China's BRICS AI open-source community pitch

Xi Jinping closed the 18th BRICS summit in New Delhi by unveiling a five-point AI plan: China will "pioneer" a BRICS AI open-source community, back cooperation on large language models, hold joint AI seminars and training courses, and build out a BRICS digital ecosystem cloud platform for skills training and industrial alignment ^[1]. He paired it with a proposed BRICS engineer cultivation alliance for mutual recognition of competency standards, timed to a moment when China is about to take over the BRICS chairship and host the next summit ^[2].

The staging looked designed to make China's AI vision the story of the summit's close. But the joint New Delhi Declaration, a 140-paragraph document tied to the summit ^[3], never actually named it. It called only for generic "international cooperation in enhancing accessibility of AI resources while ensuring its safety, security, inclusiveness and reliability" ^[4]- diplomatic language that could describe almost any country's AI policy, and stopped well short of endorsing a China-led open-source community. The declaration did flag one specific AI risk: the misappropriation of knowledge, heritage, and cultural values in training data, a caveat aimed more broadly than at any single initiative ^[6]. Host Modi added his own dose of caution from the summit stage, warning that "the weaponisation of technology and critical minerals can hinder our shared progress" ^[5]- not a direct rebuke of Xi's plan, but a reminder that BRICS members are not uniformly comfortable ceding AI leadership to any one member, China included.

The declaration's silence is easier to understand alongside a contradiction inside China itself. Even as Xi pitches open-source AI as a global public good, China's Ministry of Commerce has reportedly been convening the makers of the country's leading open-weight models - Alibaba's Qwen, ByteDance's Doubao, and Z.ai's GLM-5.2 - to discuss a possible tiered regime that could restrict overseas access to the most capable versions, including bars on publicly releasing the most sensitive frontier weights ^[7]. That sits awkwardly next to the diplomatic framing of open-source as an act of generosity toward the Global South.

The contrast sharpens against the EU, which gained real enforcement teeth for its own AI Act on August 2, 2026, with fines of up to 3% of global turnover for violations - while still carving out an exemption for genuinely open models whose weights, parameters, and architecture are public and licensed for free use and modification ^[7]. Brussels has drawn a bright legal line around what counts as "open." Beijing has not yet drawn one at all, which means the same government promoting openness abroad retains full discretion to redefine or restrict it at home whenever it judges the models too strategically sensitive to release.

Community reaction converged on a blunter reading than Beijing's five-point framing suggests. On Reddit, discussion of the World AI Cooperation Organization pact repeatedly reframed "open source" as a matter of necessity rather than magnanimity: commenters argued that flooding the market with free, permissively licensed weights is essentially the only way Chinese labs can get global users onto their models while they still trail US frontier labs on raw capability - open access read as compelled, not generous. Others pushed back that the strategy is already working in practice, pointing to real-world adoption such as Siemens using Chinese open models for cost and customization reasons, and Russian consumer AI products built on top of Qwen and DeepSeek weights because US export controls cut off GPU-heavy alternatives.

That tension - compelled openness that still produces genuine adoption - maps onto China's own public justification. Beijing's stated argument is that blacklists, closed-source systems, and technology restrictions risk fragmenting global AI supply chains and widening the gap between countries that can build frontier models and those that can't ^[8]. Critics elsewhere online read the same fact pattern as a hedge against US chip controls: if Nvidia export restrictions make it harder for China to compete on raw compute, giving away model weights - and locking in the governance relationships that come with it - is a way to make those restrictions less decisive over time.

BRICS's non-committal declaration may matter less than it looks, because the capital and coalition-building are already moving through channels that don't require BRICS consensus at all. Commentary around the summit noted China has been pushing its AI agenda through the Asian Infrastructure Investment Bank's roughly $100 billion fund and through its own standalone World AI Conference (WAIC) in Shanghai, independent of whatever BRICS collectively agrees to - with one warning being that if the bloc doesn't move fast, the real deals get concluded outside BRICS altogether. The BRICS-linked New Development Bank has separately launched a $5 billion "digital sovereignty fund" earmarked for AI research, data centers, and chip manufacturing across member states - a concrete capital commitment that outsizes anything spelled out in the New Delhi Declaration's text.

Other members are hardly waiting on Beijing either: Russia is leading its own 10-nation AI alliance with a combined GDP near $28 trillion, and Xi's most quotable framing of the mission - encouraging "open source and open cooperation" to avoid "a world where only a few nations control the most powerful technology" - was delivered at Shanghai's WAIC, not at BRICS. The World AI Cooperation Organization, with its nearly 30 signatories and zero EU members ^[4], looks less like a BRICS initiative than a parallel structure that BRICS members happen to be joining alongside it.

Underneath the summit theatrics is a genuine framing contest over what AI governance should even optimize for. Fudan University's Zhang Weiwei has argued that global AI governance has "long been dominated by a small group of Western countries whose approaches conflict with one another," casting the US as treating AI as a strategic weapon in pursuit of absolute superiority while positioning China's approach as depoliticized and human-centered - AI as a public good rather than a competitive asset - and has proposed a "federal model" letting countries keep raw data domestic while models still cooperate across jurisdictions ^[9]. It is a framing built to make BRICS members feel like co-authors of the rules rather than adopters of China's technology.

Not everyone buys that BRICS can act as the unified rule-definer this framing requires. Coverage framing China as dueling with the US over AI vision also carried a more skeptical read from the International Crisis Group's Praveen Donthi, who noted that "the Global South is not a homogeneous entity" ^[10]- a reminder that a bloc spanning India (which just cautioned against weaponizing technology from the summit stage), Russia (running its own separate AI alliance), and other members with their own interests has no obvious mechanism for turning Xi's pitch into common policy. China gets to keep proposing the architecture; whether BRICS as a body ever ratifies it remains a separate and far less certain question.