China used the World AI Conference in Shanghai to turn open-source AI into foreign policy. If Washington keeps treating the race as a chip fight alone, it will miss where developers are actually choosing sides.
The clearest sign of China's AI strategy came on July 17, when Xi Jinping told the 2026 World AI Conference that countries should "seize this rare, historic opportunity" and encourage open source, openness, collaboration and sharing. That line matters because Beijing paired it with machinery. A day earlier, China's foreign ministry said representatives of 29 countries signed the agreement creating the World Artificial Intelligence Cooperation Organization, an intergovernmental body headquartered in Shanghai.
This is not a technical standards club. It is infrastructure diplomacy.
The founding members included China, Brazil, Indonesia, Malaysia, Pakistan, Russia, Senegal and South Africa, according to Indonesia's communications ministry and China's foreign ministry. You can read that list as geopolitics, but developers will feel it first as access: models they can download, tune, translate and run without asking OpenAI, Anthropic or a U.S. cloud provider for permission.
The strategy did not begin at WAIC. It became visible in January 2025, when DeepSeek's R1 release rattled the American AI market by showing that a Chinese lab could get close enough to the frontier while spending far less on training and inference. The U.S.-China Economic and Security Review Commission sharpened the point in its March 2026 staff paper, "Two Loops: How China's Open AI Strategy Reinforces Its Industrial Dominance." The commission described one loop of open-model innovation and another loop of deployment across manufacturing, logistics and robotics.
That is the part Washington keeps underestimating. Export controls can limit access to Nvidia's most advanced chips. They cannot stop a Qwen model from being downloaded in Nairobi, Jakarta or Sao Paulo.
The software layer is moving faster
The numbers around Alibaba's Qwen are difficult to wave away. Alibaba said Qwen passed one billion cumulative Hugging Face downloads as of January 21, 2026. South China Morning Post later reported, citing Interconnects AI, that Qwen captured more than 50 percent of global open-source model downloads as of March and generated 153.6 million downloads in February alone. That is distribution at a scale most closed-model companies can only buy through enterprise sales teams.
Alibaba has also made the ladder easy to climb. The Qwen3 release included dense models from 0.6 billion to 32 billion parameters and two mixture-of-experts models, all open-weighted under Apache 2.0, according to Qwen's own release notes. Qualcomm's AI Hub lists Qwen3-0.6B as supported for phone and tablet form factors. That is not a frontier model. It does not have to be. A model that runs locally and speaks enough languages can be more useful to a small team than a better model trapped behind a dollar-priced API.
DeepSeek is playing the same access game through pricing. Its current official API table lists DeepSeek V4 Flash at $0.14 per million uncached input tokens and $0.28 per million output tokens, with cached input falling to $0.0028. V4 Pro is higher, but still cheap by Western frontier standards. If you are building in a market where cloud bills are the difference between shipping and stopping, that price gap is not theoretical.
The Global South angle is already showing up in real projects. Foreign Policy reported in June that Ernest Mwebaze built Sunflower LLM, a model covering 31 Ugandan languages, on Alibaba's Qwen 3 rather than Google, Microsoft or OpenAI. That is exactly how defaults get made. One researcher chooses a base model because it is available and good enough, then students, startups and government projects around him inherit the stack.
Closed models have a distribution problem
Xi also promised 5,000 AI training and seminar opportunities for developing countries over five years, along with cooperation centers involving ASEAN, the African Union and BRICS. Don't confuse that with charity. Training developers on your stack is how you turn software into dependence. The United States did this with cloud infrastructure in the 2010s, and Amazon, Microsoft and Google are still collecting the returns.
Here's the thing: U.S. AI companies are not wrong to sell closed models to rich enterprise customers. OpenAI and Anthropic have built real products, real revenue and strong safety arguments around controlled access. But a government ministry in Indonesia or a startup in Uganda doesn't reach for a Washington policy memo on day one. It asks what runs locally, what the licence allows, and whether the cloud bill is survivable.
RAND's March 2026 report, originally published through CSET, recommended stronger U.S. support for open models, a recalibration of export controls and incentives for permissive licensing. That is a practical agenda. It is also moving slowly. Meanwhile, AP reported this week that Xi used WAIC to announce 5,000 training spots and cooperation with developing regions while criticizing restrictions on technology sharing.
China is not winning the AI race on raw capability. NIST's CAISI evaluation in May found DeepSeek V4 Pro still lagged leading U.S. models by about eight months. Fine. Capability matters. But if you are only counting benchmark distance, you are watching the wrong scoreboard.
Distribution has its own gravity. A model that thousands of developers can download, alter and deploy starts shaping habits before it beats every closed system. Beijing understands that. Washington had better understand it too.
Also read: Elon Musk wants rival AI labs to vet each other's models before release • The five biggest cloud builders are spending every dollar they earn on AI and then some • New York becomes the first US state to ban new hyperscale data centers and other states are watching
