Baidu

Phase A — Understand the business

Lens 1 · Company Overview

Baidu, Inc. (Cayman holdco; ADS on Nasdaq = 8 Class A shares; dual-primary on HKEX 9888) is China's dominant search engine pivoting hard into a full-stack AI company. It reports two segments: Baidu General Business (renamed from "Baidu Core" in Q4 2025) and iQIYI (majority-owned streaming sub).

FY2025 revenue RMB129.1B (US$18.5B), −3% YoY. The mix splits two ways:

• By segment: Baidu General Business RMB102.5B (−2%, ~79% of total) + iQIYI RMB27.3B (−7%).
• By type (the real story): Online marketing (ads) RMB67.8B vs Others RMB61.2B. Ads are now barely half of revenue and shrinking; "Others" (cloud + iQIYI membership) is +12% and closing in.

Three business areas inside General Business:

1. Mobile Ecosystem — Baidu App (search-plus-feed, 679M MAU Dec 2025), Haokan, Baidu Post, ERNIE Bot. The legacy P4P cost-per-click ad engine lives here.
2. AI Cloud — full four-layer stack: AI Cloud Infrastructure (training/inference compute) + AI Applications (Wenku, Drive, Digital Employee, on subscription). "No.1 AI public cloud in China six consecutive years" per IDC.
3. Intelligent Driving & Other Growth Initiatives — Apollo Go robotaxi + DuerOS/Xiaodu smart devices.

Customers/contract structure: SME-heavy ad base via third-party agents + direct sales (healthcare, retail, e-commerce, games). No single customer >10% of revenue in any year presented — low customer-concentration risk. Ads are cost-per-click (no take-or-pay); cloud is subscription/consumption; iQIYI is recurring membership. >96% of revenue from Chinese mainland.

Bottom line: This is a holdco where a still-huge, still-profitable but structurally declining search-ad utility funds a three-pronged AI bet. The investable question is entirely about the transition, not the current P&L.

Lens 2 · Supply Chain

Baidu sits at an unusual node: it is both a buyer of compute and an emerging maker of it. Named stakeholders along the chain:

Upstream (compute & infra):

• AI accelerators — historically Nvidia (now export-restricted into China); Baidu is vertically integrating with its own Kunlun chips (M100 launching 2026, M300 2027); a 30,000-chip Kunlun training cluster came online April 2025. JPMorgan models Baidu chip sales 6× to ~RMB8B (US$1.1B) in 2026.
• Deep-learning framework — in-house PaddlePaddle 3.0 (the CUDA-substitute layer that makes Kunlun usable without Nvidia's ecosystem).
• Data centers / bandwidth — servers hosted at China Telecom, China Unicom, China Mobile IDCs across 10+ cities; completed third-phase cloud-compute-center build in 2025.

Midstream (the platform): ERNIE foundation models (5.0 omni-modal, Nov 2025; updated Jan 2026) → Qianfan MaaS platform → applications.

Downstream (demand):

• Ad customers — millions of SMEs + Baidu Union partner sites (third-party websites/apps that embed Baidu search, revenue-shared).
• Cloud customers — enterprises + public sector across manufacturing, energy/utilities, financial services, internet/media.
• Robotaxi — direct consumers via the Luobokuaipao app, plus Uber and Lyft as distribution partners internationally (deploy Apollo Go AVs on their platforms across Middle East, Asia, Europe).

Chokepoints / single-source dependencies:

• Advanced-node fabrication for Kunlun. Baidu designs chips but does not fabricate leading-edge silicon; China foundry capacity (SMIC) under US export pressure is the real bottleneck — the 20-F does not name a foundry, and that opacity is the risk.
• Nvidia-gap dependency runs both ways: export controls hurt Baidu-the-buyer but create the TAM for Baidu-the-maker and for its GPU Cloud (+184% YoY in Q1-26). China's GPU cloud is consolidating around Baidu + Huawei.

Names present, so the lens holds. The distinctive feature: Baidu is one of the few names where Nvidia export controls are net-ambiguous rather than purely negative.

Lens 3 · Competitive Advantages (moats)

Where the moat is real:

• Search distribution + 679M-MAU funnel — still the default information-retrieval surface in China, now being rebuilt AI-native (≈70% of mobile search result pages carried AI-generated content by Oct 2025). This is a genuine distribution moat, but it is eroding as the query monetization model changes (more on this in the bear case).
• Full four-layer AI stack — Baidu is "one of very few companies in the world" owning chip → framework → model → app. Vertical integration (Kunlun + PaddlePaddle + ERNIE + Qianfan) lets it sidestep CUDA and offer cost-controlled domestic compute. In an export-controlled China, being your own Nvidia + your own OpenAI is a structural advantage few peers match (only Huawei and Alibaba are comparable).
• Apollo Go data + regulatory moat — first driverless licenses in China and the US; 100% fully-driverless ops across 8+ China cities; 20M+ cumulative rides (Feb 2026) generating the largest real-world L4 dataset of any single operator by ride volume. Regulatory permits are slow, city-by-city, and incumbency-advantaged — a real barrier.
• Cost-leadership in inference — ERNIE 4.5 reportedly beats DeepSeek V3 on 22/28 benchmarks at ~half the parameters, i.e. cheaper to serve. Efficiency is the China-AI battleground.

Bargaining power:

• Over ad customers: weakening — SMEs have substitutes (Tencent, ByteDance/Douyin, Alibaba) and macro is squeezing budgets; Baidu cannot raise ad prices into a demand decline.
• Over cloud customers: moderate and improving — full-stack lock-in (once an enterprise builds agents on Qianfan + Kunlun, switching is costly), but it competes head-on with Alibaba Cloud and Huawei Cloud on price.
• Over compute suppliers: improving via Kunlun self-supply.

Net: The moat is bifurcating — the old search-ad moat is weakening while a new AI-infrastructure + robotaxi moat is forming. The bet is that the second is wider than the first was. Ground-truth from kb/robotics/wiki/positioning.md does not apply (humanoid matrix); moat read is built from the filing + competitive web.

Lens 4 · Segments

By reportable segment (operating results):

Critical adjustment: Baidu General Business's RMB(6.0)B operating loss is entirely an artifact of the RMB16,190M impairment of long-lived assets booked against the Core asset group. Ex-impairment, Baidu Core operating income ≈ RMB10.1B. So the core business is still solidly operating-profitable; the headline loss is a one-time non-cash write-down (analyzed in L10).

By revenue type — the trend that matters:

This is the whole thesis in two rows: the ad line is in accelerating structural decline (−3% in '24 → −14% in '25) while cloud/other is accelerating up. They crossed toward parity in FY2025. iQIYI is a secular-declining drag (lighter content slate, streaming competition) but barely profitable and self-funding.

Geography: not broken out separately — "substantially all" revenue + long-lived assets are PRC; >96% mainland revenue, >73% of assets in mainland.

---

Phase B — Measure performance

Lens 5 · Earnings Result

Most recent annual (FY2025, the 20-F):

• Revenue RMB129.1B (US$18.5B), −3% YoY.
• Operating loss RMB(5.8)B — but driven 100% by the RMB16.2B Core impairment; ex-impairment operating income RMB10.4B (US$1.5B).
• Net income attributable to Baidu RMB5.6B (US$799M); ex-impairment RMB19.4B (US$2.8B).
• Cost lines: Cost of revenue +10% (bandwidth/depreciation/server custody for cloud); SG&A +9% (channel spend + credit losses); R&D −8% to RMB20.4B (16% of revenue).
• The standout red flag: operating cash flow turned NEGATIVE at RMB(3.0)B (US$431M) vs +RMB21.2B (FY24) and +RMB36.6B (FY23) — a RMB14.5B working-capital swing. This is the single most important number in the filing (see L10).
• CapEx RMB12.1B (9% of revenue, up from 6%) — server build-out for Gen-AI.

Most recent quarter (Q1 2026, reported May 2026 — 6-K, post-20-F):

• Revenue RMB26.0B, +2% YoY — return to growth.
• Core AI-powered business RMB13.6B, +49% YoY, crossing 52% of Baidu Core revenue for the first time.
• AI Cloud revenue RMB11.3B; AI Cloud Infra +79% YoY; GPU Cloud +184% YoY.
• Operating cash flow back POSITIVE at RMB2.7B — the FY2025 cash bleed reversed in the very next quarter.
• Caveat: Q1-26 EPS down ~34% vs the unusually strong Q1-25 comp — margin still pressured by cloud mix + ERNIE-Bot-free monetization reset.

Market reaction / what's priced: stock ~$104–107 (late June 2026), +~50% over 12M on AI-transition confidence but −17% YTD and far below 2021 highs. The tape says: the market half-believes the pivot but is still discounting the ad decline and demanding proof the consolidated line inflects.

Lens 6 · Earnings Calls (sentiment trend)

No transcripts on the research-layer shelf (transcripts=0); sentiment is reconstructed from the 20-F MD&A + Q1-26 call coverage.

Tone arc (last several quarters):

• 2024–early 2025: defensive — "AI-driven search renovation," "near-term pressure on monetization," macro blamed for ad weakness. Management asking for patience on the search rebuild.
• Mid–late 2025: pivot language hardens — open-sourcing ERNIE 4.5 (a strategic reversal; Robin Li had publicly opposed open-sourcing a year earlier, then said "one thing we learned from DeepSeek is that open-sourcing the best models can greatly help adoption"). Chip narrative (Kunlun) steps forward. Q4-2025: rebrands "Core" → "General Business" and introduces an "AI-native view" cutting across segments — a deliberate effort to get investors to value the AI mix, not the legacy frame.
• Q1 2026: confident inflection tone — "AI has become the core driver," AI revenue >50% of Core, cloud growth "leads the Big Three," OCF back positive.

Recurring phrases now: "full-stack, end-to-end AI," "agent-centric," "AI-native," "cost-effective domestic compute." Phrases retired: the old "search + feed monetization" framing and ERNIE-Bot-subscription revenue (made free April 2025). The shift from defending search to selling the AI stack is the clearest sentiment signal — and it is corroborated by the numbers (L5), not just rhetoric.

Lens 7 · Comps

Peer set = Baidu + tracked census names (pony-ai, xpeng) + the obvious China-internet comps the index misses (Alibaba, Tencent) + Alphabet (the global search analog). Multiples are `` with source/date or n/a. None fabricated.

Reading: Baidu's forward P/E (~13–20x) sits below Alphabet (~24x) and roughly in line with Alibaba/Tencent — a China-discount plus an ad-decline discount. The reported 59x EV/EBITDA is a data artifact: the RMB16.2B impairment + cloud-mix margin compression crushed FY2025 EBITDA, so the trailing multiple is mechanically inflated and should be ignored in favor of forward/ex-impairment figures. The sum-of-the-parts gap is the comp story: a ~$34B market cap against US$34B of gross cash + investments implies the market is assigning close to zero net value to the operating businesses (search + a #1 China cloud + the world's highest-volume robotaxi) after backing out the balance sheet — the classic China-internet SOTP setup. (Note: gross cash ≠ net-of-debt-and-VIE-claims equity value; see L13 caveat — but even net, the discount is steep.)

Consensus / price targets: average 12M target $150 (13 analysts, WallStreetZen) to $180 (32 analysts), range $92–$274; consensus rating "Strong Buy." Caveat: the "$59 2026 EPS" figure circulating in screeners is internally inconsistent with a ~$104 ADS price (would imply <2x P/E) — almost certainly an ADS-vs-ordinary or currency data error; EPS consensus is n/a — not cleanly sourced and is not used in L11.

Lens 8 · Stock-Price Catalysts (last ~5 years)

Mostly ``; pattern over 2021→2026:

• 2021 peak → multi-year de-rate. Stock is "well below 2021 highs". Drivers: China ADR regulatory crackdown (2021–22), HFCAA delisting fear, the broad China-internet de-rate, and the secular ad slowdown.
• DeepSeek shock (early 2025) reset the entire China-AI complex — forced Baidu to open-source ERNIE and compete on cost; net catalyst for the "China-AI is real and cheap" narrative.
• AI-transition re-rate (2025–26): +~50% over the trailing 12 months as AI-Cloud growth + Apollo Go scale earned credibility.
• Robotaxi milestones move the stock: Uber partnership (Jul 2025), Lyft (Aug 2025), Middle East/Hong Kong/Switzerland L4 permits — each a discrete catalyst.
• Earnings reaction is now mix-driven: the market reacts to (a) AI-Cloud growth rate, (b) whether ad decline is stabilizing, and (c) Apollo Go ride volume — not to consolidated revenue, which the ad drag keeps roughly flat.

What the pattern reveals: for this name the market reacts to the AI/robotaxi proof-points and the China-policy backdrop far more than to the headline P&L. It is a "show me the transition + don't get delisted" stock. The −17% YTD into mid-2026 says near-term sentiment is cautious despite the structural progress.

---

Phase C — Judge people & books

Lens 9 · Management

• Robin Yanhong Li (57) — co-founder, CEO since Feb 2004, Chairman since 2000. Owns 18.6% economically but controls 59.9% of votes via Class B (10 votes/share, held through Handsome Reward Ltd). This is a founder-controlled company — Li can execute a multi-year pivot without activist or board interference, for better and worse.

• Track record: built China's dominant search engine; navigated the 2010 Google-China exit into a ~20-year franchise; but has overseen a half-decade of flat-to-declining revenue and a lost mobile-social war to Tencent/ByteDance. Mixed: a brilliant search founder who has not produced a second growth engine until (maybe) now.

• Skin in the game: very high and long-tenured; recent option/RSU grants are modest relative to his founder stake. Insider ownership is concentrated and aligned. (No insider-transactions.csv on shelf — ownership read is from the 20-F directly.)

• Haijian He (44) — new CFO since July 2025 (ex-Goldman Sachs TMT/M&A, ex-CFO of Kingsoft Cloud); also chairman of iQIYI. A capital-markets-savvy, cloud-literate CFO installed right as the AI-Cloud story becomes the equity story — a deliberate signal. CFA charterholder, Chicago MBA, HBS AMP.

• Board quality is genuinely strong for a China name: independents include Yang Yuanqing (Chairman/CEO of Lenovo), Sandy Ran Xu (CEO of JD.com, ex-PwC audit partner — now audit-committee-adjacent governance heft), Jixun Foo (Senior Managing Partner, Granite Asia/ex-GGV; XPeng board), Xiaodan Liu (PE, ex-CSRC M&A committee, chairs audit committee). This is a credible, finance-and-tech-deep board.

• Capital-allocation history — the inflection:

• First-ever dividend policy adopted Q1 2026; first payment expected 2026. After 20+ years of never paying a dividend, this signals a maturing, cash-returning posture — and notably the filing says dividends may be "supplemented by proceeds from non-core asset disposals," hinting at portfolio/iQIYI monetization.

• New US$5.0B buyback authorized Q1 2026 (through end-2028); FY2025 actual buybacks RMB5.5B (US$792M), modest vs the cash pile.

• March 2025: US$2B exchangeable bonds referencing Trip.com shares (monetizing a held stake's value while retaining upside) + RMB onshore notes (cheaper funding) — sophisticated, opportunistic financing.

• YY Live acquired Feb 2025 (US$2.14B, RMB14.2B goodwill) — a questionable use of capital into livestreaming the same year they wrote off RMB16.2B on Core; the synergy case ("integrate with mobile ecosystem") is thin.

• ROE/ROIC distorted by the impairment year, but pre-2025 the business compounded at low-double-digit operating margins on a fortress balance sheet.

• Archetype: founder-CEO, technologist, long-tenured, control-entrenched. Implication: high execution autonomy for the AI pivot; low external accountability if it stalls. The new CFO + first dividend + buyback together read as a real shift toward shareholder-friendliness.

Lens 10 · Forensic Red Flags

Forensic lens — every figure labeled.

The headline events (FY2025):

1. RMB16.2B (US$2.3B) impairment of the "Core asset group". Recoverability test failed at Sep 30, 2025 — undiscounted cash flows below carrying value. Fair value set by DCF (independent third-party valuation firm) at 13% discount rate, 8–18% revenue CAGR assumptions. Forensic read: a non-cash write-down is not fraud, but management's own model concluded the legacy Core long-lived assets are worth less than book — a candid admission that the old search/ad asset base is impaired by the AI transition. It is honest, but it is bad news, and it sits inside the segment that also houses the cash flows.
2. Operating cash flow turned NEGATIVE: RMB(3.0)B vs +RMB21.2B prior year. Driven by a RMB14.5B working-capital deterioration. This is the real flag — a non-cash impairment is benign, but a RMB24B swing in operating cash generation is not, even if Q1-2026 reversed it (+RMB2.7B). Watch whether the FY2025 OCF drain was a one-off working-capital timing issue (collections/credit terms tightening into a weak macro) or the start of cash-conversion deterioration. The convergence of an impairment and negative OCF in the same year warrants scrutiny of receivables quality (SG&A rose partly on "expected credit losses").
3. R&D cut 8% to RMB20.4B (16% of revenue) while claiming an aggressive AI transformation. Cutting research spend during a stated platform shift is a yellow flag — either efficiency (Kunlun/PaddlePaddle lowering cost) or under-investment. Given the cloud/chip output, lean toward efficiency, but monitor.

Other accounting considerations:

• VIE structure: Baidu derives 50% of external revenue from variable interest entities (up from 44–45%) controlled by contracts, not equity — Baidu Netcom (99.5% owned by Robin Li personally) + Beijing Perusal + iQIYI VIEs. RMB19.6B of loans to nominee shareholders, no repayment schedule. Standard China-ADR structural risk, but the rising VIE revenue share (50%) and personal ownership by the CEO concentrate enforcement risk. Consolidation is ASC 810 contractual — investors own a Cayman holdco's claim on contracts, not the operating assets.
• SBC: RMB3.6B (down from RMB6.3B in '23) — declining, not flattering non-GAAP unusually. Clean.
• Goodwill: YY Live added RMB14.2B goodwill (US$2.0B) in 2025; no goodwill impairment taken, but a fresh deal into a declining segment bears watching for a future write-down.
• Investment portfolio fair-value: large book of Level-3 non-marketable equity investments measured by management judgment + a consolidated investment company marking unlisted stakes to "fair value" through earnings — a discretionary, opaque line. Other income RMB12.5B (up from RMB7.4B) flattered pre-tax income; quality-of-earnings caveat.

Regulatory findings (required sub-section):

• SEC Litigation Releases / AAERs: None. "No LR found" and "No AAER found" for Baidu in the 2021–2026 search window.
• Item 3 / Legal Proceedings (10-K/20-F own disclosure): 2,378 complaints filed against Baidu in China courts in 2025 (RMB3.3B / US$477M damages sought); 2,054 pending — management assesses aggregate impact as immaterial. The 2020 US securities class actions (In re Baidu, Inc. and In re iQIYI Securities Litigation, EDNY) were both DISMISSED Sept 30, 2024 — a clean win, removes a multi-year overhang. SAMR antitrust: only minor RMB500,000 fines for failure to file concentration notifications — immaterial.
• Non-SEC enforcement (web): No material US/EU regulatory action found. Standard PRC data-security/cybersecurity (CAC, PIPL, DSL) compliance burden, but no headline penalty surfaced. HFCAA: Baidu was a Commission-Identified Issuer in 2022, but PCAOB regained China access Dec 2022; Baidu is not currently identified and has the HKEX 9888 dual listing as a fungible backstop.
• Conclusion: No material accounting-fraud or enforcement findings — verified via SEC EDGAR EFTS (LR + AAER), web search, and 20-F Item 8 Legal Proceedings as of 2026-06-29. The real forensic concerns are operational/quality-of-earnings (negative OCF, impairment, VIE concentration, opaque investment marks), not fraud.

---

Phase D — Project & stress-test

Lens 11 · Forward Projection

Bottom-up from FY2025 actuals + the Q1-2026 inflection. Every input labeled; output ``. Per SKILL --watchlist rules, NO forecast.ts forecast logged (breadth mode). Note: clean EPS consensus could not be sourced (the screener "$59" figure is a data error), so projection is framed on operating income / FCF trajectory, not a precise EPS print.

Anchor (FY2025, ex-impairment): revenue RMB129.1B; ex-impairment operating income RMB10.4B; net income attributable ex-impairment RMB19.4B (US$2.8B).

Driver mix (the engine):

• Online marketing (ads): declining. Model −8% to −12% annually near-term as the AI-search monetization reset continues and macro stays soft; the −13.6% FY2025 decline should moderate (not reverse) as AI-native ad formats (Agents) mature.
• AI Cloud: accelerating. Q1-26 infra +79%, GPU cloud +184%. Model the cloud/"Others" line +15% to +30% annually — the swing factor.
• Kunlun chips: new revenue line, ~RMB8B in 2026 per JPMorgan (6× growth), small but high-strategic-value.
• Apollo Go: still sub-scale on revenue but ride volume +triple-digit; asset-light Uber/Lyft model improves unit economics; not yet a P&L mover, all option value.
• Margins: cloud mix is lower-margin than ads near-term (cost of revenue rising), but operating leverage + Kunlun cost advantage + ERNIE inference efficiency should stabilize margins as cloud scales. R&D held flat-to-down.

Three-year operating-income path (FY2026 / FY2027 / FY2028), ``:

• Base: consolidated revenue re-accelerates from −3% to low-single-digit growth (Q1-26 already +2%), as cloud growth out-runs the moderating ad decline. Operating income (clean, ex one-offs) grows mid-to-high single digits as mix-shift margin drag offsets revenue growth → roughly RMB11–13B FY26 → RMB13–16B FY27 → RMB16–20B FY28 clean operating income. Net income attributable swings widely with the volatile other-income/investment line.
• Bull: ad decline bottoms by FY2026, cloud sustains +25–30%, Kunlun + GPU cloud inflect, Apollo Go starts contributing — consolidated revenue +mid-single-digit and accelerating; clean operating income compounds 15%+ → the SOTP discount closes.
• Bear: ad decline does not moderate (stays −12%+), cloud growth decelerates as price competition with Alibaba/Huawei intensifies, a second impairment (iQIYI/goodwill) lands, OCF stays weak — consolidated revenue flat-to-down for another two years, multiple stays compressed.

The single number to watch: the quarter the consolidated revenue YoY turns and stays positive AND OCF stays positive. Q1-2026 (+2% revenue, +RMB2.7B OCF) is the first data point that it may already be happening. Confirmation across 2–3 quarters is the re-rating trigger.

Lens 12 · Bull vs Bear

Bull case. Baidu is a sum-of-the-parts re-rating coiled spring. At ~$34B market cap against ~US$34B gross cash + investments, the market is paying almost nothing for: (1) China's #1 AI cloud growing infra +79%/GPU +184%; (2) a vertically-integrated AI stack (Kunlun chips + PaddlePaddle + ERNIE-5 + Qianfan) that is one of only ~3 full-stack players in the world and the prime beneficiary of China's Nvidia-substitution; (3) Apollo Go, the highest-ride-volume robotaxi operator globally (20M+ cumulative, ~250K+ weekly rides, 26 cities, Uber/Lyft distribution) — a multi-hundred-billion TAM optionally worth more than the whole current market cap. Capital allocation just turned shareholder-friendly (first-ever dividend + US$5B buyback). The Q1-2026 print — AI >50% of Core, revenue back to +2%, OCF back positive — suggests the transition is inflecting now. Earnings surprise vector: any quarter where the ad drag stops masking cloud + robotaxi.

Bear case. Three things could permanently impair the thesis: (1) The search-ad cash cow is in secular, possibly terminal, decline — −14% in FY2025, and Barchart's framing is blunt: "AI Cloud momentum meets the reality of advertising revenue that isn't coming back". AI-native search may structurally monetize worse than 10-blue-links forever (the innovator's-dilemma trap Google also faces). (2) AI Cloud is a low-margin, brutally competitive commodity — Alibaba Cloud + Huawei Cloud fight on price; growth may come at permanently thin margins, so revenue mix-shift destroys group margins even as it grows. (3) It's still a China ADR on a VIE — 50% of revenue via contractual VIEs, HFCAA tail risk, RMB/capital-control friction, and a CEO who controls 59.9% of votes (minorities cannot force change). Pre-mortem (18 months out, thesis broken): ad decline never moderated, cloud growth decelerated to ~15% at thin margins, a second impairment hit (iQIYI or YY Live goodwill), OCF relapsed negative, robotaxi stayed a cash-burning science project, and the SOTP discount widened because the market re-rated down the value of a shrinking-ad + commodity-cloud combo. Are multiples too high? No — forward P/E ~13–20x is cheap; the bear case is a value trap (cheap because the business is structurally challenged), not an overvaluation.

Contrarian view (what the market refuses to see): The market is anchored on Baidu-as-declining-search and is not pricing Apollo Go as a real asset. If robotaxi unit economics turn (RT6 cost-down + asset-light Uber/Lyft scaling + Gulf/Europe expansion while US players retreat internationally), Baidu owns the single largest pool of commercial L4 ride data and a global distribution footprint — and the option is being given away inside a "melting ad stock" multiple. The same is true of Kunlun: the market prices Baidu as a buyer hurt by Nvidia controls, not as a maker whose TAM those same controls create.

Lens 13 · Devil's Advocate (short-seller)

Dismantling the bull case.

• The structural break: Baidu makes most of its cash from search advertising, and generative AI is dissolving the search-ad business model from the inside. Every AI-generated answer that satisfies a query without a sponsored click is lost revenue — Baidu is cannibalizing its own monetization (≈70% of mobile SERPs now have AI content; ads −14%). This is not cyclical macro weakness; it is the search-ad model breaking. The bull's "cloud + robotaxi replace it" requires both new engines to scale faster than the cash cow shrinks — and so far the consolidated line is still down.
• Revenue concentration / what shifts: revenue is concentrated in (a) China SME ad spend (macro-sensitive, competitively contested by ByteDance/Tencent/Alibaba) and (b) the same domestic economy for cloud. There is ~96% single-country concentration and 50% of revenue runs through contractual VIEs — if Beijing ever tightens VIE enforcement or the HK/US listing arbitrage breaks, the equity claim is structurally fragile.
• Weakest-link moat: the search moat is the one eroding, and AI Cloud's "moat" is thin — it's a price war with two larger/better-capitalized rivals (Alibaba, Huawei). Cloud growth that comes at zero/negative incremental margin is revenue without value.
• Most dangerous competitor bulls underestimate: not Alibaba — ByteDance. Doubao + Douyin search is eating both Baidu's ad dollars and mindshare in consumer AI; ByteDance's distribution dwarfs Baidu's. In robotaxi, Pony.ai is scaling fleet faster (targeting 3,500+ vehicles by end-2026, fare revenue +456%) and could out-execute Apollo Go on the metric that determines profitability (fleet scale).
• Worst capital-allocation moves: buying YY Live (US$2.14B into declining livestreaming) the same year it wrote off RMB16.2B; sitting on US$34B of cash for years while the stock languished (only now returning capital); and the related-party tangle of the CEO personally owning the largest VIE (Baidu Netcom).
• Assumptions that must hold for today's price: ad decline moderates and cloud sustains 20%+ and margins stabilize and no second impairment and no China-policy/VIE/HFCAA shock. That's a long conjunction.
• If growth disappoints 20–30%: the stock is already cheap, so downside is more about no re-rating (dead money) than a crash — but a relapse to negative OCF + a second write-down could see it de-rate toward cash value with the operating businesses valued at a discount to zero, i.e. another 20–30% down.
• Single scenario that permanently impairs: AI-native search monetizes structurally below legacy search forever (à la the Google "innovator's dilemma," but Baidu lacks Google's YouTube/Android/Cloud diversification cushion), so the cash engine never stops shrinking and the cloud/robotaxi engines never reach the margin to replace it. Plausibility: moderate-to-high — this is the real risk, and it is not yet refuted by the data.

Lens 14 · Management Questions (ordered by information value)

1. Of the −13.6% FY2025 decline in online-marketing revenue, how much is macro/cyclical versus structural cannibalization by AI-generated search answers — and at what AI-SERP penetration does ad revenue per query stabilize?
2. The FY2025 RMB16.2B Core impairment used an 8–18% revenue CAGR and 13% discount rate — which assumptions, if missed, would trigger a second impairment, and does that DCF contemplate further ad decline?
3. Operating cash flow swung from +RMB21B to −RMB3B on a RMB14.5B working-capital move — what drove it (collections? credit terms? a one-off?), and is the Q1-2026 +RMB2.7B reversal durable?
4. AI Cloud is growing infra +79% — at what gross/operating margin, and where does steady-state cloud margin settle versus the legacy ad margin it's replacing?
5. What is Apollo Go's per-ride unit economic (contribution margin) today, and at what fleet size / city density does a market turn cash-flow positive?
6. With Alibaba Cloud and Huawei Cloud competing on price, what is the durable differentiation of Baidu's full-stack offer beyond cost — and what is the customer switching cost once built on Qianfan + Kunlun?
7. Kunlun M100/M300: what fab and node, what is the realistic 2026–28 volume given China foundry constraints, and how much of Baidu's own training/inference will run on Kunlun vs Nvidia/Huawei?
8. The first-ever dividend references "non-core asset disposals" — which assets (iQIYI? the investment portfolio? Trip.com stake?) are candidates for monetization, and on what timeline?
9. ByteDance's Doubao/Douyin-search: how do you defend ad dollars and consumer-AI mindshare against a distribution platform larger than yours?
10. Why acquire YY Live (US$2.14B) into a declining livestreaming market the same year you impaired Core — what is the concrete synergy and the ROI hurdle it must clear?
11. iQIYI is barely operating-profitable and secularly declining — keep, spin, or sell, and what is the strategic logic for continued consolidation?
12. The CEO personally owns 99.5% of the largest VIE (Baidu Netcom) carrying 50% of group revenue — what governance safeguards protect minority holders if that relationship were ever contested?
13. On HFCAA: what is the contingency if PCAOB China access is revoked again, and is the HKEX 9888 line liquid enough to absorb a forced ADS migration?
14. With US$34B gross cash, what is the target capital-return payout ratio and the framework balancing buyback vs dividend vs reinvestment in cloud/robotaxi?
15. What does Baidu look like in 2030 — what % of revenue and profit is AI Cloud + Apollo Go + Kunlun vs legacy search advertising, and what is the bridge to get there?

---