The Real Reason Chinese AI Is Winning the Global South

Chinese open-weight models—artificial intelligence systems whose underlying weights are publicly released—are having a moment. Alibaba’s Qwen, for example, boasts 151,448 derivatives on Hugging Face. These are customized adaptations, where programmers fine-tuned the Qwen architecture for their specific needs. That gives Qwen a repository footprint 4.7 times that of Meta’s Llama and almost double that of Google’s Gemma. It also draws 39.6 million monthly downloads of GGUF, a lightweight format optimized for local hardware, far ahead of Llama’s 7.5 million. These numbers should not be dismissed. Many of the top U.S. AI firms’ business models, especially at OpenAI and Anthropic, still largely follow a walled-garden approach, whose access is metered through APIs and subscriptions. That is a harder sell in cost-sensitive markets across the global south. Chinese open-weight models—artificial intelligence systems whose underlying weights are publicly released—are having a moment. Alibaba’s Qwen, for example, boasts 151,448 derivatives on Hugging Face. These are customized adaptations, where programmers fine-tuned the Qwen architecture for their specific needs. That gives Qwen a repository footprint 4.7 times that of Meta’s Llama and almost double that of Google’s Gemma. It also draws 39.6 million monthly downloads of GGUF, a lightweight format optimized for local hardware, far ahead of Llama’s 7.5 million. These numbers should not be dismissed. Many of the top U.S. AI firms’ business models, especially at OpenAI and Anthropic, still largely follow a walled-garden approach, whose access is metered through APIs and subscriptions. That is a harder sell in cost-sensitive markets across the global south. Chinese open-weight models provide a feasible alternative. Analysts at the Stanford Institute for Human-Centered Artificial Intelligence have already pointed to this as a next generation of digital infrastructure with geopolitical implications built in. But the key difference from, for example, the Belt and Road Initiative, China’s much-ballyhooed global investment initiative, is that Chinese AI does not need to be pushed by the central government. If anything, Beijing is being pulled in by developers globally. Adoption began well before Chinese President Xi Jinping delivered his now-famous pitch to global south leaders at the Shanghai World AI Conference in July. This is critical: The geopolitical influence may be more organic and therefore much harder to dislodge. There is a useful precedent here from an earlier generation of open-source technology. The open-source operating system Linux, an alternative to Windows, revolutionized enterprise computing without directly selling the software itself. Companies such as Red Hat built profitable businesses around enterprise subscriptions and services layered on top of the Linux ecosystem. By giving users the flexibility to adapt the system, Linux became an indispensable global standard. Chinese open-weight AI models could follow a similar path, effectively becoming the Linux of the generative AI era while driving demand for the hardware and cloud services needed to run them locally. Many are already doing this. In fact, many Chinese big tech companies, Alibaba most notably, are simultaneously open-weight model developers and cloud providers. Interestingly, this playbook is not unique to open source. It is already visible in another technology where China leads: electric vehicles. AI may be next. The Chinese EV story in Africa, for example, is not as simple as a story of dumping Chinese finished vehicles. Much of the action is in commercial “boda-boda” motorcycle taxis, tailored to local realities of battery access, electricity supply, and price sensitivity. In sub-Saharan Africa, grid instability is a major constraint, and many operators cannot afford to buy a finished EV that requires reliable grids and high upfront costs. Chinese suppliers have approached this market by providing modular components such as lithium batteries to local companies such as Zeno Uganda. Zeno then uses those parts to build a localized “Battery-as-a-Service” battery-swapping ecosystem around local infrastructure constraints. In Ghana, start-ups such as SolarTaxi import Chinese solar panels, batteries, and electric motors and then use them to modify affordable two-wheelers and tuk-tuks for local demand. This is the Lego-like model. Chinese technology arrives as swappable components; local businesses turn those parts into products that fit their markets, often bundling them with local microfinance to enable grassroots access and training local teams in maintenance and technology development. The adoption logic of open-weight AI in the global south is similar and in some ways goes one step further. The AI models are Lego blocks that developers can use, adapt, assemble, and, importantly, innovate on top of to create entirely new products. That creates an enormous opportunity for local economies to benefit from global AI development while also helping to build local technical talent. Chinese tech companies developed their products under tight computing and hardware constraints, pushing them toward algorithmic efficiency and models that can run locally on less demanding hardware. This is an advantage in power-constrained regions that cannot sustain large-scale compute clusters. Chinese AI architectures also perform well with multibyte, non-Latin scripts such as Burmese, Indonesian, Malay, Filipino, Tamil, Thai, and Vietnamese. Developers with access to the architecture can train regional-language models and fine-tune them for local use. Obviously, cost matters, too. Many Chinese models are dramatically cheaper to access than the U.S. proprietary models, while open weights also give developers the option to self-host. For small developers operating on tight budgets, premium subscription fees are difficult to justify for marginal performance gains. Then there is sovereignty concern. Open-weight architectures give governments more control over where data is stored and processed. That matters in countries wary of routing sensitive information through U.S.-controlled cloud providers, particularly as debates over the reach of laws such as the U.S. CLOUD Act continue. This is already showing up in practice. Several countries are using Chinese technology to build local AI systems. Singapore launched SEA-LION, designed as a regional AI system for Southeast Asia. The project aims to give Southeast Asian countries a sovereign foundation model with better support for local languages and fine-tuning. Its best-performing model, SEA-LION v4.5, is built on Qwen. Malaysia is evaluating a $494 million sovereign AI initiative powered by Huawei’s Ascend 910C chips to gain more control over national data, despite U.S. warnings over semiconductor export controls. The trend now extends beyond Southeast Asia and into the U.S. backyard. Brazil has actually gone further: In August, its government announced a $251 million supercomputing project in Rio de Janeiro developed with Huawei and iFlytek, alongside a separate $193 million supercomputer tender, as it seeks to balance U.S. influence. Rio-3.5-Open-397B, released by Rio’s municipal IT company, is also based on Qwen 3.5. None of this requires governments or developers to “choose China over the United States” in a geopolitical or ideological sense. Even as Pax Silica attempts to force U.S. partners toward Washington’s preferred technology supply chains, developers can adopt Chinese models for entirely prosaic reasons: The price is right, flexibility is high, and they are simply the better choice for certain needs. And adoption is not binary. Developers in the United States, for example, already use combinations of closed and open-weight models depending on the workload. That is simply business logic. But once applications, data pipelines, hardware, and services grow around those models, switching becomes expensive. A practical choice today can become structural dependence tomorrow, and that has geopolitical implications. That is the part Washington should watch. The stickiest technological dependence is the kind users choose for themselves. If emerging markets shape their digital futures around Chinese open-weight models, China could do in AI what it has already done in EVs: gain lasting influence by supplying the bricks others build with.