China won open source while you were sleeping
Eighty percent of developers worldwide who use open-source AI tools are building with Chinese models. Alibaba's Qwen family has crossed nearly one billion cumulative downloads. US policymakers spent two crucial years treating open-source AI as a national security threat, and China filled the vacuum with the most aggressive distribution play the tech industry has seen since Android. This isn't a story about who built the better model. It's a story about who understood distribution.
The numbers that should worry Washington
In January 2026, Alibaba's Qwen surpassed 700 million downloads on Hugging Face, making it the most widely adopted open-source AI system on the planet. By March 2026, after the release of Qwen 3.5, that number blew past the 50% mark of all global open-source model downloads. For context, Meta's Llama, once the undisputed leader, has fallen to second place in cumulative downloads. A joint study by MIT and Hugging Face found that Chinese-developed open models accounted for about 17% of total downloads through August 2025, narrowly surpassing the US share of 15.8%, the first time China had edged ahead on that measure. But the gap has only widened since. Between February 2025 and February 2026, Chinese models captured 41% of downloads compared to 36.5% for US models. Some estimates suggest around 80% of US AI startups now use Chinese open-source AI models. And it's not just Qwen. DeepSeek's R1 reasoning model briefly overtook ChatGPT as the most downloaded app on the US App Store when it launched in January 2025. Companies like MiniMax, Moonshot AI, Z.ai, and Tencent have all released competitive open models. The Chinese open-source AI ecosystem has gone from a single standout to a deep bench.
How the US handed China the advantage
The irony is almost too neat. US export controls on advanced AI chips, first imposed in October 2022 and expanded in 2023 and 2024, were designed to slow China's AI progress. In practice, they accelerated exactly the kind of innovation the US should have feared most. Cut off from the latest Nvidia hardware, Chinese labs were forced to optimize. DeepSeek demonstrated that you could train competitive models with far fewer chips than anyone thought possible. The constraint didn't kill Chinese AI; it made it leaner and more efficient. Chinese developers circumvented hardware restrictions by training models on chips located in data centers across Southeast Asia, renting capacity rather than owning it. Meanwhile, US policymakers spent those same years debating whether open-source AI itself was dangerous. The National Telecommunications and Information Administration published reports urging caution on open-weight models. Regulatory uncertainty made American companies hesitant to go fully open. Meta released Llama with restrictive licenses. OpenAI and Anthropic kept their best models proprietary. As a Wall Street Journal opinion piece by Jai Ramaswamy and Matt Perault put it: "This didn't happen because China out-engineered the U.S. It happened because U.S. policymakers spent two crucial years treating open-source AI as a threat." China read the room differently. Open source wasn't a risk to manage. It was a distribution strategy to execute.
Distribution beats product, again
This is a pattern we've seen before. Google gave away Android for free and won mobile. The product wasn't necessarily better than iOS, but it was more available, more customizable, and it met developers where they were. China is running the same playbook with AI. Qwen isn't necessarily the best model in the world. On many benchmarks, frontier proprietary models from Anthropic and OpenAI still lead. But Qwen is free, it supports 119 languages, it comes in sizes from 600 million to tens of billions of parameters, and it's available to anyone with an internet connection. For a developer in Jakarta, Lagos, or São Paulo who needs a capable model and doesn't have the budget for API calls to GPT-5, the choice is obvious. The Chinese government appears to be supporting this expansion deliberately. The cost to build and deploy these models in China may be as little as a third of US-based alternatives. Alibaba's chairman Joe Tsai has said openly that open-sourcing AI brings global benefits by lowering costs, making it a natural fit for talent-poor and cash-poor countries. That framing is generous, but it's also strategic. Every developer who builds on Qwen is a developer who isn't building on Llama or Mistral. And once you build on a model, switching costs are real. You've fine-tuned it. You've built your pipeline around its tokenizer, its context window, its quirks. Developer lock-in through open source is one of the oldest plays in tech, and it works every time.
The governance question nobody wants to answer
Here's where the conversation gets uncomfortable. When the most-used AI models globally are trained under different values, governance frameworks, and regulatory regimes, what does that mean for the information ecosystem? Researchers at CSIS found that Chinese models like DeepSeek and Qwen exhibit measurable biases toward more hawkish, escalatory policy recommendations, particularly in scenarios involving Western democracies. China requires AI systems to pass an ideological test before public release. Training data must be filtered for political sensitivity, with companies barred from using any source unless 96% of its content is deemed "safe" by state standards. This doesn't mean every app built on Qwen will parrot Chinese state media. Open-weight models can be fine-tuned, and most developers strip out or override default behaviors. But the base layer matters. The assumptions baked into training data shape outputs in subtle ways, and most end users will never know which model powers the application they're using. The US advisory body that tracks AI competition warned in March 2026 that China's open-source dominance threatens US AI leadership. The report highlighted that as AI shifts from large language models toward agentic AI and embodied AI, China may be even better positioned to capitalize, given its mass data collection infrastructure and its lead in manufacturing hardware like humanoid robots.
What this looks like from a small state
For a country like Singapore, caught between two AI superpowers, the calculus is particularly fraught. Singapore has positioned itself as a neutral hub, a place where both American and Chinese AI companies can operate without triggering bilateral tensions. Its national AI strategy emphasizes governance, applied research, and being a staging post for real-world deployment rather than competing on raw compute. But neutrality gets harder when you have to pick an ecosystem. If your developers are building on Qwen because it's free and good enough, you're implicitly tilting toward the Chinese stack. If you mandate American models, you're paying more and limiting access. Singapore's prime minister Lawrence Wong acknowledged at the 2025 Summer Davos that the world is entering a transitional period without a clear new world order, a dangerous time for small states if hard power and the law of the jungle return. The strategic question for any small country isn't really which model is better. It's which dependency is more manageable.
What happens next
The US is not standing still. Nvidia launched the Nemotron Coalition to develop frontier-level open models with partners including Mistral, Perplexity, and Cursor. OpenAI has started releasing some open-weight models. The Stanford AI Index Report for 2026 notes that US and Chinese models have traded the lead multiple times since early 2025, with the performance gap narrowing to single-digit percentages. But catching up on distribution is harder than catching up on benchmarks. Qwen's ecosystem now includes over 170,000 derivative models. More than 90,000 enterprises deploy Qwen through Alibaba Cloud. That's an installed base, and installed bases have gravity. The lesson here isn't that open source is dangerous or that China cheated. Open source is, on net, a massive positive for developers everywhere. The lesson is that the US made a strategic miscalculation. It treated openness as a vulnerability while its competitor treated it as a weapon. By the time Washington figured out the difference, 80% of the world's open-source AI developers were already building with Chinese models. The race isn't over. But the starting positions have shifted in ways that will take years to undo.