August 31, 2026, and FactorsToday BIDU factor-model data, dated July/August 2026 — third-party market-data CSV and statistical API.
22. Baidu Q2 2022 results and Baidu Q3 2024 results — company earnings releases used to cross-check the five-year event map.

Data limitations. Baidu is a foreign private issuer and reports interim results on Form 6-K rather than Form 10-Q. AI-powered Business categories are unaudited management records. No public segment economics exist for search, AI cloud, Apollo, or Kunlunxin. Factor and price data are statistical and do not establish causality. All currency translations use the Q2 filing’s RMB6.7851 per U.S. dollar unless stated otherwise.