To understand China’s leading AI company Baidu today, it is worth asking a simple question: Why would an internet company spend a decade building something so expensive, so technically demanding, and so slow to pay off?
The answer is artificial intelligence. On September 1, Baidu completed the conversion of its Hong Kong listing from secondary to primary status, making it a dual-primary-listed company alongside its Nasdaq listing.
The move does not itself change the company’s business, nor does it raise new capital. But its timing is significant. AI-related businesses have now accounted for more than half of Baidu’s general business revenue for two consecutive quarters.
AI is no longer simply Baidu’s bet on the future. It is becoming the company’s operating reality. The more important question, however, is what Baidu has spent the past decade building.
Around 2016, the mobile internet was still at the height of its expansion. Traffic, users and monetization dominated the strategy of Chinese internet companies. Li Yanhong, Baidu’s founder and CEO, chose a different path. He began redirecting the company’s resources toward AI and, rather than starting with consumer applications, chose to build from the bottom of the technology stack upward.
Chips. Cloud infrastructure. Foundation models. Applications. It was a particularly demanding route. Chip development requires years of R&D and close coordination across a complex supply chain. Cloud infrastructure requires enormous capital investment. Large models require sustained spending on computing power, algorithms and data. Applications ultimately have to prove their value in the real world.
Failure at any one layer can turn the idea of “full-stack AI” into little more than a corporate slogan. That is why Baidu’s AI strategy is better understood not as a bet on any single model or product, but as an attempt to build an integrated technological system.
The value of full-stack AI does not come simply from owning multiple businesses. It comes from making those businesses reinforce one another. Computing chips determine efficiency. Cloud infrastructure provides the foundation for training and inference. Models turn computing power into intelligence. Applications turn intelligence into user demand and revenue. The requirements generated by applications, in turn, feed back into the development of models, infrastructure and chips. If that cycle works, technology investment stops being merely a cost. It becomes a compounding capability.
Li has described Baidu’s architecture in terms of “chips, cloud, models and agents.” The idea is fundamentally different from the growth flywheel of the mobile internet. In the previous era, scale often came from acquiring users, generating traffic and improving monetization. In AI, the flywheel is more engineering-driven: better computing efficiency lowers training costs; better models expand the range of applications; larger applications generate more real-world demand; and that demand drives further optimization.
This may be the most important, and least visible, part of Baidu’s transformation.
For years, the market was accustomed to viewing Baidu through the lens of search, advertising and the traditional Chinese internet. Under that framework, AI research and infrastructure spending could easily appear as costs weighing on earnings.
But if Baidu is building the infrastructure of an AI company, the same spending has a different meaning. Over the past several years, that transformation has begun moving from laboratories into commercial markets.
Kunlunxin, Baidu’s AI chip business, has continued to advance its commercialization. Baidu Intelligent Cloud has benefited from rapidly increasing demand for AI computing infrastructure. The company’s Ernie family of foundation models has gone through multiple generations of development. AI productivity products and intelligent agents are beginning to attract significant user activity. Apollo Go, Baidu’s autonomous-driving business, has pushed its technology into real-world roads and commercial operations.
These businesses are at different stages of maturity and face very different competitive challenges. But together they form the practical foundation of Baidu’s full-stack strategy.
More importantly, the layers are beginning to interact.
Large-model training creates new demands for computing infrastructure. Computing infrastructure affects the cost and speed of model development. Better models expand the range of AI applications. Real-world applications generate new requirements and data that can feed back into model development.
The potential moat is therefore not simply that one layer is better than its competitors. It is whether the entire system can improve faster because the layers are connected. This is also why AI may demand more patience than the mobile internet ever did.
A consumer application can be tested, adjusted or abandoned within months. Chips, data centers and foundation models operate on entirely different timelines. They require years of accumulated engineering expertise and can absorb enormous resources before producing visible returns.
The difficult part is not recognizing that AI matters. The difficult part is continuing to invest when the payoff remains uncertain. Li has articulated a simple principle for deciding which technologies Baidu should pursue: if success depends heavily on technological advancement and requires long-term iteration, the company should be willing to invest. If technology is not the decisive factor and the attraction is primarily the size of the market, entering the business may not be worthwhile.
That philosophy helps explain why Baidu chose such a difficult path.Of course, technological depth does not automatically translate into commercial success. Full-stack capabilities are not inherently a moat. Chips must compete on performance, cost and ecosystem compatibility. Foundation models face intense competition. Cloud businesses must demonstrate sustainable economics. AI applications must prove that users will return and pay. Autonomous driving must confront safety, regulation and the difficult economics of operating in the physical world.
Ultimately, the market will judge Baidu on those outcomes. But the past decade has established something important: Baidu was not simply chasing an AI trend. It was attempting to build an integrated AI technology stack. That distinction matters as investors reassess the company.
The conversion of its Hong Kong listing into primary status may broaden its investor base. Potential inclusion in the Stock Connect system could bring additional liquidity. A sum-of-the-parts valuation could also allow investors to assess its cloud, chip, autonomous-driving and AI businesses through different industry lenses.
But none of those developments creates value by itself.
The central question is whether the technological system Baidu spent a decade building can generate sustainable growth over the next decade. The first ten years were largely about construction: putting the chips, cloud, models and applications in place and making them work together.
The next ten years will be about proving the economic value of that integration. That is the real meaning of full-stack AI. It is not a list of businesses or a collection of technology buzzwords. It is a willingness to spend years building the hardest capabilities first, and to wait for technology, products and economics to converge.
A decade ago, Baidu made its most consequential decision by betting on AI. Today, the harder test begins: turning that technological foundation into durable growth. Time has already shown that patience can be a competitive advantage. The next chapter will require something more difficult, not merely patience, but returns.
Source; baidu cloud, leidacj, 36kr, gdec
