Robin Li — Serving the Billion
The number that orients Robin Li's thinking about AI is not the number of parameters in his best model, or the inference cost per thousand tokens, or any of the benchmark scores that dominate the global AI press. It is the number of Chinese people who need to find something, learn something, do something, or decide something — and who will, if the interface is right and the price is accessible, use an intelligent system to help them. That number is in the hundreds of millions, on its way toward most of 1.4 billion. It is not a market segment. It is a civilization seeking to use a new kind of tool.
Li has articulated, more consistently than any other major AI executive in either civilization, the civilizational dimension of deploying AI in China. The frame is not primarily economic, though the economic argument is present. It is demographic and social. A country with 1.4 billion people, a rapidly aging population, a healthcare system under structural pressure, an educational system that is excellent at producing credentialed graduates but less well-optimized for the continuous learning a changing economy requires, and a workforce in the middle of an enormous transition from manufacturing to services. In each of these domains, Li argues, intelligent systems can provide leverage that is simply not available through any other means at an affordable cost.
A doctor using AI to extend her diagnostic capacity can effectively serve more patients than a doctor without it. A teacher using AI to personalize instruction can reach students who would otherwise be underserved. A factory worker using AI to navigate a more complex logistics task can remain productive in a sector that is, without AI assistance, displacing her. The argument is not unique to China — every AI company with a conscience-marketing budget makes some version of it. What distinguishes Li's version is the specificity of the constraints he has accepted and the rigor with which Baidu's strategy operationalizes them.
Chinese-language AI deployment is genuinely harder than English-language deployment. The Chinese linguistic tradition is richer in ambiguity, more deeply embedded in cultural context, and less well-served by the training corpora that have driven Western model development. Baidu's investment in Chinese-language NLP extends back to the company's earliest days, when building effective Chinese-language search required building the Chinese-language text analysis tools that did not yet exist. That investment now provides a foundation for Chinese-language AI deployment that competitors cannot easily replicate.
The regulatory dimension of serving Chinese users adds a further layer of specificity that Li has navigated throughout Baidu's history. The Chinese information environment is governed by a complex set of content standards, data sovereignty requirements, and algorithmic regulations that have no precise equivalent in Western markets. Operating within these constraints is not a neutral technical challenge. It requires ongoing negotiation between what the technology can do and what the regulatory environment permits. Li has, throughout his career, managed this negotiation with more success than many of his peers, partly because he began it early and established trust with regulators before the stakes were high, and partly because his commitment to beneficial AI applications gives him more credibility in those conversations than a purely commercial actor would have.
The billion-person frame has implications for the business model of AI that Li has been exploring. A service that costs ten dollars per month is accessible to a relatively affluent user base. A service that costs one dollar per month, or ten cents per interaction, is accessible to a much larger population — but requires a very different cost structure and a very different approach to monetization. Li has been explicit that the commercial sustainability of AI for mass deployment in China requires cost curves that fall much faster than current inference economics permit. This is one of the reasons his full-stack strategy — including the chip program — is not merely a supply-chain hedge but a genuine attempt to control the cost structure of the inference layer at scale.
The strategic frame Li has placed at the center of Baidu's positioning is the application layer thesis. Foundation models, in the long run, will behave like all foundational technologies: they will commoditize. The cost of inference falls by an order of magnitude every eighteen months. Models improve through architectural refinements, distillation techniques, and the accumulated feedback of deployed applications. Open-source models exert downward pressure on the rents available to closed-source providers. The model layer is, in technical economics terms, a market tending toward contestability. A perfectly contestable market generates no long-run excess returns. The returns, therefore, will eventually flow to whoever has built durable differentiation at the application layer — where user trust, domain expertise, language specificity, and deep integration with real workflows create the kind of switching costs models, by themselves, cannot supply.
This is not an eccentric position. It is broadly consistent with how value has distributed in previous platform revolutions. The mobile internet is Li's favored comparison. In 2007, when Apple introduced the iPhone, the dominant strategic question in the technology industry was which company would own the smartphone platform. By 2015, it was apparent that the platform layer, though enormously profitable, had distributed the largest share of value not to itself but to the applications that ran on it. Uber, Instagram, WhatsApp, and a thousand vertical SaaS businesses created value the platform layer could not have created alone. Li's argument is that the AI moment is structurally identical: the model is the smartphone, and the applications are everything that made the smartphone matter.
Where Li's position diverges from the mobile internet analogy is in the specificity of the claim about China. The applications that will create value in the Chinese AI economy are not the same applications that created value in the American smartphone economy. The Chinese user base has different literacy patterns, different workflow habits, different relationships to privacy, commerce, and government service, and a fundamentally different linguistic and cultural context than the populations for which most American AI applications are optimized. Li's argument is that this specificity is not a constraint but an opportunity. Baidu's advantage is not that it is building better models than OpenAI. Its advantage is that it understands, with two decades of log data and deployment experience, what Chinese users actually need from an intelligent interface — and that understanding is not transferable to an American competitor who arrives with a better model but no context.
Then the most counterintuitive position Li has taken, and the one that most clearly distinguishes him from the open-source consensus that has formed around figures like Zuckerberg in the American AI ecosystem: the closed-source contrarian position. Li believes closed-source models are generally superior to open-source models at comparable scale, and he has said so in terms that leave no room for charitable reinterpretation. His phrase — that open-source AI is "a kind of IQ tax" — was the most quoted thing he said in 2024, and it provoked the kind of reaction that precise contrarian positions always provoke: a mixture of outrage from those who found the framing dismissive, and quiet agreement from those who had been thinking the same thing but found it professionally inconvenient to say so.
The argument behind the phrase is more careful than the phrase itself. Li's claim is not that open-source software is bad, or that open AI research is unimportant. His claim is that the choice to deploy open-source models over closed-source models, in an enterprise context, is often made on the basis of ideological preference rather than technical analysis — and that ideological preference, in this case, leads to worse outcomes. Closed-source models benefit from shared inference infrastructure that allows the provider to serve all customers with a single high-quality deployment. Closed-source models benefit from feedback loops that are, in aggregate, orders of magnitude larger than any individual deployer's feedback loop. Closed-source models tend to have lower inference costs per unit of capability because the provider can amortize the fixed cost of the deployment infrastructure across a very large customer base.
The IQ tax framing implies that choosing open-source over closed-source in this context is not just a neutral preference but a mistake — an error that costs the chooser measurable performance and economic value. This is a characteristically Li move: converting a question that is often treated as a matter of values and political economy (freedom of software, concentration of AI power, dependence on corporate providers) into a question of empirical fact that can be assessed by measuring outcomes.
The position is complicated by Baidu's own open-source contributions and by the release of DeepSeek's R1 model in January 2025 — an open-source Chinese model that demonstrated performance competitive with American frontier models at a fraction of the reported training cost. Li welcomed DeepSeek as proof that Chinese AI development was on a strong trajectory while maintaining Baidu's closed ERNIE models were still superior to DeepSeek at comparable scale. The claim is difficult to verify independently. Baidu has an obvious interest in making it.
What makes Li's framework distinctive in the Chinese cohort — and what distinguishes him from Yi Zeng, Bing Song, Zheng Yongnian, and Kai-Fu Lee, all of whom operate in registers more philosophical or political — is that Li is a builder. He runs a company. He has to make payroll. He has to ship products. He has to navigate a regulatory environment, a competitive market, and a capital structure that all impose binding constraints on what is possible to do and what is possible to say. His articulation of the Chinese AI position is the articulation of a person who has to operationalize the position in working software for hundreds of millions of users every day.
This is the contribution Li makes that the more philosophical voices in the Chinese cohort cannot make. He demonstrates that the Confucian frame — the emphasis on serving the relational fabric, on the substantive theory of the public good, on the specificity of the user as embedded in a particular civilization rather than as a generic individual — can be operationalized in commercial AI development at scale. It is not just a philosophical alternative. It is a strategic alternative with measurable economic consequences. Baidu's product portfolio, the application-layer focus, the full-stack vertical integration, the closed-source insistence on shared inference infrastructure — all of these are the Confucian frame translated into the contemporary AI commercial environment.
The orange pill, applied to enterprise AI strategy, reveals that the question of what to build for whom is not a generic question with a universal answer. The answer depends on what kind of user the system is being built for, and "what kind of user" depends, in turn, on what conception of the user — individual, relational, civilizational — the strategist is operating with. Li's contribution is to have made the Chinese answer to that question commercially legible. Whether the answer prevails in the Chinese market — and whether anything analogous prevails in markets that have not yet been confronted with the same question — is the open commercial question on which Baidu's next decade depends.
[YOU] on AI argues that human value comes from deciding what things are worth building. Li agrees and adds: deciding what things are worth building depends on knowing who you are building for, and "who you are building for" is a substantive answer with civilizational content, not a neutral demographic question. That is the bridge from his position to the broader Confucian frame. It is also the bridge from the Confucian frame to the commercial strategy that has, against significant odds, kept Baidu competitive in a market dominated by American open-source ecosystems and Chinese state-backed competitors.
Li serves the billion. The billion is not a demographic. The billion is a civilization. That is the answer his entire strategic framework is built around, and it is the answer that distinguishes Chinese application-layer AI from anything currently being built in California.