The Democratisation of AI

China’s Alternative Path

(This article was originally published in the magazine Latin America on the move No. 559 )

In July 2025, two visions for artificial intelligence emerged from the world’s technological superpowers. The United States released “Winning the Race: America’s AI Action Plan,” framing AI as a zero-sum competition where America must achieve “unquestioned and unchallenged global technological dominance.” Days later, China unveiled its “Global AI Governance Action Plan,” positioning AI as “an international public good that benefits humanity” and calling for inclusive development that supports the Global South. One commentator described the American approach as “a digital Monroe Doctrine”; China’s read like a manifesto for technological multilateralism.

Six months later, the divergence has become concrete. Chinese open-source AI ecosystems have grown to dominate global downloads. Government regulation has prevented monopolistic concentration while steering technology toward public benefit. And a national strategy called “AI+” has pushed artificial intelligence into factories, farms, and energy grids—not just the virtual playgrounds of Silicon Valley. What emerges is not merely a different set of policies, but a different answer to a fundamental question: who should AI serve?

The answer, China’s practice suggests, is everyone. This democratisation unfolds across three dimensions: open-source technology that lets all nations participate, governance that prevents capture by the few, and applications that benefit all of society.

Opening the Table: Technology Democratisation Through Open Source

The Silicon Valley narrative has long held that frontier AI requires frontier capital—billions of dollars, tens of thousands of cutting-edge chips, proprietary moats protecting competitive advantage. DeepSeek’s trajectory challenges each assumption.

In January 2025, the Hangzhou-based company released its R1 model under an MIT licence, permitting unrestricted use, modification, and commercial deployment. The model matched capabilities that American competitors had achieved only through massive resource concentration—yet DeepSeek trained it for approximately $5.5 million using 2,000 GPUs, roughly one-hundredth the cost of comparable American efforts. By December, DeepSeek-V3.2 had earned gold medals at both the International Mathematical Olympiad and the International Olympiad in Informatics, competitions where only about 8% of human participants achieve gold.

Hugging Face, the leading platform for open AI models, documented the aftermath. What it termed the “DeepSeek Moment” triggered an explosion of open-source activity. Alibaba’s Qwen family reached 700 million cumulative downloads, becoming the world’s most widely used open-source AI ecosystem—surpassing Meta’s Llama. Baidu went from zero releases on Hugging Face in 2024 to over 100 in 2025. Chinese newly created models now exceed American ones in global downloads.

Open-source matters not merely because code becomes available, but because knowledge circulates. DeepSeek published its methods in Nature, subjecting them to peer review that confirmed reproducibility—breaking the black box that had kept frontier AI opaque. This transparency enabled a global community of researchers and developers to build upon the work, creating what Hugging Face researchers called “the ability to replicate itself.” Competition shifted from raw capability to ecosystem integration; the question changed from “can we do this?” to “how do we do this well?”

The cost implications are equally transformative. DeepSeek’s API charges $0.28 per million input tokens—roughly sixteen times cheaper than comparable American services. What once required national-level resources is now accessible within realistic budgets for universities, startups, and government agencies across the developing world.

For nations previously excluded from AI development, the implications are immediate. India has announced plans to build local models based on DeepSeek technology, with initial applications focused on agriculture and climate adaptation. Brazil’s $4 billion AI plan emphasises domestic models and compute. The door that resource constraints had kept closed has opened—not through charity, but through engineering that makes participation economically viable.

Preventing Capture: Governance Democratisation Through Regulation

Technology alone does not guarantee democratic outcomes. Without governance, AI’s benefits concentrate among those who control it. China’s regulatory approach—often characterised in Western media as authoritarian restriction—functions in practice as anti-monopoly enforcement.

Consider the pattern. Ant Group’s proposed IPO was halted in 2020 amid concerns about systemic financial risk and data concentration. Didi faced investigation in 2021 after its US listing raised questions about cross-border data flows. These interventions prevented the formation of private data monopolies that, left unchecked, would have accumulated unprecedented power over citizens’ information and economic activity.

The regulatory philosophy extends beyond reactive enforcement. China’s AI content labelling requirements, effective September 2025, mandate clear identification of AI-generated text, images, and video—a transparency measure that addresses concerns about synthetic media without prohibiting the technology. The physical “kill switch” designed into AI agent hardware reflects a pragmatic approach to autonomy: enable capability while preserving human override. The pattern is “pilot first, legislate later”—allowing innovation while developing evidence-based governance.

China’s Global AI Governance Action Plan codifies this philosophy internationally. Its thirteen points call for building cross-border open-source communities, lowering thresholds for technological innovation, and supporting developing countries in building AI capacity. The explicit commitment to assist the Global South in “truly accessing and utilising AI” contrasts sharply with American export controls designed to restrict access.

The contrast illuminates different theories of how technology serves public interest. The American approach trusts market competition among private giants to deliver innovation; the Chinese approach treats unchecked concentration as the threat requiring active intervention. Neither is neutral; both reflect political choices about who technology should empower.

For developing nations watching this divergence, the governance question is not abstract. Data extracted by American platforms from Latin American users generates value captured elsewhere—what some analysts have termed “data colonialism.” China’s framework, whatever its domestic motivations, offers a vocabulary and precedent for treating data as a sovereign resource requiring protection.

Serving All: Application Democratisation Through “AI+”

Perhaps the starkest divergence concerns what AI actually does.

American AI development has concentrated in virtual domains: coding assistants, text generation, video synthesis. These applications serve software developers and content creators—valuable, but narrow. The real economy of agriculture, manufacturing, and energy remains largely untouched. When Oracle announced a $300 billion commitment to OpenAI, analysts noted the figure was discounted directly to current-period revenue—a bubble warning if ever there was one.

China’s “AI+” strategy pursues a different trajectory. Announced as national policy in 2025, it mandates AI integration across energy, manufacturing, agriculture, ocean industries, and logistics. The implementation is concrete: CATL deploys AI agents for 24-hour visual inspection in battery manufacturing; Mengniu uses AI monitoring for livestock health and feeding; DingTalk’s enterprise platform serves over 3 million companies with AI-powered workflow automation.

The scale is significant. China’s AI core industry exceeded one trillion yuan in 2025. Over 600 national-level intelligent computing centres now operate across the country. Two hundred million industrial robots work in Chinese factories—more than any other nation.

The underlying logic treats AI not as a product category but as infrastructure. Just as electricity transformed every industry it touched, AI integration aims to raise productivity across the entire economy. The benefits accrue not to a few platform companies but to manufacturers, farmers, and service providers who adopt the technology.

This represents a fundamentally different answer to the question of who AI serves. In the American model, AI primarily empowers technology companies and their shareholders; users are customers, and often products. In the Chinese “AI+” model, AI empowers the broader economy; technology companies are enablers rather than extractors. This is application democratisation: AI serving all of society, not just those who build it.

What This Means for Latin America

Latin America stands at a crossroads. The ILIA 2025 index, published by ECLAC and Chile’s CENIA, documents the region’s position: Latin America attracts only 1.12% of global AI investment despite representing 6.6% of global GDP. Yet the region ranks third worldwide in downloads of generative AI applications, suggesting appetite far exceeds current capacity.

The democratisation unfolding elsewhere offers lessons, though not a template. Consider Cuba’s Cecilia project. Developed by the Artificial Intelligence Research Group at the University of Havana—in a country experiencing twenty-hour blackouts and unstable internet—Cecilia is a 2-billion-parameter language model trained specifically on Cuban text: ten years of newspapers, the national encyclopedia, four hundred literary works, collections of Cuban laws, and popular song lyrics. The team built on Salamandra, an open-source Spanish model from Barcelona, adapting it through continual pretraining to capture Cuban linguistic and cultural nuances.

Cecilia required no Chinese infrastructure, no American cloud services, no billion-dollar budgets. What it required was open-source foundations that could be adapted locally, academic collaboration with Spain’s University of Alicante, and the determination to build something that serves Cuban society rather than importing solutions designed elsewhere. The model is released under a Creative Commons licence, allowing others to learn from and build upon the work. This is AI democratisation in practice: constrained resources producing genuine capability through intelligent adaptation rather than brute-force scaling.

Yet clear-eyed assessment requires acknowledging limitations. As ECLAC’s executive secretary noted, closing gaps in “infrastructure, talent, innovation and governance” remains essential—open-source models alone cannot substitute for genuine local capacity. Hardware dependencies persist; chips and cloud platforms remain controlled primarily by American and, increasingly, Chinese corporations. The path to digital sovereignty is longer than downloading a model.

The choice between Washington’s AI vision and Beijing’s is, as one Brookings scholar observed, “itself a trap.” What matters is not which superpower to follow but whether Latin America develops the indigenous capabilities that make choice meaningful. Regional coordination—shared standards, pooled research resources, joint procurement—offers leverage that individual nations cannot achieve alone.

China’s experience demonstrates that AI democratisation is possible—through open technology that lowers barriers to participation, governance that prevents capture by the few, and applications that serve all of society rather than narrow interests. The rules of the game are being rewritten. The question for Latin America is not whether to play, but on whose terms.