Baidu spent two years telling investors it was becoming an AI company. On 18 August it published a quarter where that claim turned into arithmetic: RMB 12.5 billion (≈US$1.74 billion / HK$13.6 billion) of revenue from what it calls its core AI-powered business (AI业务), or roughly half of its general business for a second consecutive quarter.
The other side of the ledger is less comfortable. What Baidu calls its legacy business — traditional search and feed advertising, plus the older internet products attached to it — brought in RMB 10.4 billion, down 23% year on year. The two blocks do not add up to all of general business: a residual sits in lines Baidu does not break out, which is why "roughly half" is the honest description rather than an exact split. Total revenue was RMB 31.3 billion (≈US$4.35 billion / HK$34 billion), 4% lower than a year earlier.
That is the shape of the problem. Baidu is not growing its way into an AI business. It is swapping one business for another, and the replacement is smaller, far more capital-hungry, and not yet earning the margin the old one did.
What actually grew
Strip out the labels and Baidu's AI revenue is three very different businesses wearing one name:
- AI Cloud Infra (AI云基础设施): RMB 7.3 billion (≈US$1.01 billion / HK$7.9 billion), up 50% year on year. Inside it, GPU cloud (GPU云) revenue rose 283%, accelerating from 184% growth in the previous quarter.
- AI Applications: RMB 2.5 billion, up 3%. Essentially flat.
- AI-native marketing services: RMB 2.6 billion, flat year on year.
So the fastest-growing part of the AI story is not software. It is renting out graphics processors by the hour against multi-year contracts, in a market where Alibaba Cloud, Tencent Cloud, Huawei Cloud and a queue of smaller GPU providers are all adding capacity at the same time.
The parts that look like a software business — the applications Baidu bundles for consumers and offices, including DuMate, whose enterprise edition launched in June 2026 — grew 3%. Baidu's own chief executive, Robin Li, has been candid that the company made mistakes in earlier model generations and is now reorganising to iterate faster.
The capex bill behind the GPU cloud
GPU rental is a good business to be in when supply is short. It is a punishing one to be in when you have to buy the hardware yourself.
Morgan Stanley lifted its 2026 capital expenditure estimate for Baidu from RMB 24 billion to RMB 35 billion (≈US$4.86 billion / HK$38 billion), according to Chinese financial press coverage of the note. That is about a third of the roughly RMB 100 billion of general business revenue Baidu generates in a year, spent on assets that depreciate on a three-to-five-year clock while chip generations turn over faster.
Baidu does have room to spend. Operating cash flow was positive RMB 3.4 billion in the second quarter, the fourth consecutive quarter in the black, and the company still holds a large cash pile. But a cash-generating search engine and a cash-consuming compute landlord are different animals, and the market prices them differently.
The advertising line is not coming back the way it left
Online marketing revenue was RMB 13.1 billion (≈US$1.82 billion / HK$14.2 billion), down 19% year on year. Its share of Baidu's general business fell from 62% to 52% in twelve months.
Part of that is macro weakness in Chinese advertising. Part of it is a deliberate product decision: Baidu has pushed AI-produced answer blocks into search results, and answer blocks do not carry the same density of ad slots as ten blue links. Users get a better experience; the company gets a smaller inventory to sell.
The audience has not gone anywhere. Baidu App had 644 million monthly active users in June 2026. The gap between 644 million users and a shrinking ad line is the whole strategic question in one sentence: Baidu still owns the demand, but it has not yet found the format that monetises it as well as the old one did.
What the next set of numbers will show
Baidu reports third-quarter results on 16 November. Notes published in late October by CLSA, Nomura and Goldman Sachs point in the same direction: core revenue of about RMB 24.7 billion (≈US$3.43 billion / HK$26.8 billion) for the quarter, up around 1%; AI cloud infrastructure around RMB 8.0 billion, up more than 90%; and online marketing around RMB 11.8 billion, down 22% to 25%.
If those estimates hold, AI revenue would overtake advertising revenue outright for the first time. That would make the crossover look structural rather than a one-off quarter.
It is worth being precise about what those figures are: analyst forecasts, not reported results. Baidu has not published third-quarter numbers at the time of writing, and forecasts at this level of detail are routinely off by several percent in either direction.
Why this matters outside China
The same experiment is running at Google, and the two are not at the same stage. Google's search revenue is still growing while it inserts AI answers; Baidu's is shrinking. Baidu moved earlier and harder into AI-produced results, which means it is absorbing the cannibalisation cost earlier, and in a weaker ad market.
For anyone outside China, Baidu is the cleanest available case study of what happens when a search monopoly decides to disrupt itself on purpose. The lesson is not that the strategy fails. It is that the timing of the cost and the timing of the replacement revenue do not line up, and the gap has to be financed with capital expenditure at exactly the moment advertising is weakest.
Honest limitations
This article relies on Baidu's own second-quarter release and on analyst notes reported by Chinese financial media for the third quarter. It has no access to Baidu's GPU cloud contracts, customer concentration, or gross margins — none of which the company discloses. Segment definitions are Baidu's own, and its reorganisation of reporting lines means year-on-year comparisons are cleaner than three-year comparisons. We did not model depreciation schedules, which materially affect how profitable GPU rental looks in any given quarter. Nothing here is investment advice.
What readers can do now
- If you are tracking the "AI eats search" thesis, watch the advertising line rather than the AI line. AI revenue is flattered by a low base and by hardware resale economics; the ad decline is the number that tells you whether the core franchise is holding.
- If you buy cloud capacity in China, check the contract structure. Multi-year GPU commitments with prepayment are becoming standard, and the exit and price-reset clauses matter more than the headline discount.
- If you build AI products, note that Baidu's applications line grew 3% while its infrastructure line grew 50%. Capability is not the constraint. Distribution and willingness to pay are.
