Baidu's AI infrastructure gains momentum amid advertising decline

Baidu's strongest card right now is not Wenxin, but its AI infrastructure.

Author and source: 0x9999in1, ME News



• Baidu is no longer facing the question of whether to shift toward AI, but whether AI growth can promptly fill the gap left by the decline in traditional search advertising. In the second quarter of 2026, Baidu’s core AI business revenue reached RMB 12.5 billion, a 25% year-over-year increase, and has accounted for approximately half of Baidu’s core business revenue for two consecutive quarters; however, the company’s total revenue still declined by 4% year-over-year.
• Baidu's strongest asset right now is not Ernie Bot, but its AI infrastructure. In the second quarter, AI cloud infrastructure revenue reached RMB 7.3 billion, a 50% year-over-year growth, with GPU cloud revenue rising 283% year over year; in comparison, AI application revenue was RMB 2.5 billion, up only 3% year over year, and AI-native marketing revenue was RMB 2.6 billion, roughly flat year over year. Baidu’s AI transformation already has a solid foundation, but the upper layers have yet to truly take off.
• Li Yanhong proposed letting Wenxin return to the top tier of foundational models; what matters is not merely a declaration of catching up, but Baidu’s strategic recalibration: moving away from pursuing leadership in every dimension, and instead focusing on core products such as search, digital humans, Miaoda, and general-purpose agents, reintegrating real user feedback into model training.
• This strategy is logically sound. Baidu possesses a full stack of resources—search, Baike, Netdisk, cloud computing, autonomous driving, Kunlun chips—that few Chinese tech companies can match. But the question remains sharp: having more resources doesn’t mean faster iteration; more use cases don’t guarantee stronger products. In the AI era, the most brutal metric is speed.
• Therefore, Baidu’s next decisive move is not to release another model with larger parameters, but to establish a rapidly accelerating feedback loop: model upgrades → product improvements → user growth → increased inference requests → higher revenue → reinvestment into models. Once this flywheel starts spinning, Baidu still has a chance; if it doesn’t, even the most comprehensive stack of capabilities will merely be an expensive list of assets.

Li Yanhong said this time that he wants to bring Wenxin back to the top tier of foundational large models.

What truly matters is not the phrase “top tier,” but rather “returning to.”

This means Baidu has not avoided a reality already felt by the market: although Baidu was an early entrant in China’s current wave of rapid foundational model iteration, it has not consistently maintained the strongest voice or pace. Li Yanhong himself acknowledged that Baidu is among the first companies in China to invest in foundational models, and along the way, “there have been missteps”; he also believes that the competition among models is far from over, as new models may lead in certain capabilities every few months. Therefore, Baidu will continue to invest resources, restructure its organization, and attract top AI talent to accelerate the iteration of Wenxin.

Three years ago, this statement was merely an expression of a technical roadmap.

By August 2026, it will already be carrying significant financial pressure.

Baidu's total revenue for the second quarter was RMB 31.3 billion, a 4% year-over-year decline; net profit attributable to Baidu was RMB 2.3 billion, significantly lower year-over-year. Baidu's core business revenue amounted to RMB 25.2 billion, with online marketing services revenue at RMB 13.1 billion, down 19% year-over-year. Meanwhile, Baidu's traditional business revenue was RMB 10.4 billion, down 23% year-over-year. On the other hand, core AI business revenue reached RMB 12.5 billion, up 25% year-over-year, accounting for 50% of Baidu's core business revenue.

Putting these numbers together almost spells out Baidu’s current situation on the wall:

The old engine is losing RPM, the new engine has ignited, but the thrust is still insufficient.

In the past, Baidu could discuss when AI would contribute to revenue; now, it cannot. After the continuous contraction of its advertising business, AI is no longer a distant prospect—it has become a new core business that must be rapidly scaled up.

Therefore, understanding Baidu's AI strategy today cannot focus solely on ERNIE or merely ask "What rank is the model?"

The real question is: Can Baidu successfully replace the engine of its business system without stalling mid-flight?

Objectively speaking, Wenxin is not a product line that has fallen so far behind technologically that it is no longer visible.

In its Q1 2026 earnings report, Baidu disclosed that it launched ERNIE 5.1 in May, enhancing its text and reasoning capabilities while reducing model size. According to the LMArena rankings disclosed by Baidu at the time, ERNIE 5.1 ranked first among Chinese models on the text benchmark and first among Chinese models and fourth globally on search-related benchmarks. While such rankings can fluctuate rapidly and this leadership does not directly equate to long-term model dominance, it at least demonstrates that Baidu still possesses the technical capability to push specific capabilities to the forefront.

The problem is that today’s large model competition is no longer a once-a-year college entrance exam.

It's becoming more like F1.

Getting pole position this time isn't meaningless—of course it is meaningful. But next time, the architecture changes, the data changes, the inference paradigm shifts, and Agent capabilities are completely reshuffled; yesterday’s advantage could vanish quickly.

This also explains why Li Yanhong did not emphasize this time that "Wenxin's capabilities must all be number one."

The direction he proposed is more pragmatic: Baidu will prioritize enhancing capabilities most critical to its own applications—such as AI search, digital humans, Miaoda, and general-purpose agents—and then feed behavioral data and performance feedback from real-world products back into model training.

This is actually the most significant shift in Baidu's AI strategy.

In the past, the industry tended to fall into a form of "model-centric thinking": if the model is strong enough, applications will naturally emerge.

By 2026, this logic is clearly no longer sufficient.

The real competition is becoming: who has larger real-world scenarios, who can obtain higher-quality feedback, and who can faster turn that feedback back into model capabilities.

What Baidu should truly bet on is precisely this closed loop, rather than simply chasing rankings.

Why does Baidu still have the right to talk about "coming back"?

Because search remains an exceptionally unique asset.

A user saying “Help me write an article” in a chatbot may or may not be satisfied with the result—this satisfaction is sometimes hard to quantify. But search behavior is different. What the user types, what they click on, whether they ask follow-up questions, whether they revise their keywords, whether they exit quickly, and whether the response resolved their issue—all of these form continuous feedback signals.

This is precisely the logic Li Yanhong emphasized on the call: after Wenxin improves user intent understanding and content quality assessment, it can immediately be integrated into search and information streams; once the product team observes the results, they identify issues and feed relevant data back into model training.

The model isn't trained in isolation in a lab and then released; instead, it enters real-world products and continuously learns from every challenge it faces.

This is where Baidu is most likely to establish differentiation.

In the second quarter, Baidu further integrated its AI search with Wenxin Assistant, extending single-turn search responses to multi-turn conversations, while continuing to enhance tool usage, multi-step planning, and complex task execution. Baidu’s management stated that user satisfaction, search intent, and retention are improving; however, the company is still deliberately slowing the monetization of AI search to prioritize user experience, leading to continued pressure on its traditional advertising business until the second half of 2026.

This is, in fact, a very painful self-revolution.

The previous business model for search result pages was: users generate demand, advertisers purchase traffic, and Baidu facilitates the matching.

The endgame of AI search may very well be: users pose questions, and AI directly performs analysis, comparison, planning, and even executes tasks.

Here’s the question—if answers become increasingly complete, why would users still click so many links? And if traditional clicks decline, what happens to the original ad placements?

So Baidu is currently experiencing a classic "innovator's dilemma": the more seriously it transforms search, the more likely it is to inadvertently harm its old search business model.

But it must be changed.

Because if Baidu doesn’t fix its old search itself, someone else will do it for them.

If you look only at current revenue, the most compelling part of Baidu AI is not ERNIE nor Agent, but the cloud.

In the second quarter of 2026, Baidu AI Cloud's infrastructure revenue reached RMB 7.3 billion, a 50% year-over-year increase; among this, GPU cloud revenue grew by 283% year-over-year, following a 184% increase in the first quarter. Management stated that GPU cloud has achieved triple-digit growth for four consecutive quarters.

This is very important data.

Because it demonstrates that at least some AI demand is genuinely reflected in Baidu’s financial statements, rather than remaining confined to product launches, model rankings, and developer demos.

Especially as AI transitions from the training phase to larger-scale inference, what businesses truly need is not just "renting a few GPUs," but overall efficiency across models, inference frameworks, chips, scheduling, and cloud services.

This precisely aligns with Baidu's long-standing emphasis on "full-stack AI."

At the bottom is Kunlun芯, above which is cloud infrastructure, and above that are Qianfan, Wenxin models, and applications. In theory, Baidu can perform end-to-end optimization across chips, frameworks, models, and inference to reduce costs and improve utilization.

In the second quarter, management also explicitly stated that GPU cloud profit margins are higher than those of traditional CPU cloud; with the increasing proportion of GPU revenue, improved resource utilization, and expanded scale of model-as-a-service usage, AI cloud profit margins still have further room for growth.

If this line continues to generate revenue, Baidu will have a very solid AI cash flow entry point.

But it also reveals another issue:

Baidu's fastest-growing business is now "selling shovels," rather than "digging for the biggest gold mine."

In the second quarter, Baidu's AI application revenue reached RMB 2.5 billion, a year-over-year growth of only 3%; AI-native marketing services revenue was RMB 2.6 billion, flat compared to the same period last year.

When placed alongside the 283% growth of GPU cloud, the contrast is stark.

This is the most crucial data when assessing how far Baidu AI has progressed.

Rapid cloud growth indicates that the industry requires computing power; it does not automatically prove that Baidu already possesses the next-generation super app.

For an internet giant with hundreds of millions of users, what’s truly exciting isn’t just the existence of “AI business,” but rather one or more AI products beginning to drive platform-level growth.

No such surge has been observed yet.

Of course, Baidu has no shortage of product offerings. Wenku and Wangpan continue to integrate AI features, GenFlow is undergoing continuous iteration, Miaoda is expanding into AI-native application development, and after DuMate's launch in March 2026, its enterprise version was released in June. Baidu disclosed that in June, the AI daily active user penetration rate in Wenku and Wangpan increased by 27.4% year-over-year.

These numbers show that users are indeed using AI.

However, there is still a long way to go between using AI features and generating an entirely new growth curve.

Baidu needs to answer: Is there an AI product that can help users form high-frequency habits that didn't exist before? Can an agent truly integrate into enterprise workflows and generate ongoing revenue? Can digital humans evolve from demonstration technologies into scalable commercial tools? Can Miaoda become more than just “generating an app” and instead build its own development and distribution ecosystem?

If these questions remain unanswered, even the most robust foundational models will struggle to achieve meaningful repricing.

Baidu has two other assets worth reevaluating from a time-based perspective.

One is Kunlun芯, and the other is Luobo Kuai Pao.

Their commonality is that none of them were hastily created after the surge of generative AI; instead, they have been built through long-term technological investment.

As of the second quarter of 2026, Baidu stated that Kunlun芯 has completed the development and commercialization of three generations of AI chips and continues to advance subsequent products designed for large-scale inference. Management emphasized that the significance of Kunlun芯 extends beyond selling chips independently—it enables joint optimization with Baidu Cloud, models, and applications, strengthening Baidu’s cost and supply chain control capabilities amid sustained growth in domestic AI computing demand.

RoboTaxi represents another type of AI: not chat-based, but AI directly entering the physical world.

In the first quarter of 2026, RoboTaxi completed 3.2 million fully driverless rides, representing over 120% year-over-year growth. By April, the total number of public ride orders provided exceeded 22 million. In the second quarter, Baidu disclosed that RoboTaxi’s operations have expanded to 28 cities worldwide, with the fleet accumulating over 350 million kilometers of autonomous driving, including more than 240 million kilometers of fully driverless driving, and is advancing testing or commercial operations in markets such as London, Dubai, Hong Kong, and Switzerland.

It is difficult to definitively judge how far it is from becoming a profit pillar on Baidu’s financial statements.

But it at least shows one thing: Baidu is not just an AI company with language models.

From chips to cloud, from models to search and autonomous driving, Baidu has many cards in hand.

And this is precisely both an advantage and a risk.

Full-stack sounds great.

But full-stack companies have an inherent issue: complexity.

Chips require resources, models require resources, cloud services require resources, search needs to transform, agents are fighting for entry points, autonomous driving aims to expand overseas, digital humans need to be commercialized—each of these areas alone is enough for a startup to bet its entire future on.

Baidu doing everything means it must have stronger organizational coordination than others.

Otherwise, so-called full-stack can easily become "everything is there, but nothing is fast enough."

This is what Baidu needs to be most cautious about today.

In the AI era, company size and historical accumulation still matter, but the weight of speed is rising sharply. Models can undergo capability shifts every few months, product interfaces are continuously redesigned, developers quickly migrate to platforms with lower costs and better performance, and users have no patience for a tech giant to complete internal coordination.

Reuters noted after Baidu's second-quarter earnings report that ERNIE Bot had lacked significant upgrades for some time, while other Chinese model providers continued to roll out new versions; at the same time, Baidu's declining advertising revenue means it must simultaneously bear the pressures of investing in AI talent and infrastructure, as well as its traditional business.

Therefore, Li Yanhong emphasized that adjusting the organization and bringing in top talent is more important than simply returning to the top tier.

Because what Baidu may have lacked all along was never a more beautiful strategic PowerPoint.

What it lacks is the ability to compress six months into three months, and three months into six weeks.

So, can Baidu be unblocked again?

I still believe the answer is: yes, but the window won't stay open forever.

Baidu possesses cash, users, search, cloud, chips, models, and autonomous driving—resources that most domestic AI startups cannot simultaneously hold. By the end of June 2026, Baidu’s total cash and investments still amounted to RMB 283.1 billion, providing it with a substantial buffer to continue competing in this highly capital-intensive arena.

But resources can only give you entry to the competition, not the championship.

In the next one to two years, you can determine whether Baidu's AI has truly transformed by watching a few very straightforward questions.

Has the major update to Wenxin significantly improved speed? Is AI search enhancing user experience while discovering new business models? Can AI application revenue move beyond single-digit growth? How long can the triple-digit growth rate of GPU cloud services be sustained? Can Kunlun Chip truly generate scalable external commercial revenue? Will Luobo Kuipao’s global expansion ultimately achieve a healthy unit economics model?

The most important thing, however, is one word: closed loop.

Stronger models lead to better search experiences; more high-quality feedback from search and agents further improves the models; increased model usage drives demand for cloud infrastructure; expanded cloud scale reduces inference costs; lower costs spur more applications; applications generate revenue, which is reinvested into chips and models.

If this circle were to truly come to life, Baidu’s accumulated assets over the past two decades—which now seem somewhat scattered—would suddenly reconnect.

At that time, Baidu was no longer just a company with search, cloud, chips, Wenxin, and Robotaxi—it was a truly self-reinforcing AI machine.

But if this circle is always missing a piece—the model exists, but applications haven’t exploded; cloud services have grown, yet search revenue continues to shrink; there are many products, but no new user entry points—then Baidu isn’t facing the question of “Where does Wenxin rank?” but something far more brutal:

Can a company that once defined China’s internet search gateway reinvent itself a second time before AI redefines the gateway?

Therefore, "bringing Wenxin back to the top tier" is certainly important.

But what Baidu truly aims to reclaim is never just a position on a model ranking chart.

It's racing against time.

In this round of AI competition, time is precisely the most expensive and irreplaceable resource.

1. Baidu, Inc.: "Baidu Announces Second Quarter 2026 Results," August 18, 2026, Baidu Investor Relations website.
2. Baidu, Inc.: "Baidu Announces First Quarter 2026 Results," May 18, 2026, Baidu Investor Relations website.
3. Baidu, Inc.: "Baidu Announces Fourth Quarter and Fiscal Year 2025 Results," February 26, 2026, Baidu Investor Relations website.
4. Sina Tech: "Full Text | Baidu Q2 Earnings Call Transcript: Wenxin to Return to the Top Tier, Complete Primary Listing in Hong Kong by Year-End," August 18, 2026.
5. Blue Whale News: "Baidu's Li Yanhong: Aiming to Return Wenxin to the Top Tier," August 18, 2026.
6. Reuters: "Baidu CEO Vows to Return Ernie to AI Frontier as Revenue Miss Sinks Shares," August 18, 2026.
7. TIME: “We’re Not That Far Behind.” Baidu’s Robin Li on China’s Push to Diffuse AI Throughout Society, January 25, 2026.