Baidu, Inc. (NASDAQ: BIDU) — Winning the Search War, Losing the Ad Business: A Cash-Backed Bet That AI Refills a Draining Annuity

Independent equity research · 2026-07-02 The analysis below (Business Overview through What Must Be True) is deliberately written without a buy/sell recommendation and carries no price target. The one exception is the clearly-labeled opinion block immediately below.

⚡ Claude’s Take

This block is the author’s own independent opinion and general information only — not investment advice. Everything below it (the Business Overview through What Must Be True sections) is written without a position and without a price target.

Verdict: HOLD / accumulate-on-weakness / NOT a short. Conviction: LOW–MEDIUM. A cheap, asset-backed China-internet optionality basket where a genuinely melting search-advertising annuity is wrapped around ~$40B of cash and investments and a set of AI/robotaxi call options the market is pricing at roughly zero. This is a deep-value / sum-of-the-parts situation, not a quality compounder — the crux is that you are paid to wait via the balance sheet, but there is no forcing mechanism to close the discount and the core profit engine is visibly draining. My fair-value zone is ~$105–140/ADS (the base-case band); I would accumulate below ~$95–100 (where you pay almost nothing for a large operating business plus free AI options on top of ~$75/ADS of net financial assets) and would not chase above ~$140. Not a short at $113 — it sits ~30% below its 52-week high and ~67% below its 2021 peak, carries ~66% of its market cap in financial assets, just initiated a dividend and a fresh $5B buyback, and any China-AI or robotaxi headline can squeeze it violently.

The market is pricing this correctly and incorrectly at once. Correctly: the search-ad annuity is not merely cyclically soft — online-marketing revenue fell −13.6% in FY2025 and −22% YoY in Q1 2026, and management wrote down the Baidu Core search asset group by ¥16.2B in Q3 2025 on a DCF that permanently lowered its cash-flow assumptions. That is management conceding the core is structurally impaired, and it is the single most important fact in this report. Incorrectly: at ~1.0x book and ~3.7x EV/EBITDA once you net out the investment portfolio, the market credits essentially nothing for the pieces that are working — #1 share in China GPU cloud (~40%), AI-Cloud infrastructure growing +79% YoY, Apollo Go past 20M cumulative robotaxi rides across 26 cities, and a re-rating of the whole China-AI complex that could return at any time. The framing is a value/special-situation with embedded free optionality, not momentum: the tape is a China-internet beta vehicle that has destroyed value for a decade (−9.5%/yr over five years), briefly woke up in 2025 (+37%), and is now consolidating below its moving averages. The single fact that flips me bullish: evidence that AI-generated search results can monetize at anything close to legacy ad margins (i.e., the annuity stabilizes rather than terminally declines). The single fact that flips me bearish: ad revenue decline steepening past ~−25% with AI-Cloud margins staying thin, confirming the transition destroys more profit than it creates — a value trap dressed as a turnaround.

Tag: “The moat held; the money leaked out the bottom.”

📈 Stock Price Action — Five-Year Event Map

Baidu’s ADS is a two-thirds round-trip. From an all-time high near ~$340 (Feb-2021) the stock collapsed to a five-year low around ~$77 (Oct-2022), then spent three years as dead money before ripping +37% over the trailing twelve months in the 2025 China-AI re-rate to a 52-week high of ~$162.52 (Jan-2026), and then rolling over. It trades at $113.30 (2026-07-02) — ≈−67% below the 2021 peak and ≈−30% off the 52-week high — sitting below both its 50-day and 200-day EMAs (clustered ~$121.5). 52-week range $85.86–$162.52. (Price levels are FACT from the adjusted daily price series; the attributed drivers are INTERPRETATION.)

Cycle narrative. (1) Baidu rode the 2020–21 reflation-and-robotaxi mania to a peak near $340. (2) It then shed three-quarters of its value as Beijing’s platform crackdown, HFCAA delisting fears, and zero-COVID crushed every China-internet name. (3) The Nov-2022 PCAOB audit agreement plus the reopening pivot sparked a sharp snapback. (4) 2023 gave the gains back twice as the first ERNIE demo underwhelmed and the ad recovery stalled. (5) A Sept-2024 stimulus rally and (6) the Sept-2025 DeepSeek-driven China-AI re-rate combined for a +37% twelve-month move, peaking (7) near $163 in Jan-2026. (8) Since then the ADS has fallen ~30% as the AI trade cooled and FY2025 results confirmed an earnings trough — GAAP operating income swung to a loss on a large Q3 impairment — leaving the stock below its moving averages and back at a China-internet-beta multiple. Each move is cross-referenced against the earnings calendar, the 20-F/6-K record, and the news feed; the price move is fact, the cause is interpretation.

1. Executive Summary

Baidu is China’s dominant search engine attempting a violent pivot into artificial intelligence — cloud, large language models (the ERNIE family), and autonomous driving (Apollo Go robotaxi) — while its legacy profit engine drains. The investment question is not whether Baidu has a moat in search; it plainly does (it holds ~64% of China search queries). The question is whether that moat still translates into cash. In FY2025 it did not: search share rose roughly 20 points year-over-year even as online-marketing (advertising) revenue fell 13.6%, and in Q1 2026 ad revenue fell 22% YoY. The moat has detached from its financial outcome — the textbook signal of a franchise being disintermediated, here by the company’s own AI-generated search results (70% of mobile results by late 2025) cannibalizing the paid-link ad unit.

The FY2025 numbers are ugly and partly distorted. Total revenue fell ~3% to ¥129.1B (~$18B). GAAP operating income swung from +¥21.3B to a −¥5.8B loss, driven overwhelmingly by a ¥16.2B non-cash impairment of the Baidu Core search asset group booked in Q3 2025 — a DCF-based write-down that is, in effect, management’s admission of permanent deterioration in the core’s cash flows. Strip the impairment and normalized non-GAAP net income was ¥18.9B (¥53.41/ADS, ~$7.64), down ~30%. Operating cash flow turned negative (−¥3.0B) on a large working-capital swing and a 46% capex ramp for AI infrastructure — a real ramp, but the negative OCF is a timing/consolidation artifact (H2 2025 and Q1 2026 OCF were both positive), not a cash burn.

Against that melting core sits a genuine, largely-free set of AI options: AI-Cloud infrastructure revenue +79% YoY in Q1 2026, #1 share in China’s GPU cloud (~40%), self-designed Kunlun AI chips (a real edge under US export controls), and Apollo Go, the volume leader in Chinese robotaxi with >20M cumulative rides. None is yet a proven profit engine, and all sit inside an over-capitalized China-AI arms race where API prices have already collapsed (DeepSeek’s 2025 shock commoditized frontier LLMs and forced Baidu to open-source ERNIE and make its consumer chatbot free).

Valuation is where the bull case lives. At $113.30 the ADS trades at ~1.0x book, ~1.35x tangible book, and — netting out the ~$22–23B long-term investment portfolio — roughly 3.7x EV/EBITDA on a large operating business. Total cash and investments (~$42B) approximate the entire market cap (~$38.6B). But that headline overstates liquidity: net cash excluding the illiquid investment portfolio is only ~$2.5–3B, and the portfolio (Du Xiaoman 42%, listed China-tech equities, PE/VC) deserves a China/illiquidity haircut. Capital allocation is a cautious cash-steward’s, not a value-maximizer’s: an under-executed buyback, a sprawling low-return investment book, a ¥2.1B YY Live acquisition into a fading vertical, and founder Robin Li’s 60.2% voting control removing external discipline — offset by a first-ever dividend and a fresh $5B repurchase authorization.

The verdict of the body below: a structurally challenged industry position, an eroding (not durable) core franchise, low-quality declining growth partially offset by genuine AI optionality, balance-sheet-heavy but low-return financials, mixed capital allocation, and a deeply-discounted valuation that is cheap for identifiable reasons. The bull needs the AI transition to refill the annuity faster than GenAI drains it; the bear needs only for the draining to continue. No recommendation and no price target appear below this line.

2. Business Overview

Baidu operates two reportable segments: Baidu Core and iQIYI.

Baidu Core (FY2025 revenue ¥102.5B, ~79% of total) is the historical search business plus the AI franchise being built on top of it. Within Core, management now splits revenue two ways:

• Online marketing (advertising) — the legacy annuity. Pay-for-performance search and feed advertising: advertisers bid for placement against user queries and content-feed impressions. This was ¥81.2B in FY2023, ¥78.6B in FY2024, and ¥67.8B in FY2025 (−13.6%) — a franchise in outright decline. It remains the profit engine (high-margin, cash-generative) but is shrinking fast: Q1 2026 ad revenue fell ~22% YoY, with the “legacy/traditional” ad component down ~29%.
• Non-marketing / AI-powered revenue — the offset. Chiefly AI Cloud (enterprise cloud infrastructure and AI/GenAI services, including the ERNIE model family and PaddlePaddle developer platform), plus intelligent-driving/Apollo, and other applications. Management’s “Core AI-powered business” reached ¥40B in FY2025 (+48%) and crossed 50% of Baidu Core’s “General Business” revenue for the first time in Q1 2026 (52%, +49% YoY). Within it, AI Cloud infrastructure grew +34% in FY2025 and +79% YoY in Q1 2026; AI-native marketing tools grew off a small base (+301%).

iQIYI (NASDAQ: IQ; FY2025 revenue ¥27.3B, ~21% of total, −7% YoY) is Baidu’s majority-owned, consolidated online video-streaming platform — the “Netflix of China” — monetizing via subscriptions and advertising. It is a structurally declining, low-margin drag: operating income collapsed from ¥1.8B (FY2024) to ¥0.2B (FY2025), a 0.8% margin. Because it is consolidated, it inflates Baidu’s reported revenue and depresses blended margins; minority interest on the balance sheet was ¥23.4B. Baidu has not sold down the stake.

Revenue by line (RMB billions):

The table captures the entire thesis in six rows: the high-margin ad line is falling double-digits while the lower-margin AI line grows ~50%, and the crossover has already flipped Baidu Core’s mix past 50% AI for the first time. Revenue is roughly flat because the two are canceling out — but each canceling dollar trades high margin for low.

Beyond the segments, Baidu holds a large strategic-investment portfolio (~¥169B), including a ~42% equity-method stake in fintech affiliate Du Xiaoman (deconsolidated 2018), residual listed-equity stakes, and PE/VC positions.

How it makes money: the overwhelming majority of profit still comes from search advertising — a high-margin, recurring, auction-based annuity tied to Chinese consumer/commercial search demand. Revenue recurrence is moderate: advertising is economically sensitive and increasingly contested; cloud is contract-and-consumption based and growing; iQIYI subscriptions are recurring but declining. The strategic tension of the entire business is that Baidu is deliberately trading down its margin structure — using a shrinking high-margin ad annuity to fund a growing, capital-intensive, lower-margin cloud/AI business — and betting the AI franchise ultimately re-monetizes at attractive economics before the annuity runs dry.

Structure/mechanics an investor must internalize: Baidu is a Cayman Islands holding company that controls its Chinese operating entities through a variable-interest-entity (VIE) contractual structure, not direct equity ownership; it lists as a US ADR (foreign private issuer) filing 20-F/6-K (not 10-K/10-Q), with a secondary Hong Kong listing (9888.HK). Each ADS represents 8 Class A ordinary shares. The share structure is dual-class: Class A (1 vote) and Class B (10 votes), with founder/chairman/CEO Robin (Yanhong) Li controlling ~60.2% of votes on ~19.2% of economics. Statements are reported in RMB.

3. Industry Dynamics

China digital advertising is a large, still-growing market (~$141B in 2025, roughly 13% forward CAGR toward ~$267B by 2030) — but Baidu is a share loser inside a growing market, which is the worst place to sit. The profit pool has migrated decisively toward short-video, social, and content-commerce: ByteDance (Douyin) is the #1 digital advertiser in China (~26% share), followed by Alibaba and Tencent, with Kuaishou now generating more ad revenue than Baidu. Search-as-a-category is being disintermediated at both ends — by short-video/social discovery (Douyin, Kuaishou, WeChat Video Accounts, RedNote/Xiaohongshu) capturing attention, and by GenAI assistants (including Baidu’s own AI Overviews) capturing informational queries and collapsing the paid-link inventory that search advertising monetizes. Structural verdict on the ad industry for Baidu specifically: unfavorable — the category is growing but Baidu’s slice and its monetization intensity are both eroding.

China cloud computing is large and growing but oligopolistic and, for Baidu, sub-scale. In general (IaaS/PaaS) cloud, Baidu is a #4–5 also-ran at roughly 6% share, behind Alibaba Cloud (~36%), ByteDance’s Volcano Engine (~15%), Huawei Cloud (~13%), and Tencent Cloud (~7%). Baidu’s genuine strength is the narrower GPU/AI cloud, where it is #1 at ~40% share (Huawei ~30%; the two together >70%), differentiated by a full stack — ERNIE models, PaddlePaddle, and self-designed Kunlun AI accelerators, a real cost-and-supply edge in a world where US export controls throttle Chinese access to Nvidia. The catch: this is a small, capital-hungry, currently margin-dilutive niche inside a market where capital is flooding in.

The Marathon capital-cycle read is negative. Chinese hyperscalers and AI labs are pouring capital into AI/cloud/LLMs — Alibaba has guided to RMB380B+ of AI/cloud capex, Tencent is ramping hard, and a swarm of well-funded model labs (DeepSeek, Moonshot/Kimi, Zhipu/GLM, MiniMax) has already collapsed API pricing. High expected returns are attracting capital that competes those returns away before they are earned. Baidu is a mid-pack participant in an over-capitalized race — precisely the point in the capital cycle where supply-side discipline is worst and incremental returns are most suspect.

Robotaxi/autonomous driving is the one industry vector where the structure is arguably improving: Chinese regulators have been comparatively permissive on driverless permitting, and Baidu’s Apollo Go is the domestic volume leader. But the profit pool is unproven, competition is intensifying (Pony.ai, WeRide, plus global entrants), and unit economics remain pre-scale.

The VIE/ADR regulatory overhang is a permanent tail risk. Baidu operates its Chinese businesses through contractual VIEs (US holders own the Cayman shell, not the operating companies); it has been on the SEC’s HFCAA provisional list since 2022; a Feb-2025 US executive-branch posture revived HFCAA-style enforcement scrutiny of China ADRs; and Baidu appeared on the June-2026 US Department of Defense “1260H” list (reputational/sanctions-adjacent, not an investment ban, but an escalation path). The Hong Kong secondary listing partially de-fangs a US-delisting scenario (holders can convert), but the underlying VIE-enforceability and China-policy risks remain and are the reason China-internet equities carry a structural discount.

Verdict: structurally challenged industry position. Baidu’s large legacy franchise sits in a profit pool draining away from it; the niches it genuinely leads (GPU cloud, robotaxi scale) are small, capital-intensive, and either margin-dilutive or pre-profit; and the whole complex trades under a persistent geopolitical/VIE discount. This is not a structurally good industry for Baidu as currently constituted.

4. Competitive Position

Name the moat, then test whether it still pays. Baidu Core’s historical advantage is a classic pairing of intangibles (brand/habit — “Baidu it” is a verb in Chinese) and scale (the largest Chinese search index and the richest query-log/behavioral dataset). On the Greenwald market-share-stability test, the franchise looks intact: Baidu holds roughly 64% of China search queries as of late 2025, rebounded ~20 points from ~45% a year earlier after it rolled out AI-generated “AI Overview” answers in early 2025 (mobile share ~78%, desktop ~39%). Stable-to-rising share is normally the fingerprint of a durable moat.

But the moat has detached from its financial outcome — and that is dispositive. Share rose ~20 points year-over-year while online-marketing revenue fell ~14% for the year and ~22% in the most recent quarter. A moat that does not defend the cash flow it exists to protect is, in Greenwald’s own framing, failing the only test that matters. The mechanism is self-inflicted: by late 2025 ~70% of Baidu’s mobile search results were AI-generated (up from ~35% in April), and an AI-authored answer displaces the paid sponsored link that is the ad unit. Management explicitly guided ad revenue down, citing an inability (so far) to monetize AI search at legacy density. Baidu is caught in the incumbent’s dilemma made acute: it must cannibalize its only large profit pool to defend query share against DeepSeek/ByteDance-style assistants, with no assurance the new format ever re-monetizes at the old margin.

ERNIE (the LLM) is not a moat. After DeepSeek’s Jan-2025 release matched frontier performance at a fraction of the cost and collapsed API pricing, Baidu made its Ernie Bot consumer app free (April 2025), open-sourced ERNIE 4.5 under Apache-2.0 (July 2025), and shipped ERNIE 5.0 (a 2.4-trillion-parameter MoE) in Jan 2026. Open-sourcing a flagship model is a tacit admission that the model itself is not a defensible product — the value, if any, has moved to distribution and full-stack cost. As of 2026 ERNIE is not the benchmark leader among Chinese models (DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu’s GLM variously lead). The LLM is table stakes, not an edge.

AI Cloud is a real but narrow advantage. Baidu’s ~40% GPU-cloud share and Kunlun-chip full stack are genuine and scarce under export controls — a supply/cost advantage in a specific niche. But it is sub-scale in general cloud (~6%), currently gross-margin-dilutive (consolidated gross margin fell from 50.3% to 43.9% as cloud COGS scaled), and its standalone profitability is undisclosed — a red flag on quality. It is an advantage that has not yet shown up as attractive returns.

Apollo Go is a genuine option, not yet a moat. Baidu is the Chinese robotaxi volume leader — >20M cumulative rides across 26 cities, ~3.4M fully-driverless rides in Q4 2025 (+~200% YoY), >1,000 driverless vehicles in Wuhan, with a purpose-built RT6 vehicle at ~$28–30K (roughly 60% cheaper than prior generations) and international pilots via Uber (Abu Dhabi), plus Lyft/PostBus tie-ups in Europe. Management claims per-ride breakeven excluding depreciation in Wuhan. Against Pony.ai (~1,000-vehicle fleet, ~16% gross margin, claimed Guangzhou breakeven) and WeRide (~1,200 vehicles, ~30% gross margin), Baidu leads on scale/volume but not clearly on unit economics, and a April-2026 mass system failure that froze 100+ Wuhan vehicles is a reminder that the technology and its liability tail are unproven at scale. Fully-loaded (post-depreciation) profitability is not demonstrated.

Direct comparison vs peers. Unlike Alphabet — a search incumbent navigating the same GenAI disruption but cushioned by a vast, diversified profit pool and a scaled, profitable cloud — Baidu has no cushion: it must cannibalize its only large profit engine to fund the transition. Unlike Tencent (WeChat’s social/gaming fortress) or Alibaba (scaled commerce + the leading cloud), Baidu lacks an adjacent, defended cash machine to bankroll the AI build-out from strength. That is the competitive crux.

Verdict: an eroding franchise, not a durable advantage. The search moat is real on the query-share dimension but broken on the dollar dimension; ERNIE is commoditized; AI Cloud is a narrow, sub-scale, margin-dilutive edge; Apollo Go is a real but unproven option. A moat that no longer defends its cash flow is, for investment purposes, a moat in name only.

5. Growth History and Forward Opportunities

History: a franchise that peaked and rolled over. Revenue climbed from ¥107.1B (FY2020) to a peak of ¥134.6B (FY2023), then declined to ¥133.1B (FY2024, −1.1%) and ¥129.1B (FY2025, −3.0%). This is not a growth company at the consolidated level; it is a business whose legacy engine is shrinking faster than its new engine can offset. Decomposed:

• Online marketing (ads): ¥81.2B → ¥78.6B → ¥67.8B (FY23→FY25), i.e., a −16% two-year decline that accelerated to −13.6% in FY2025 and −22% YoY in Q1 2026. Quality of this decline is poor — it is structural (GenAI cannibalization + share of attention lost to short-video), not merely cyclical.
• AI-powered / non-marketing revenue: the offset, +48% to ¥40B in FY2025, led by AI Cloud infrastructure (+34% FY2025, +79% YoY Q1 2026). This is higher-quality growth (real demand, scarce GPU capacity) but lower-quality economics (margin-dilutive, capital-intensive, undisclosed profitability).
• iQIYI: declining (−7% FY2025), a drag with collapsing operating income.

The one encouraging inflection: Baidu Core revenue returned to +2% YoY in Q1 2026 — its first positive print in some time — entirely on AI-Cloud strength offsetting the ad decline. Whether that marks a stabilization or a brief crossover is the central forward question.

Forward opportunities (ranked by credibility):

1. AI Cloud scaling — the most credible near-term growth driver; real demand, #1 GPU-cloud position, Kunlun cost edge. Risk: margins and an over-capitalized market.
2. Apollo Go commercialization — the highest-optionality driver; volume leadership and a cheap purpose-built vehicle, with international expansion via Uber/Lyft. Risk: unproven fully-loaded economics, safety/liability, regulatory pace.
3. AI-native marketing / agent monetization — the bridge that must work for the ad annuity to stabilize: re-monetizing AI search results and selling AI agents/tools to advertisers. Off a small base (+301% FY2025) but from a category that is simultaneously shrinking the legacy ad unit.
4. ERNIE ecosystem / API — least credible as a standalone driver post-commoditization; more useful as a cloud-demand and search-defense enabler than a profit center.

Verdict: low-quality growth at the consolidated level. The company is not growing (revenue down two straight years); the growth that exists is concentrated in a lower-margin, capital-hungry cloud business and a pre-profit robotaxi option, funded by a high-margin annuity in structural decline. The forward case rests on a mix shift re-rating the economics — plausible but unproven, and dependent on the very AI monetization that is currently dilutive.

6. Financial Quality

Five-year financial summary (RMB billions unless noted):

The shape is unmistakable: a business that peaked in FY2023 and has been sliding on the top line while its margin structure compresses and its earnings quality is buffeted by a one-time impairment and investment marks. The one bright spot in the table is the non-GAAP line — normalized earning power near ¥19B — which is what a valuation should anchor to, not the impairment-wrecked GAAP EPS.

Revenue and margins. FY2025 revenue ¥129.1B (−3%); gross margin compressed hard from 50.3% to 43.9% as lower-margin AI-Cloud COGS scaled (Baidu Core COGS +15% on −2% revenue). This is the central financial-quality story: the mix shift away from high-margin advertising into capital-intensive cloud is structurally compressing profitability, and it will continue as long as cloud out-grows ads.

The GAAP earnings collapse is real but partly one-time. Consolidated GAAP operating income went from +¥21.3B (FY2024) to a −¥5.8B loss (FY2025). The swing is dominated by a ¥16.19B non-cash impairment of the Baidu Core (search/ad) long-lived-asset group, booked entirely in Q3 2025, flagged as a Critical Audit Matter — a DCF-based write-down reflecting management’s own lowered growth assumptions for the core. Excluding the impairment, Baidu Core operating income was ~+¥10.1B (a still-halved figure vs the ~¥19.5B ex-items in FY2024). GAAP net income to Baidu fell from ¥23.8B to ¥5.6B; GAAP diluted EPS/ADS from ¥65.91 to ¥11.78.

Normalized (non-GAAP) run-rate. The GAAP→non-GAAP bridge for FY2025 adds back the ¥16.19B impairment and ¥3.6B SBC (net of amortization, investment fair-value marks, and tax): non-GAAP net income to Baidu ¥18.9B (−30% YoY), non-GAAP diluted EPS/ADS ¥53.41 (~$7.64). Q1 2026: GAAP net income ¥3.45B (EPS/ADS ¥8.76), non-GAAP net income ¥4.3B (EPS/ADS ¥12.06, ~$1.75), adjusted EBITDA ~¥6.0B. So the “real” run-rate earnings power sits near ¥19B non-GAAP net income — an ~$8/ADS normalized figure that puts the ADS at roughly a low-teens normalized P/E, not the depressed GAAP optics.

Cash flow — the negative-OCF scare is a timing/consolidation artifact, not a burn. Operating cash flow went +¥36.6B (FY23) → +¥21.2B (FY24) → −¥3.0B (FY25); FCF −¥16.4B after capex jumped from ¥8.3B to ¥13.4B (+46%, AI infrastructure). The negative OCF is explained by a ~−¥42B working-capital drag: “other assets” −¥22.4B (including ~+¥9.1B of long-term supplier advances = prepaid AI-chip/compute capacity, plus a PAG-related loan and VAT/inventory timing), iQIYI content spend −¥11.6B, and receivables −¥4.5B. Critically, H2 2025 OCF was +¥3.9B and Q1 2026 OCF +¥2.7B — the drain reversed, confirming a timing/prepayment effect rather than structural cash consumption. This is an important QoE point: the scariest line in the FY2025 statements is largely mechanical.

Quality-of-earnings flags to carry forward:

• “Other income” is low-quality and volatile: FY2025 total other income +¥12.5B included +¥5.8B of “others, net” fair-value gains on the investment portfolio — mark-to-market noise (ASU 2016-01) that inflates GAAP net income and is rightly excluded from non-GAAP. Watch that GAAP net income is not being flattered by investment marks in “good” quarters.
• Two data-source traps (filing wins): some data aggregators report FY2025 operating income as +¥10.4B (reclassifying the impairment below the line) — the filed GAAP figure is a −¥5.8B loss. Their diluted EPS (¥16.29) also omits accretion of redeemable non-controlling interest — the filed figure is ¥11.78. And some aggregators’ per-share book/sales figures are garbled for BIDU (they do not reconcile to the ADS) — do not anchor on their percentiles.

Returns and balance sheet. GAAP ROE was 3.1% in FY2025 (impairment-depressed) vs 13.9% in FY2024; on normalized earnings ROE is ~7%, and ROIC is a low ~2–3% — but these are distorted by a balance sheet bloated with ~¥294B of cash and financial investments that earn low returns and mechanically depress every return ratio. On the operating capital actually deployed in search/cloud, returns are far higher; the low headline ROIC is an artifact of the cash/investment hoard, not of a bad operating business. Balance sheet: cash + short-term investments ¥115.3B (~$16.5B); total debt ¥97.1B (~$13.5B); long-term investments ¥168.8B (~$23B); net cash excluding the LT portfolio only ~+¥18B (~$2.5–3B); total cash + investments ¥294.1B (~$42B). Per ADS: cash + ST investments ~$48; total cash + investments ~$115–124; book value ~$112; tangible book ~$86.

Verdict: do economics improve with scale? No — they are compressing. The mix shift is structurally lowering margins and returns, and the low headline ROE/ROIC (even normalized) reflects both the diluting cloud transition and an inefficient, over-capitalized balance sheet. The offsetting quality point is that the normalized earnings power (~¥19B non-GAAP net income, positive underlying cash generation once the FY2025 timing effects wash out) is real and the balance sheet is unusually strong — this is a low-return but not a distressed business.

7. Capital Allocation

The scorecard is mixed-to-below-average. Management is a competent, conservative cash-steward but not a value-maximizing allocator, and founder super-voting control removes the external discipline that might force a better outcome.

M&A — the one recent large deal was defensive-at-best. The ~¥14.2B FY2025 goodwill increase is entirely the YY Live acquisition — JOYY’s China live-streaming business. Originally agreed in 2020 at US$3.6B, the deal was terminated in Jan-2024 and then re-signed and closed in Feb-2025 at a reduced ~US$2.1B (¥15.2B consideration; ~93% booked as non-tax-deductible goodwill). Buying a fading, regulated live-streaming vertical while the core melts is questionable capital allocation and carries clear future-impairment risk. Baidu’s acquisitive history is uneven: iQIYI (majority stake, consolidated, now a declining drag) and the 2018 deconsolidation of Du Xiaoman fintech (now a ~42% equity-method holding) sit alongside a long tail of minority investments.

The investment portfolio (~¥169B) is a low-return distraction. It comprises the Du Xiaoman stake, listed China-tech equity securities (a residual Ctrip/Trip.com position among them), and PE/VC holdings. Fair-value marks flow through the GAAP P&L, injecting recurring earnings volatility. This is capital parked at low, uncertain returns and a persistent source of “other income” noise — a value store, perhaps, but not high-ROIC reinvestment, and a candidate for return of capital.

Buybacks — tepid relative to capacity. Against a 2023 $5B authorization, only ~$2.3B was executed by mid-2025 (under half). Cash-flow repurchases were ¥1.9B / ¥4.8B / ¥6.4B / ¥5.5B (~$0.79B) across FY22–FY25; diluted ADS count has fallen from ~354.6M to ~343.0M (~3%/yr) as buybacks modestly exceeded SBC. A new $5B authorization (announced Jan-2026, running through Dec-2028) was opened, with ~$172M repurchased in Q1 2026. Given ~$42B of cash and investments and a sub-book share price, the pace of return is conservative to a fault — the single clearest lever to close the discount, only lightly pulled.

Dividend — a first-ever, and a signal. Baidu adopted its first-ever dividend policy in Q1 2026 (announced Jan-2026), with an initial payment expected around year-end 2026 (amount at board discretion). After paying nothing FY2020–2025, this — alongside the fresh buyback — suggests a nascent shift toward shareholder returns, consistent with the broader China-ADR capital-return wave. It is a positive at the margin but not yet material.

R&D and capex — disciplined on R&D, ramping on capex into an arms race. R&D was ¥20.4B in FY2025 (15.8% of revenue), notably declining in absolute terms (~¥24B → ¥22B → ¥20.4B) even as the AI narrative intensified — arguably efficient, arguably under-investing versus better-funded rivals. Capex, by contrast, jumped +46% to ¥12.1B (~9–10% of revenue) for AI-Cloud GPU/datacenter build-out. Through the Marathon lens, ramping capital into an over-capitalized AI/cloud market with unproven above-WACC returns while consolidated revenue shrinks is the classic late-cycle misstep — the discipline here is unproven.

SBC — a genuine positive. Stock-based comp fell from ¥6.3B to ¥3.6B (~2.8% of revenue) and is not the source of the GAAP/non-GAAP gap (that is the impairment and investment marks). Low and falling dilution, with buybacks shrinking the count — better than most of the peer group.

Governance/incentives. Robin Li controls ~60.2% of votes on ~19.2% of economics via Class B shares; the float has no governance leverage, and key-person dependence on Li is acute. Executive compensation for Chinese ADRs of this type is typically discretionary/option-heavy rather than ROIC/EPS/TSR-linked (full Item-6 metric detail is an open item). The VIE contractual-control structure is the standard China-ADR arrangement, with its attendant enforceability risk. There is no disclosed insider open-market buying or selling signal (FPIs file no Form 4s); Li’s stake has been stable.

Verdict: has management allocated capital intelligently? Mixed, tilting negative on the forward read. Positives: low/falling SBC, a net share-count shrink, an inaugural dividend and fresh buyback, and a ~$42B cushion that caps downside. Negatives dominate the forward view: an under-executed buyback leaving tens of billions parked at low returns, a sprawling low-ROIC investment book that injects GAAP noise, the ¥2.1B YY Live purchase plus the ¥16.2B core impairment evidencing prior misallocation, a capex ramp into an AI arms race as the core shrinks, and founder super-voting removing the discipline that might force capital back to owners.

8. Changes and Headwinds — Last Two Years

Strategic:

• The AI pivot went from narrative to margin reality. Through 2024–2026 Baidu shifted from “we have ERNIE” to actually re-plumbing search around AI (AI Overviews reached ~70% of mobile results) and scaling AI Cloud — with the cost now visible in a −6.4pt gross-margin compression and an ad-revenue decline that steepened to −22% YoY.
• DeepSeek shock (Jan-2025) reset the entire Chinese AI competitive landscape: frontier performance at a fraction of the cost, collapsing API prices, and forcing Baidu to make Ernie Bot free (Apr-2025), open-source ERNIE 4.5 (Jul-2025), and ship ERNIE 5.0 (Jan-2026). Baidu’s paid-LLM monetization thesis was largely destroyed.
• YY Live closed (Feb-2025) at a cut ~$2.1B — a defensive, questionable use of capital.
• Apollo Go scaled and internationalized — past 20M cumulative rides, 26 cities, RT6 vehicle, Uber/Lyft/PostBus partnerships — but suffered an April-2026 mass system failure freezing 100+ Wuhan vehicles.

Financial:

• ¥16.2B Q3-2025 impairment of the Baidu Core asset group — the defining event of FY2025 and a management admission of permanent core deterioration.
• First-ever dividend policy + new $5B buyback (Jan-2026) — a capital-return inflection.
• Negative FY2025 OCF (timing/prepayment-driven; reversed in H2-2025 and Q1-2026).
• Q1 2026 stabilization signal: Baidu Core revenue +2% YoY (first positive in a while), operating margin recovering off the Q3-2025 trough.

Regulatory/geopolitical:

• Revived HFCAA enforcement scrutiny (Feb-2025) of China ADRs and appearance on the June-2026 DoD 1260H list — reputational/escalation risk, not an investment ban.
• Ongoing VIE-enforceability, cybersecurity-review, and data-governance overhangs common to the complex; the HK secondary listing partially mitigates a US-delisting scenario.

Leadership: Robin Li remains chairman/CEO with entrenched control; no board/management change alters the governance calculus.

Verdict: on balance, the last two years weakened the thesis on fundamentals (accelerating ad decline, margin compression, the core impairment, DeepSeek commoditizing the LLM) while modestly strengthening the shareholder-return and stabilization angles (dividend, buyback, Q1-2026 Core inflection, AI-Cloud +79%). The net is a business whose downside is better protected (cash, capital returns) but whose earning power has structurally deteriorated.

9. Risk Analysis (Risk Matrix)

Risk of catastrophic loss / total loss: low. The realistic tail is a VIE-enforceability or forced-delisting scenario, and even there the Hong Kong secondary listing plus ~$42B of cash/investments provide substantial recovery value. The greater risk is not zero, but a slow bleed: the ad annuity draining faster than the AI franchise re-monetizes, leaving a low-return, cheap-forever “value trap” with no forcing mechanism to close the discount.

10. Valuation Discussion (Embedded Expectations)

No price target, no recommendation. Valuation is discussed as embedded expectations and scenarios.

The setup. At $113.30 the ADS carries a market cap of ~$38.6B on ~340M ADS. Against that, total cash and investments are ~$42B — approximately the entire market cap. Book value is ~$112/ADS (P/B ~1.0x) and tangible book ~$86/ADS (P/TBV ~1.35x). On own-history multiples the ADS is cheap: EV/EBITDA ~8.7x on a standard basis, and ~3.7x once the ~$23B long-term investment portfolio is netted out — placing the large operating business (search + cloud + iQIYI + Apollo) at a mid-single-digit EV/EBITDA and ~1.3x operating revenue.

Important caveat on the “cash ≈ market cap” line. It is true but misleading if left there. Liquid net cash (cash + ST investments − total debt) is only ~$2.5–3B; the bulk of the cushion is the ~$23B long-term investment portfolio (Du Xiaoman 42%, listed equities, PE/VC), which deserves a meaningful China/illiquidity/holdco discount. The honest floor is net-financial-assets of roughly $75/ADS (~66% of the price) after a haircut — still substantial asset backing, but not $124/ADS of spendable cash.

Own-history and cross-sectional context. Baidu trades near the lower half of its own 8-year EV/EBITDA range and screens as the cheapest of the major China-internet group on EV/sales and P/B, with by far the highest share of market cap in financial assets — the classic “China holdco discount + earnings trough” set-up. It trades nowhere near Alphabet’s ~14–16x EV/EBITDA / ~20x P/E search-and-AI multiple; the market assigns Baidu a China-internet-beta multiple, not a franchise or AI multiple.

Comparable multiples (approximate, mixed as-of dates — interpretation):

Sum-of-the-parts (assumption-heavy; segment splits approximate).

The current $113 sits between the bear and base SOTP — the market credits little for AI-Cloud growth, roughly nothing for Apollo Go, and haircuts the investment book heavily.

Embedded expectations. Netting ~$15B of haircut financial assets against the $38.6B cap, the operating business is priced at only ~$23–24B (~1.3x operating revenue / ~5–6x EV/EBITDA). For $113 to be “fair” with no re-rate, search need only decline low-to-mid single digits, AI Cloud grows ~15–20% at thin margin, and the multiple holds ~6–8x EV/EBITDA. The market is underwriting continued secular ad decline, a modest cloud multiple with thin margins, and essentially zero for Apollo Go and much of the portfolio. Upside requires a catalyst: AI-search monetization stabilizing the annuity, an AI-Cloud margin inflection, Apollo commercialization credit, China-risk-premium compression (Stock Connect / southbound flows), or buyback-driven share shrink.

Scenario value zones (per ADS — no target, no recommendation):

• Bear ($60–85): GenAI cannibalizes search faster than it monetizes (double-digit ad decline persists); cloud stays low-margin; China/portfolio discount widens; multiple compresses to a trough ~4–5x EV/EBITDA.
• Base ($105–140): ad decline moderates to low-single-digits with gradual GenAI monetization; AI Cloud +20% with improving margin; a modest re-rate to ~7–9x EV/EBITDA; partial investment-book credit; ~$2B Apollo option. The current $113 sits at the low end of base.
• Bull ($150–210): AI-Cloud margin inflection + ERNIE re-accelerating monetization + Apollo commercial ramp credited + China risk-premium compresses + buybacks shrink the float; re-rate to ~10–12x EV/EBITDA plus a fuller SOTP.

The valuation is genuinely cheap, but cheap for identifiable reasons; the embedded expectations are low, which is the source of the asymmetry and the reason it can stay cheap indefinitely absent a catalyst.

11. Variant Perception

Consensus view. The sell-side and much of the buy-side treat Baidu as a cheap, cash-rich, out-of-favor China-AI optionality play — “you get search for free and the AI/robotaxi upside is a call option,” with a persistent China-ADR discount and no near-term catalyst. The tape agrees: the factor model prices BIDU as China-internet beta (dominant loadings to Social-Media and Country-China; near-zero style loadings; beta ~1.06; negative alpha), a name that destroyed value for a decade (five-year return ~−9.5%/yr, max drawdown −77.5%) before a 2025 re-rate (+37% one-year) that is now consolidating below its moving averages. Consensus is neither euphoric nor capitulated — it is indifferent.

The strongest bull case. Baidu is a sum-of-the-parts mispricing: ~$42B of cash and investments against a ~$38.6B market cap, a normalized ~¥19B non-GAAP earnings stream, #1 GPU-cloud position growing +79%, and a genuine robotaxi leader — all for ~1.0x book and ~3.7x operating EV/EBITDA. The AI-search share rebound (to ~64%) shows the franchise is defensible on usage; if Baidu can monetize AI search at even half the legacy density, the ad annuity stabilizes and the mix shift toward a scaling, eventually-profitable AI Cloud re-rates the whole company. Q1 2026’s return to +2% Core growth is the first data point. A China-AI re-rate, southbound-flow inclusion, and an accelerated buyback/dividend could all act as catalysts. You are paid to wait via a fortress balance sheet.

The strongest bear case. The moat has already detached from its cash flow — share up 20 points, ad revenue down 22% — and management has conceded it with a ¥16.2B DCF impairment of the core. This is not a trough to be cycled through; it is a structural disintermediation of query-based advertising by GenAI, and Baidu must cannibalize its only large profit pool to defend query share, with no proof the new format re-monetizes at the old margin. The offsets are lower-quality: AI Cloud is sub-scale and margin-dilutive in an over-capitalized market where DeepSeek collapsed pricing; ERNIE is commoditized and open-sourced; Apollo Go is pre-profit with a fresh safety black-eye. The ~$42B “cash” is mostly an illiquid, low-return investment book behind a founder who returns capital grudgingly. Result: a low-return business that is cheap and stays cheap — a value trap with a super-voting founder and no forcing mechanism.

The 3–5 assumptions that matter most:

1. AI-search monetization — does the AI-generated result format ever re-monetize at a density approaching legacy search ads? (Bull: yes, partially; Bear: no.)
2. Ad-decline trajectory — does the −22% YoY stabilize toward low-single-digits, or steepen? (The single most important swing variable.)
3. AI-Cloud margins — does the #1 GPU-cloud position inflect to attractive profitability, or stay a dilutive share-grab?
4. Capital return — does management pivot to aggressive buyback/dividend to close the discount, or leave capital parked?
5. China risk premium — does the VIE/geopolitical discount compress (Stock Connect flows, US-China détente) or widen (HFCAA/1260H escalation)?

Falsification evidence: the bull is falsified by two more quarters of >20% ad decline with flat/negative Core growth and no AI-Cloud margin improvement; the bear is falsified by AI-search monetization data showing ad revenue stabilizing (decline < ~5%) alongside a visible AI-Cloud margin inflection. The factor-positioning read supports the “offsides consensus” thesis on the downside-protected side: a decade of value destruction and current below-EMA, negative-momentum positioning means expectations are low and crowding is absent — a name the market has given up on, which is where mispricings live if the fundamentals stabilize, and where value traps live if they do not.

12. Fact vs. Interpretation Table

13. Open Questions

1. AI-Cloud standalone profitability — Baidu does not disclose AI-Cloud segment margins. Is it gross-margin-positive and on a path to operating profit, or a dilutive share-grab?
2. AI-search monetization density — what is the revenue-per-query of the AI-Overview format vs the legacy sponsored-link unit? Management has not quantified it.
3. Ad-decline stabilization — is the Q1-2026 Core +2% a genuine inflection or a one-quarter crossover as easy comps and AI-Cloud mask the ad bleed?
4. Apollo Go fully-loaded economics — per-ride profitability including depreciation and R&D; path and timeline to segment profitability; the liability/insurance model after the April-2026 Wuhan failure.
5. Investment portfolio composition and liquidity — the mark-to-market vs realizable value of the ~¥169B book; how much is Du Xiaoman vs listed vs illiquid PE.
6. Executive incentive metrics — are there ROIC/EPS/TSR-linked comp metrics, or is it discretionary/option-based? (20-F Item 6 not fully extracted.)
7. Capital-return trajectory — will management materially accelerate buybacks/dividends against the ~$42B cushion, or leave capital parked?
8. YY Live integration — rationale, synergy, and impairment risk of a fading live-streaming asset bought as the core melts.

14. What Must Be True

Bull case — what must be true:

1. The ad annuity stabilizes. Online-marketing decline moderates from −22% YoY toward low-single-digits over the next 4–6 quarters as AI-search monetization ramps. Falsification test: two consecutive quarters of ad revenue worse than −20% YoY with no improvement in Baidu Core operating margin → bull broken.
2. AI Cloud inflects to attractive economics. Infrastructure growth (+79% Q1’26) sustains while gross margins stop compressing and turn up. Falsification: consolidated gross margin falls below ~42% for two more quarters with cloud still undisclosed/dilutive → bull broken.
3. Capital is returned. Management materially accelerates buyback/dividend against the ~$42B cushion, shrinking the float and signaling discipline. Falsification: buyback pace stays <$1B/yr and the dividend is token through 2027 → the “paid to wait” leg weakens.
4. Optionality is credited. Apollo Go demonstrates fully-loaded unit economics and/or the China risk-premium compresses. Falsification: another safety/operational failure or an HFCAA/1260H escalation → optionality and multiple both cut.

Bear case — what must be true:

1. GenAI structurally disintermediates search advertising. The AI-Overview format never re-monetizes near legacy density; ad revenue keeps falling double-digits. Falsification: AI-search monetization data showing ad revenue decline < ~5% with stable ad load → bear broken.
2. The AI offsets stay low-quality. AI Cloud remains sub-scale and margin-dilutive in an over-capitalized market; ERNIE stays commoditized; Apollo stays pre-profit. Falsification: a visible AI-Cloud operating-margin inflection or Apollo segment profitability → bear broken.
3. Capital stays parked. The founder returns capital grudgingly; the discount never closes; the low-return investment book persists. Falsification: an aggressive, sustained capital-return program → the value-trap thesis weakens.
4. The multiple is right, not wrong. A low-return, structurally-declining core in a discounted geography deserves ~1.0x book. Falsification: normalized ROE re-rating toward mid-teens on a stabilized mix → the multiple is too low.

The elegant tension: both cases share the same balance sheet and the same ~$113 price. The bull and bear diverge on a single empirical question — can AI-generated search results monetize at anything close to legacy ad margins? — and that question will be answered in the ad-revenue line over the next few quarters. Watch that line above all else.

15. Source Appendix

See the Source Appendix below for the full source list. Primary sources: Baidu FY2025 Form 20-F (filed 2026-03-17); Q4/FY2025 and Q1-2026 earnings 6-K exhibits (2026-02-26; 2026-05-18); FY2020–FY2024 20-Fs; SEC EDGAR filing index (CIK 1329099). Quantitative cross-checks: aggregated fundamental-data providers (statements, ratios, EV — reconciled to filings, with noted discrepancies on operating income and EPS); adjusted price series and public news aggregation; a public multi-factor risk model. Peer context drawn from public filings and disclosures of Alibaba (BABA), Tencent (TCEHY/0700.HK), PDD, JD.com (JD), and Alphabet (GOOGL). All third-party figures are labeled and reconciled to primary filings where they drive a conclusion.

APPENDIX A — Standard Diligence Questionnaire

Baidu, Inc. (NASDAQ: BIDU) — Report date 2026-07-02

Supplemental diligence questionnaire. Grounded in the underlying research; Fact / Interpretation / Assumption labels applied where material.

General

What thoughtful questions have other investors asked about this company? The dominant investor debate is “value trap vs. deep-value optionality”: is Baidu a cheap, cash-rich China-AI call option (search for free, cloud/robotaxi upside embedded), or a low-return franchise whose only large profit pool is being structurally disintermediated by GenAI? Sub-questions that recur: (i) Can AI-generated search results ever monetize at legacy ad density? (ii) Is AI Cloud a real profit engine or a dilutive share-grab? (iii) What is Apollo Go actually worth, and when does it make money fully-loaded? (iv) Will the founder ever return the ~$42B cash/investment hoard to owners? (v) How should one discount the VIE/HFCAA/1260H geopolitical tail? (Interpretation.)

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? A low — FY2025 was an earnings trough. GAAP operating income turned negative (−¥5.8B) on a ¥16.2B core impairment; normalized non-GAAP net income (¥18.9B) was down ~30% YoY, and margins compressed (gross margin 50.3%→43.9%). (Fact.) But it is partly structural, not merely cyclical — the ad-annuity decline is secular. (Interpretation.)

Driven by external environment or internal actions? Both. External: weak China ad macro, GenAI disruption, competitive share loss to short-video. Internal: the deliberate mix-shift into lower-margin AI Cloud, the AI-search cannibalization of the ad unit, and a 46% capex ramp. (Interpretation.)

How stable are revenues? Moderately unstable and declining — total revenue fell two straight years (FY24 −1.1%, FY25 −3.0%). Advertising (economically sensitive, contested) is falling double-digits; AI Cloud (consumption/contract) is growing +79%; iQIYI subscriptions recurring but declining. (Fact.)

Outlook for products/services? Legacy search advertising: structural decline. AI Cloud: high growth, uncertain margin. Apollo Go: high-optionality, pre-profit. iQIYI: declining. (Interpretation.)

How big will this market be? China digital ad ~$141B (2025) → ~$267B (2030) at ~13% CAGR, but Baidu is losing share within it. China cloud growing fast; GPU/AI cloud faster. Robotaxi a large future TAM if unit economics work. Predominantly domestic (China), with nascent international robotaxi pilots. (Fact / Interpretation.)

Business Quality & Competitive Moat

Is the industry getting more or less competitive? More — search is being disintermediated by short-video/social discovery and GenAI assistants; the cloud/LLM market is over-capitalized with collapsing API pricing (post-DeepSeek). (Interpretation.)

How profitable is the business (ROIC, ROE)? Low on a headline basis — FY2025 GAAP ROE 3.1% (impairment-depressed), ~7% normalized; ROIC ~2–3%. But headline returns are severely depressed by a ~¥294B cash/investment hoard; returns on operating capital in search/cloud are materially higher. (Fact / Interpretation.)

How profitable is the industry — competitors, barriers to entry? Search advertising historically very profitable but the profit pool is migrating to ByteDance/Kuaishou/Tencent. Cloud is low-margin and capital-intensive. Barriers: search has real scale/index/data barriers (Baidu ~64% share) but they no longer defend the cash flow; cloud barriers are capital + chips (Baidu’s Kunlun edge); LLM barriers have largely collapsed. (Interpretation.)

Can the business be easily understood? Reasonably — a search-advertising annuity funding an AI/cloud/robotaxi build-out, plus a consolidated streaming asset and a large investment portfolio, inside a China VIE/ADR wrapper. The complexity is in the balance sheet (investment book, VIE, dual-class) and the accounting distortions (impairment, FV marks, negative OCF). (Interpretation.)

Can it be undermined by foreign low-cost labor? Not directly relevant — it is a domestic-China digital-services business. The analog risk is domestic low-cost competition (DeepSeek collapsing LLM pricing). (Interpretation.)

Do brands matter? Yes — “Baidu” is a verb for search in Chinese (intangible/habit moat), but brand equity is not translating into ad dollars as GenAI erodes the format. (Interpretation.)

Nature of competition? Attention (vs Douyin/Kuaishou/WeChat/RedNote), query monetization (vs GenAI assistants), cloud (vs Alibaba/Huawei/Tencent/ByteDance), models (vs DeepSeek/Qwen/Kimi/GLM), robotaxi (vs Pony.ai/WeRide). (Fact.)

Customers’ switching costs? Low for consumers (search is a habit, not a lock-in) and moderate for cloud/enterprise (integration, data gravity). (Interpretation.)

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? The ~¥169B investment portfolio (incl. ~42% Du Xiaoman, equity-method) may hold hidden value or hidden marks; Apollo Go’s value is not separately carried; the ERNIE/PaddlePaddle IP is expensed. (Interpretation.)

Off-balance-sheet liabilities? The VIE contractual structure is the key structural exposure (control by contract, not equity). Standard operating-lease and content-commitment (iQIYI) obligations. (Interpretation.)

How conservative is the accounting? Mixed. Conservative in taking the large Q3-2025 impairment promptly; less conservative in that GAAP net income is inflated by investment fair-value marks (ASU 2016-01) in “other income.” Non-GAAP reconciliation is standard. (Interpretation.)

How CapEx-hungry is the business? Increasingly — capex jumped +46% to ¥12.1B (~9–10% of revenue) for AI-Cloud GPU/datacenter build-out, and rising as cloud scales. Historically the search business was capital-light. (Fact.)

Capital Allocation & Management

How much FCF, and how is it used? Normalized FCF is positive (FY2025’s −¥16.4B was a timing/capex artifact; H2-2025 and Q1-2026 OCF were positive). Uses: modest buybacks (~$0.79B FY25), a new $5B authorization (2028), a first-ever dividend (2026), heavy AI capex, and a large parked investment portfolio. Philosophy: conservative cash-steward. (Fact / Interpretation.)

Significant acquisitions recently? YY Live (~$2.1B, Feb-2025, ~93% goodwill) — a fading live-streaming vertical; questionable. (Fact / Interpretation.)

Buying back shares? Yes, but tepidly — diluted ADS down ~3%/yr; only ~$2.3B of a prior $5B authorization used by mid-2025; new $5B program through Dec-2028. (Fact.)

Issuing large amounts of new shares to insiders? No — SBC is low and falling (~2.8% of revenue), and buybacks exceed it. (Fact.)

Compensation policy of directors/management? Full metric detail not extracted; typically discretionary/option-based for China ADRs of this type (open question whether ROIC/EPS/TSR-linked). (Open question.)

Motivations of management? Founder Robin Li controls ~60.2% of votes on ~19.2% economics — entrenched, long-horizon founder control, aligned on long-term value but insulated from external discipline; capital return has been grudging. (Fact / Interpretation.)

Valuation & Market Data

Is the stock an ADR, MLP, or K-1 issuer? An ADR (foreign private issuer; files 20-F/6-K), each ADS = 8 Class A ordinary shares; secondary HK listing (9888.HK). Not an MLP; not a K-1 issuer. (Fact.)

Dividend policy? First-ever dividend policy adopted Q1-2026; first payment expected ~YE2026 (amount at board discretion). No dividends FY2020–2025. (Fact.)

How profitable is the business? Low headline returns (see above), distorted by the cash/investment hoard; normalized non-GAAP net income ~¥19B; operating-capital returns materially higher than headline. (Fact / Interpretation.)

Is net income diverging from cash from operations? Yes in FY2025 — GAAP net income positive (¥5.6B) while OCF was negative (−¥3.0B), driven by working-capital/prepayment timing and AI-chip advances; the divergence reversed in H2-2025/Q1-2026. Watch the investment fair-value marks in “other income” for GAAP/cash divergence going forward. (Fact.)

Risks & Downside

What factors would cause the stock to decline? Continued/steepening ad decline; AI-Cloud margins staying thin; a China ad-macro downturn; VIE/HFCAA/1260H escalation; an Apollo Go safety event; investment-portfolio marks/impairment; capital staying parked; RMB depreciation. (Interpretation.)

Risk of a catastrophic loss? Low — the realistic tail is VIE-invalidation or forced US delisting, materially mitigated by the HK listing and ~$42B of cash/investments. (Interpretation.)

Chance of a total loss? Very low — substantial asset backing (~$75/ADS net financial assets after haircut) and a large, if declining, cash-generative operating business. (Interpretation.)

Recent News & Events

Has the business environment changed recently? Yes materially — the DeepSeek shock (Jan-2025) commoditized LLMs and collapsed API pricing; AI-search cannibalization accelerated (ad revenue −22% YoY Q1-2026); the ¥16.2B core impairment (Q3-2025) formalized structural deterioration; and a first-ever dividend + new $5B buyback (Jan-2026) marked a capital-return inflection. (Fact.)

Significant acquisitions? YY Live closed Feb-2025 (~$2.1B). (Fact.)

Change in accounting policies? No material change; the FY2025 impairment and investment fair-value marks are the notable P&L distortions. (Fact.)

Recent changes — new markets, facilities, management? Apollo Go international robotaxi pilots (Uber/Abu Dhabi, Lyft/PostBus Europe); AI-Cloud GPU/datacenter capex ramp; ERNIE 5.0 launch (Jan-2026) and ERNIE open-sourcing (Jul-2025). No change in senior leadership (Robin Li remains chairman/CEO). (Fact.)

APPENDIX B — Source Appendix

Baidu, Inc. (NASDAQ: BIDU) — Report date 2026-07-02

Primary sources first; third-party quantitative feeds reconciled to filings where they drive a conclusion. Facts labeled in the memo; this appendix lists provenance.

Primary — SEC filings (EDGAR, CIK 0001329099)

• Form 20-F, FY2025 — filed 2026-03-17. https://www.sec.gov/Archives/edgar/data/1329099/000119312526109289/d38065d20f.htm — segments, online-marketing vs AI revenue, ¥16.19B Q3-2025 impairment (Critical Audit Matter), YY Live goodwill, balance sheet (cash/investments/debt), cash-flow statement, SBC, buybacks, dual-class ownership, VIE structure. (Saved locally: _scratch/bidu_20f_2025.htm/.txt.)
• Form 20-F, FY2024 — filed 2025-03-28. https://www.sec.gov/Archives/edgar/data/1329099/000119312525066199/d853848d20f.htm — prior-year comparatives.
• Form 20-F, FY2023 / FY2022 / FY2021 / FY2020 — 2024-03-15 / 2023-03-22 / 2022-03-28 / 2021-03-09 — multi-year revenue/margin trend.
• Form 6-K — Q1 2026 results — filed 2026-05-18. Ad revenue −22% YoY; Baidu Core +2% YoY; AI Cloud infra +79%; non-GAAP EPS/ADS ¥12.06; adj. EBITDA ~¥6.0B; positive OCF. (Saved: _scratch/q1_2026_pr.htm/.txt.)
• Form 6-K — Q4/FY2025 results — filed 2026-02-26. FY2025 segment detail; buyback + first-ever dividend policy ($5B authorization through Dec-2028). (Saved: _scratch/q4_2025_pr.htm/.txt.)
• Recent 6-Ks — 2026-07-02, 2026-06-09, 2026-04-30, 2026-04-23, 2026-03-17 (per EDGAR filing index).
• SEC EDGAR filing index — https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=1329099 (20-F/6-K enumeration; BIDU is a foreign private issuer — no 10-K/10-Q/Form 4).

Primary — company disclosures

• Baidu Investor Relations (earnings releases, ERNIE / Apollo Go operating metrics — cumulative rides, cities, RT6 vehicle, international pilots).
• iQIYI (NASDAQ: IQ) filings for the consolidated streaming segment.

Third-party quantitative (reconciled to filings)

• Aggregated fundamental data providers — income statement, balance sheet, cash flow, profitability/credit/per-share ratios, enterprise value, valuation multiples. Noted discrepancies: some aggregators reclassify the FY2025 impairment (reporting +¥10.4B operating income vs the filed −¥5.8B loss) and quote a diluted EPS (¥16.29) that omits redeemable-NCI accretion (filed ¥11.78) — filing figures used throughout.
• Daily adjusted price/OHLCV series (5-year event map, EMAs, beta) and public news aggregation for recent events.
• A public multi-factor risk model — factor loadings (China-internet beta profile), risk-adjusted returns/drawdowns by horizon, factor-similar peers (the KWEB/BABA/JD/BILI/WB complex), beta/alpha/relative strength, idiosyncratic vol. Third-party statistical estimates; facts (loadings, returns) reportable, “continue/mean-revert” flagged as interpretation.

Third-party qualitative

• China search-share trackers (StatCounter-type data) — ~64% share late-2025, ~20pt YoY rebound.
• China digital-ad and cloud market-share/size data (industry trackers) — ByteDance #1 advertiser (~26%); Baidu ~6% general cloud / ~40% GPU cloud.
• Trade/technical press on the DeepSeek shock, ERNIE open-sourcing (Apache-2.0, Jul-2025), ERNIE 5.0 (Jan-2026), and Apollo Go operations (incl. the April-2026 Wuhan mass system failure).
• US regulatory sources — SEC HFCAA provisional list (2022–), Feb-2025 enforcement posture, June-2026 DoD 1260H list.

Peer context (public filings & disclosures)

• Alibaba (NYSE: BABA / 9988.HK) — cloud/AI comp, VIE structure
• Tencent (0700.HK / TCEHY) — competitor, quality premium of the group
• PDD Holdings (PDD) and JD.com (JD) — China ad / e-commerce macro, VIE/ADR
• Alphabet (GOOGL) — the search / GenAI-cannibalization analog

Currency: statements reported in RMB; USD conversions at ~7.0 (FY2025) unless noted. Each ADS = 8 Class A ordinary shares. Price $113.30 as of 2026-07-02 (adjusted daily price series).