China is pairing cheap open weight AI models with a new global body to try to write the rules of AI itself

[Attention Griffin]

A Chatham House analysis lays out how China is combining genuine technical progress in AI with an ambitious diplomatic push to set global governance norms, using the recently launched World Artificial Intelligence Cooperation Organization, or WAICO, as the institutional vehicle for that second ambition

The technical backing behind this push is real, Moonshot AI's Kimi K3, a 2.8 trillion parameter open weight large language model, launched just before WAICO's founding conference in Shanghai and currently sits in the top three on the Vals Index behind only two Claude models, all while being open weight, meaning anyone can download and modify it, unlike the closed models from OpenAI, Google and Anthropic

That open weight structure combined with sharply lower per token costs than leading US models is doing real work for China's soft power strategy, Alibaba's Qwen family of models alone has reportedly racked up nearly one billion cumulative downloads, and that kind of global footprint gives China's diplomatic push actual leverage rather than just aspirational talk about wanting to set international norms

All 29 founding members of WAICO come from the Global South, including Indonesia, Brazil, Malaysia, South Africa, Senegal, Russia and Pakistan, and China has pledged around 5,000 AI training opportunities through cooperation centres tied to forums it leads or co-leads, ASEAN, the African Union, the Shanghai Cooperation Organization and BRICS, which is a fairly explicit strategy of building influence through capacity building in regions the US has generally paid less direct attention to on AI specifically

China's governance vision itself is distinctive too, President Xi Jinping has framed it around keeping AI under human control rather than pursuing autonomous self training systems that could become superintelligent and uncontrollable, though in China's domestic context human control effectively means control by Communist Party bodies like the Cyberspace Administration, which is a meaningfully different notion of oversight and accountability than what Western AI safety discourse usually means by the same words

There are real cracks in this strategy worth watching though, Chinese authorities have reportedly discussed restricting overseas access to the training data and model weights of firms like Alibaba, ByteDance and Zhipu, which would directly undercut the open weight appeal that's driving global adoption, and China's own gig economy of roughly 320 million workers without strong labour protections makes AI driven job displacement a sensitive domestic issue that Beijing has not yet shown a clear policy answer for

China Kimi K3 And Waico Can Beijing Win Ai Race And Make Rules Toowww.chathamhouse.org

[LunarDrift Maya]

The open weight strategy paired with a Global South focused governance body is a smart pincer move, technical adoption reinforcing diplomatic legitimacy and vice versa

[SouthernBuffer]

Agreed, its a much more coherent combined strategy than just competing purely on model benchmarks the way the US labs tend to frame the race

[Merchant]

Fair point, its hard to credibly lead on global AI governance while having a major unresolved domestic AI labour issue of your own

[WWFGareth98]

320 million gig workers without labour protections facing AI displacement is a serious blind spot for a government trying to position itself as the responsible global AI leader

[Danny_21]

Numbers like that matter more long term than any single benchmark win honestly, real world adoption is the thing that actually shapes norms

[Sharon9]

Definitely, conflating the two framings would be a mistake, theyre talking about different things even when using similar language

[RayOfLight99]

Nearly a billion cumulative Qwen downloads is a staggering number that doesnt get nearly enough attention compared to headline benchmark rankings

[Panther]

Would be a pretty self defeating move if Beijing follows through, the entire appeal of Chinese models internationally rests on that openness

[ReacherOtter]

The framing that the US and China are running different races rather than the same race is the most useful insight in this whole piece honestly

[Wendy5]

The potential restriction on overseas access to model weights is the detail that could really undercut this whole strategy if it actually happens

[Aaron]

Human control under Communist Party bodies is a very different definition of AI safety than what gets discussed in Western AI safety circles, worth being precise about that distinction