Baidu Q3: AI revenue has exceeded half of the total, and its profit contribution has not yet taken over the leading role.
Yesterday, Baidu released its third-quarter earnings report. Its total revenue reached 30.8 billion yuan, down 1% year-on-year and 2% quarter-on-quarter; net profit attributable to Baidu was 1.4 billion yuan, and net profit under non-GAAP was 2.4 billion yuan. Judging solely from the profit figures, this is hardly a standout earnings report. But when breaking down the revenue structure, a more significant signal emerges: Baidu's core new AI business revenue hit 13.2 billion yuan, accounting for 53% of its general business revenue, marking the third consecutive quarter that the proportion has exceeded 50%. This company that started with search advertising has historically shifted its revenue focus to AI.
In contrast, the capital market shows divergence in pricing this earnings report: Baidu's AI transformation has crossed the 50% threshold in terms of revenue structure, but on the income statement and cash flow statement, the costs of transformation are still being fully realized. For investors, the real question to answer is not "how high is the proportion of AI revenue", but whether this shift in revenue structure is creating value or consuming value.
I. The "two-tiered reality" in the income statement: AI is growing while profits are shrinking
When splitting Baidu's revenue into two layers, the contrast is extremely stark.
The first layer is traditional online marketing, which is still accelerating its decline. Online marketing service revenue in the third quarter was 11.9 billion yuan, down 22% year-on-year, and its proportion in general business revenue plummeted to 48% from 62% in the same period last year. What is more alarming is that this decline has expanded compared with previous quarters. CLSA estimated before the earnings release that the year-on-year decline of Baidu's general business marketing revenue in the third quarter would be about 25% (the actual decline announced in the earnings report is 22%), and attributed the decline to "the continued sluggish macroeconomy and moderate product iteration". When Fitch downgraded Baidu's long-term issuer rating from "A" to "A-" in August 2026, the core reason given was also "structural decline in the search advertising business", pointing out that AI competition is eroding the monetization capacity of Baidu's search business. This means that the decline in online marketing is not a simple cyclical fluctuation, but a structural contraction whose bottom has not yet been seen.
The second layer is the AI business, which has strong growth momentum but far thinner profit margins than traditional advertising. Smart cloud infrastructure revenue reached 8.2 billion yuan, up 94% year-on-year, significantly accelerating from the 50% growth rate of the previous quarter; among which GPU cloud revenue increased by 191% year-on-year and 43% quarter-on-quarter. AI application revenue was 2.6 billion yuan, basically flat year-on-year; AI native marketing service revenue was 2.4 billion yuan, down 15% year-on-year. AI cloud infrastructure is a clear growth engine, but its profit margin structure is completely different from that of traditional search advertising: computing power leasing and cloud services require continuous capital expenditure and depreciation and amortization, while the marginal cost of search advertising is nearly zero.
The superposition of the two layers of revenue is reflected in the contraction on the profit side. Operating profit in the third quarter was only 626 million yuan, with an operating profit margin of 2%; non-GAAP operating profit was 1.7 billion yuan, with a profit margin of 6%. Cost of sales increased by 9% year-on-year to 19.9 billion yuan, forming a sharp contrast with the 1% slight decline in revenue; R&D expenses decreased by 7% year-on-year to 4.8 billion yuan. CLSA expects profitability to remain under pressure in the fourth quarter, as the impact of "increased AI investment" has not been fully reflected.
This leads to the most critical financial logic of this earnings report: Baidu is using the shrinkage of its high-margin business to replace the expansion of its low-margin business. Every one yuan of reduced revenue from online marketing may have an impact on profits equivalent to that which can only be offset by several yuan of increased revenue from AI cloud infrastructure. The transformation has been "successful" in terms of revenue proportion, but in terms of profit contribution, the substitution effect is far from complete.
II. The high growth rate of AI cloud needs to be analyzed in detail: the "volume" of GPU cloud and the "quality" of MaaS
The 94% growth rate of smart cloud infrastructure is undoubtedly the most eye-catching highlight in the earnings report, but the growth rate alone is not enough to explain the whole picture. We need to break down its structural composition.
The 191% year-on-year growth of GPU cloud revenue is the main driving force for the acceleration of smart cloud infrastructure. The essence of GPU cloud is computing power leasing. Against the backdrop of the current explosion of domestic large model training and inference demand, the demand side is indeed strong. However, computing power leasing is a capital-intensive business with limited differentiation, the pressure of price competition always exists, and revenue quality is highly dependent on the on-shelf rate and contract term. The earnings report disclosed a positive signal: the smart cloud infrastructure has "increased customer prepayments and continued to shift to longer-term contracts". The increase in prepayments and the extension of contract terms, to a certain extent, indicate that customer stickiness and revenue predictability are improving, which has more analytical value than the mere number of GPUs put on shelves.
A more "quality-oriented" indicator comes from the Qianfan MaaS platform: the Token call revenue from external customers increased by 1159% year-on-year. Admittedly, the triple-digit growth rate is partly due to the low base, but the nature of MaaS revenue is different from that of GPU leasing — it is closer to platform-layer revenue, where customers pay for model calls, and stickiness depends on model capabilities and platform ecosystem, rather than pure computing power prices. The "Token Hub" platform launched by Baidu at its World Conference in November has attracted more than 100 settled supply resources within one month of launch, with an average daily Token call volume of over 700 billion, trying to extend itself from a "computing power provider" to a "model transaction matching party". If the Token Hub can gradually establish a credible model transaction standard and credit system, Baidu's position in the AI value chain will migrate from the heavy-asset bottom layer to the upper layer, which is meaningful for the long-term profit margin structure.
Variables at the hardware layer cannot be ignored either. The Kunlun Chip M100 is expected to be ready for mass production by the end of 2026, positioned as a chip optimized for large-scale AI inference. The significance of self-developed chips does not lie in short-term revenue contribution, but in the internalization of computing power costs. When the scale of GPU cloud continues to expand, the capital expenditure and depreciation of externally purchased chips are hard constraints on profit margins; the gradual introduction of self-developed inference chips is expected to push down the cost curve of unit computing power in the medium and long term. However, this takes time, and the mass production ramp-up and actual deployment efficiency of the M100 still need to be verified.
III. Apollo Go and Token Hub: Option value is accumulating, but has not yet entered the income statement
Strictly speaking, neither Apollo Go nor Token Hub constitutes the core variable of the current earnings report's income statement, but they determine the "upper limit" of Baidu's transformation narrative.
The most substantial progress of Apollo Go in the third quarter is the signal of improved operational efficiency: the order volume of fully driverless operations in Shenzhen increased by about 170% quarter-on-quarter, it expanded its operational area in Dubai and doubled its fleet size. As of October 2026, its global footprint covers 30 cities, with a cumulative autonomous driving mileage of over 390 million kilometers. The reason why the quarter-on-quarter jump in Shenzhen's orders is important is that it occurred in existing operating cities, pointing to the optimization of the single-city operation model rather than pure geographical expansion. Robin Li has previously confirmed that Apollo Go has achieved break-even per vehicle in Wuhan, and the quarter-on-quarter growth in Shenzhen further strengthens the assumption that "the unit economic model is replicable". However, it is necessary to be sober that the large-scale profitability of the autonomous driving business is still premised on a larger fleet size, lower vehicle costs (especially sensors and computing platforms) and a more relaxed regulatory environment. The current stage is still in the range of "verifying the model" rather than "realizing profits".
Token Hub is a more forward-looking bet by Baidu on the AI commercialization model. A core pain point in the current large model call market is that model quality is uneven, and callers lack credible evaluation standards. Token Hub focuses on "verification" and "credit investigation", trying to act as an "intermediate layer" for model transactions through model fingerprint comparison, capability testing and credit rating. If this platform can accumulate sufficient suppliers and callers, Baidu is likely to occupy a light-asset, high-stickiness node in the value chain of model distribution. However, its revenue model is still unclear, and the current average daily call volume of 700 billion Tokens is more of an indicator of ecological activity, which is still far away from identifiable profit contribution.
IV. The essence of valuation divergence: Is the market pricing for "current profits" or "future revenue"
Institutions' divergence on Baidu's current valuation status accurately maps the inherent contradiction of this earnings report.
CLSA's position leans towards the "profit realist school": with a target price of 115 US dollars, the core logic is that the decline in marketing revenue coupled with increased AI investment leads to a 21% and 18% downward revision of adjusted net profit forecasts for 2026 and 2027 respectively. Under this framework, the high growth of AI cloud is not enough to offset the shrinkage of the profit pool of traditional businesses, and Baidu's profitability inflection point is pushed back.
BOCOM International's position in its research report in June 2026 leans towards the "segment revaluation school": adopting the sum-of-the-parts (SOTP) valuation method, assigning a 3x 2026 price-to-sales ratio to the AI business, 1x price-to-sales ratio to the traditional advertising business, valuing the approximately 58% stake in Kunlun Chip at 20x 2026 price-to-sales ratio, plus net cash and short-term investments, it derives a target price of 175 US dollars / 171.7 Hong Kong dollars, maintaining the "Buy" rating. The core assumption of this framework is that the AI business should not be priced with the profit logic of traditional advertising, but with revenue scale and strategic scarcity; the independent valuation of Kunlun Chip provides an additional anchor for asset value.
The difference between the two positions is essentially different answers to "what to focus on during the transformation period". CLSA focuses on the present of the income statement, while BOCOM International focuses on the future of revenue structure and asset value. JPMorgan previously raised the Hong Kong stock target price to 225 Hong Kong dollars after the Q1 earnings report in May 2026, whose logic is closer to the latter — its SOTP valuation shows that the valuation per ADS corresponding to the AI business alone reaches 169 US dollars.
Baidu's current financial characteristics — slight revenue decline, thin profits, AI revenue accounting for more than half, total cash and investments of 276.6 billion yuan — fall right in the gap between the two pricing frameworks. For funds that prefer the certainty of current earnings, the 2% operating profit margin and the still declining online marketing are major weaknesses; for funds that are willing to price for structural transformation, the 94% growth rate of AI cloud, the triple-digit Token revenue growth of MaaS, and the 53% AI revenue proportion for three consecutive quarters provide sufficient directional evidence.
Conclusion: The "revenue evidence" of transformation is sufficient, and the "profit evidence" is yet to come in 2027
The most important message conveyed by Baidu's 2026 Q3 earnings report is not the level of profit in a single quarter, but that the shift in revenue structure has crossed an irreversible critical point. The AI business has contributed more than half of the general business revenue for three consecutive quarters, which means that Baidu is no longer a "search company that is undergoing transformation", but a company that "takes AI revenue as the mainstay, but whose profit structure is still in transition".
However, there is a significant time lag between the shift of revenue structure and the shift of profit structure. The 22% decline in online marketing is immediate and structural; the 94% growth of AI cloud is real, but its profit margin is far lower than the replaced advertising revenue; investments in model training and computing power infrastructure continue to erode current profits. This state of "revenue has shifted gears but profits have not" will most likely last from the fourth quarter of 2026 to the first half of 2027.
The real verification node is in 2027: whether the revenue scale of AI cloud infrastructure can pull the operating profit margin from the current low single-digit level to a more meaningful level without a significant stall in growth rate; whether the mass production of Kunlun Chip M100 can begin to substantially improve the computing power cost structure; and whether online marketing revenue can show an early signal of stabilizing decline in a certain quarter. Before at least two of these three conditions show verifiable improvement, Baidu's stock will most likely remain in a valuation tug-of-war where "the direction is recognized but the pace is questioned". The 53% AI revenue proportion proves the direction of transformation, but the 2% operating profit margin reminds the market: from "earning revenue like an AI company" to "earning profits like an excellent company", Baidu still has a gap to fill with both execution and time.
This article is from the WeChat official account "Tech Business", written by Bowen, and published with authorization by 36Kr.