Baidu's AI Revenue Surpasses Advertising for the First Time, but Profit Halved: How Much Longer Can Baidu Hold On?

AI is getting more and more ferocious—how much longer can Baidu hold out? Jiexin October 30, 2026 12:22:30

Author: Jiexin Jiegexin

Let's put two sets of numbers side by side first.

In the first quarter of 2026, Baidu's total revenue was 32.1 billion yuan, a slight year-on-year decline of 1%, and net profit was 3.445 billion yuan, shrinking by more than half year-on-year. But at the same time, the share of AI-related revenue in total revenue exceeded 50% for the first time, surpassing advertising for the first time.

In the second quarter of 2026, total revenue was 31.3 billion yuan, down 4% year-on-year, and net profit attributable to shareholders was 2.319 billion yuan, down 68.33% year-on-year. On the night the earnings report was released, Baidu's US shares fell 12.73%.

This is very much worth pondering.

There must be something wrong somewhere in between.

That statement is correct, but it is not the crux.

Over the past two decades, when users searched for a term on Baidu, ten blue links came up. These ten links were ordered, and the order could be sold. This is the physical basis of bidding ranking. It does not require users to like ads; it only requires users to scroll down.

Bidding ranking is one of the most profitable inventions of the Chinese internet. It turned "users' confusion" into a commodity that could be priced, charged per click, bid on in real time, with extremely high gross margins, and Baidu relied on it for more than a decade.

But the same mechanism also locked Baidu in. If a company makes money from only one thing for twenty years, all its muscles will grow around that thing. So when the ad slots disappeared, what Baidu lost was not a business line, but its only muscle memory.

This is not a case of Baidu doing badly; it is that the carrier of this business itself has disappeared. Pay-per-click presupposes that users have to click, and AI does not give users the chance to click. When users no longer click, CPC as a billing method loses its unit of measurement.

So Baidu's real crisis is not "users have run away," but "users are still asking, but the mechanism that could resell those questions to advertisers is gone."

This year, Baidu App's monthly active users have declined for three consecutive quarters, down by tens of millions year-on-year, and this trend is real. But more fatal than the numbers is the way it is declining: users do not no longer need search; they no longer use "this search called Baidu," because the answers are elsewhere.

Second, Baidu's blood transfusion has been very successful, but what it swapped was gross margin

Next is the second, more hidden problem.

Look at the contrast between two sets of numbers: some analyses split the second quarter into two parts, with AI cloud revenue earning about 2.4 billion yuan more year-on-year, and online advertising earning about 3.1 billion yuan less year-on-year. In the quarterly report, revenue was basically flat, but profit collapsed.

Why? Because these two yuan are not the same kind of money.

Advertising is a typical asset-light, high-gross-margin business. Adding two more ad slots to a search results page costs almost nothing extra. Its gross margin has stayed high for years.

AI cloud is an asset-heavy business. You have to buy cards, build data centers, pay electricity bills, and maintain delivery teams of hundreds of people. Its gross margin is much lower, and it requires upfront capital expenditure. So what Baidu is doing is using one yuan of high-gross-margin revenue to exchange for one yuan of low-gross-margin revenue. The revenue statement looks barely changed, but the profit statement has collapsed by half.

Revenue flat, profit halved—this is not the pain of transformation; it is changing from collecting money while lying down to collecting money while standing and sweating.

This sentence sounds like wordplay, but it determines one thing: if Baidu's transformation succeeds, it will become another company, one with a completely different profit margin and valuation logic.

And this matter may make management more uncomfortable than losing money.

Third, Baidu has never lacked "seeing it five years early"; what it lacks is "lasting until then"

I want to say something fair, and also the place where Baidu is most easily underestimated.

Baidu's judgment on technology has actually always been accurate.

Ernie was among the earliest batch of large models in China; it was not made by following behind others. When Kunlun Chip was making AI chips, many people still thought buying Nvidia was enough. When Apollo Go was running fully driverless cars in Wuhan, most Robotaxi companies around the world were still in the safety-driver stage.

Look at the cards it now holds: the Kunlun Chip P800 has already delivered multiple 10,000-card clusters, and important versions of Ernie 5.1 were trained on fully domestic clusters. Apollo Go completed 3.2 million fully driverless trips in the first quarter of this year, up more than 120% year-on-year; in March, weekly orders peaked at more than 350,000, and cumulative trips have exceeded 17 million. Its overseas fleet exceeds one thousand vehicles; at the end of March it launched fully driverless commercial operations in Dubai, and in August it also connected to Uber. These are not PowerPoint slides; they are real things.

So Baidu's problem has never been that it cannot see. Its pattern is: every time it sees things three to five years earlier than others, and then during those three to five years, it waits for the market to mature.

And its problem has always been here too: it cannot last until then.

In the mobile internet round, it saw it, but WeChat and ByteDance intercepted it. In the O2O round, it saw it, burned tens of billions, and then shut it down. Short video, livestreaming, food delivery—it opened its mouth earlier than others every time, and every time it failed to wait until harvest day.

Baidu's real organizational capability is the ability to "discover opportunities." Its shortcoming is the ability to "boil opportunities down into cash."

There is also corroborating evidence. Reports count that in recent years the cumulative number of internal corruption cases investigated at Baidu has reached more than seven hundred people. This matter has no direct relationship with the business model, but the mechanism it reflects is the same: when the main business stops growing, the internal distribution logic of the organization quietly changes from "making the cake bigger together" to "fighting over the existing cake."

Moreover, the shape of this cycle works in Baidu's favor. Cloud contracts start at three to five years, chip production lines take ten years to pay back, and Robotaxi licenses are long-term authorizations at the city level. These are all slow businesses, and what slow businesses fear most has never been losses—it is switching tracks midway.

IV. Three Ways to Break Through—Which Is Most Reliable

Now let's talk about what to do. I don't intend to offer a pile of correct but useless advice; I'll just name three, and spell out the cost of each.

The first, and the most reliable: turn itself into the water seller.

Cloud plus chips are the most structurally solid assets in Baidu's hands. On full-stack in-house development, few companies in China can do all three layers—chips, cloud, and models—themselves. And this path has two natural tailwinds: first, the compute shortage is a long-term contradiction; second, domestic substitution is a certain direction.

Add to that the rumors of Kunlunxin being spun off for an IPO, and this path also has a layer of monetization room at the capital-market level.

The cost is also clear: once this path works, Baidu will no longer be "a company that sells answers" but "a company that sells compute." The former commands an internet valuation; the latter, a manufacturing valuation. It may earn more money, but the story's multiple will come down.

The second, Apollo Go.

This is the card that makes Baidu least resemble a Chinese internet company. Its operating scale leads globally, and Wuhan already achieved per-vehicle break-even in the first quarter of this year.

But I don't think this is Baidu's lifeline, because its volume is still too small to matter. The figure of 250,000 weekly orders is almost negligible when placed in the national ride-hailing market.

Its value is not in today, but in that it is a ten-year business—and a global one. ZTE back then did not survive by making money in its first year either.

The ad slots are gone, but the demand remains. Users still ask hundreds of millions of questions every day; they just don't want to click links for them.

So only one path remains: from "traffic referral" to "getting things done."

In the past, search sent users elsewhere to complete transactions and earned a toll fee. The search agent of the future will directly complete the task for the user and take a cut of the transaction.

That sounds beautiful, but it means Baidu must rebuild an entire set of capabilities: payments, fulfillment, merchant relationships, after-sales. In the past twenty years, Baidu has not built any of these solidly.

This is the only path that can let Baidu keep its identity as an "entry point," and also the one most likely to fail.

Finally: what it must pass is not a technology test but an identity test

It does not lack technology, it does not lack a head start, and even its luck is not too bad, because this time the cycle is long enough.

What it lacks is a new question: besides selling confusion, what else can I sell?

Advertising companies sell attention; infrastructure companies sell certainty and scarcity. The former makes fast money; the latter makes slow money.

Baidu now stands between the two. It has money, technology, licenses, a fleet, chips—it lacks almost nothing. The only thing it has not yet decided is: whether to give up the identity that made it glorious for twenty years.

That decision is harder than any technology transition.

Baidu's history is in fact a history of "starting earliest and leaving earliest." Every time, it was the first to enter, and every time, it was the first to leave.

Because a failed technology transition can be retried, but if an identity transition fails, you no longer know who you are.