Superintelligence for all

Agentic AI represents an opportunity for shared global development as it can become a tool for leapfrogging legacy digital divides rather than widening them

The 2026 World Artificial Intelligence Conference convenes in Shanghai on Friday under the theme “Intelligent Partners, Co-create the Future”. The central question facing the world is no longer whether AI will transform societies, but who gets to shape and benefit from that transformation.

The current trajectory of AI development, in which the frontier autonomous systems carry out long and complex tasks under or without human supervision, is unprecedented in its capacity to improve the very AI training process itself, triggering a cycle of recursive self-improvement and snowballing small initial advantages into decisive leads.

Concerns that this dynamic will entrench monopoly are real. The United States began restricting exports of advanced GPUs in 2019 and has since proposed controls that would limit AI chip exports to all Global South countries, even to ostensible allies. A further means of control has been for the US to keep its trained AI models closed and to not publish any technical details on how the models are produced or how they operate. Most recently, it has restricted access to Anthropic’s frontier Mythos model citing cybersecurity reasons.

Due to the US’ unipolar approach restricting access to compute, shunning open source and even restricting the ability to run frontier models, most countries in the world face the risk of being pushed into a permanent “consumer” role. In this role, countries such as India and the Philippines risk being forced to simply buy access to Anthropic and OpenAI’s models without developing any domestic AI skills or tech stack of their own. They would also permanently give up control over their own data, weakening their economic position and introducing national security and privacy threats.

Yet this outcome is not inevitable. If paired with open-weight model sharing, capacity-building partnerships and locally hosted inference infrastructure, agentic AI can become a genuine win-win: a tool for leapfrogging legacy digital divides rather than widening them.

The attempted US tech blockade has not produced the intended isolation. In response, China has embraced an ambitious research program, which aims to not just develop GPUs, but respond with parallel innovation — Huawei’s Ascend series, multi-patterning techniques — to extend pre-EUV lithography and “Tau-scaling”, a 3D chip-layout optimization approach. Industry observers note Huawei plans to offer Ascend GPUs in Asia-Pacific markets such as the Republic of Korea in Q4 2026, signaling growing global competitiveness. Apple has begun to lobby the US government for permission to buy memory chips from China.

More significantly, top Chinese labs have released full model weights, technical papers and code, making Chinese open-weight models the most widely adopted foundation models in global research labs. Industry surveys suggested that around 80 percent of US AI startups either experiment with or deploy Chinese open-weight models for inference and fine-tuning.

This openness, and not the mere narrowing of a performance gap, is the distinguishing feature of a multilateral approach to AI governance. It creates the possibility of a shared AI commons, in line with the Shanghai Declaration on Global AI Governance issued in 2024, which commits signatories to ensure equal rights, equal opportunities and equal rules for all countries in developing and using AI technologies. As a supplement to the declaration and practical guide to action, China’s Global AI Governance Action Plan adopted at the 2025 WAIC directs its commitment to strengthening South-South cooperation, enhancing the representation and voice of developing countries. Article 5 of the plan explicitly calls for open-source platforms, basic resource sharing, and lowering barriers to innovation and application.

Concrete mechanisms are already emerging. Launched in September 2025 at the China-Association of Southeast Asian Nations AI ministerial roundtable in Nanning, the Guangxi Zhuang autonomous region, the China-ASEAN Countries AI Application Cooperation Center is designed to strengthen the foundation for AI development, provide open-source services, facilitate industrial cooperation and promote talent cultivation across member states. Since 2019, the WAIC side-events have trained hundreds of officials and engineers from Africa, Southeast Asia, Central Asia and Latin America in large language model deployment, data governance and responsible AI use. The hope for 2026 expansion is high.

For Global South partners, the pathway to AI sovereignty follows a natural progression: first, host open-weight inference on local or regionally-shared hardware; next, fine-tune and adapt models to national needs, languages and regulatory contexts; and finally, as domestic datasets and compute capacity mature, pre-train original models from scratch. This is not a theoretical ladder. Tsinghua University has already begun training students from Southeast Asia and West Asia, who return to work at Z.AI offices in their home countries — bringing hands-on experience with open-weight systems that can be modified to reflect local values and priorities. Because the models and research are open, each country retains the freedom to customize, adapt and eventually build its own stack.

Crucially, this ecosystem is built on comparative advantage rather than unilateral monopoly. The Global South contributes the majority of the world’s STEM talent and nearly half of global GDP. India offers an exceptionally large pool of young AI researchers and software engineers. Vietnam’s rapidly expanding manufacturing sector provides a natural test bed for embodied AI. The Middle East states possess unrivaled energy and land resources for sustainable compute. When these strengths are combined with open models and targeted capacity-building — such as the joint laboratories and training programs outlined in the Action Plan — the result is a genuinely reciprocal partnership, not a donor-client relationship. The goal is not for every country to build the largest model, but for every country to build the model that fits its own needs, talents and future.

None of this negates the need for vigilance on safety, bias or misuse. But it reframes the question: the risk is not that agentic AI will inevitably concentrate power, but that exclusive, closed ecosystems could make it do so. The alternative — codified in the Shanghai commitments — is an inclusive architecture where frontier labs, Global South governments, and open-source communities co-develop norms and share benefits.

As delegates gather in Shanghai, the message from Global South countries is unambiguous: They do not seek to be permanent consumers of black-box intelligence. Through open models, shared datasets, joint laboratories and transparent governance, agentic AI can be domesticated to local needs and values.

China’s experience offers a model for a common path: AI must be a public good of a shared future for humanity. Whether that aspiration is realized depends on choices made now — by governments, companies, and research communities — to treat AI not as a zero-sum asset but as a cooperative infrastructure. In the age of agentic AI, the dividing line will not be between those who build the smartest agents and those who use them, but between those who lock intelligence behind walls and those who build the bridges that let it circulate. The latter is the only path consistent with a multipolar, equitable digital future.

The author is an assistant professor at the College of AI at Tsinghua University.

The author contributed this article to China Watch, a think tank powered by China Daily. The views do not necessarily reflect those of China Daily.

Contact the editor at editor@chinawatch.cn.