Beijing’s newly approved 15th Five-Year Plan landed with very little fanfare in the Western press. It should not have. In the 141-page planning document submitted to the National People’s Congress in March, the word for lithography machine does not appear once. Neither does wafer fab, extreme ultraviolet, or chip manufacturing. The entire Western vocabulary of the semiconductor war is absent. What took its place is a strategy built on something the West did not expect China to choose: openness.
The dominant assumption in Western AI circles has always been that the country with the biggest models wins. More parameters, more compute, more proprietary control. OpenAI, Google, Anthropic — the model is the moat. China looked at that framework and walked in the opposite direction.
Since DeepSeek released its R1 reasoning model in January 2025, Chinese companies have repeatedly delivered AI models that match the performance of leading Western models at a fraction of the cost. DeepSeek did not just match competitors. It exposed how much of Western AI spending was structural waste. R1 was trained for approximately six million dollars, compared to roughly one hundred million for GPT-4, using an architecture that activated only 37 billion of its 671 billion parameters per inference. The model was then released under an open MIT licence, free for anyone to download, modify, and deploy.
That decision turned out to be a weapon.
According to a study of 100 trillion tokens processed through OpenRouter and Andreessen Horowitz, Chinese open-source models went from 1.2 percent of global AI usage in late 2024 to nearly 30 percent within months. Alibaba’s Qwen followed the same logic. The Qwen model family had surpassed 700 million downloads by January 2026 on Hugging Face, making it the world’s most widely used open-source AI system. These are not niche adoption figures. One partner at Andreessen Horowitz estimated that 80 percent of US startups use Chinese base models to develop derivatives for their businesses. American companies, built on American capital, running their core AI workloads on Chinese infrastructure. The irony is structural and deliberate.
Why does this work? Because openness compounds. Frontier labs refine each other’s base models, enterprises adapt them for niche applications, and deployment data feeds back into capability improvements. Proprietary models cannot participate in this loop. Every time OpenAI improves GPT-5, that improvement stays inside OpenAI. Every time DeepSeek improves R1, the entire global developer ecosystem absorbs it, builds on it, and sends improvements back. The gap between these two approaches grows with each iteration.
The Five-Year Plan encodes this logic formally. A new planning term appears in the document for the first time in Five-Year Plan history: 模芯云用, meaning “model-chip-cloud-application,” a four-character compound that places the model layer as the primary strategic layer. Chip production remains a goal. But Beijing is no longer measuring success by semiconductors manufactured. It is measuring success by how deeply computing infrastructure penetrates the economy, with a target of digital economy value-added at 12.5 percent of GDP by 2030.
There is a counter-argument worth taking seriously. Proprietary frontier models remain more capable at the absolute cutting edge. GPT-5 can do things that Qwen cannot. The West still leads on the hardest benchmarks. But that argument misunderstands where the real competition is happening. China is not competing for the top of the leaderboard.
It is competing to become the default AI infrastructure layer for the developing world, where cost sensitivity and deployment flexibility matter far more than marginal benchmark performance. Singapore, Indonesia, Malaysia — the adoption numbers from Southeast Asia alone suggest this strategy is working faster than most Western analysts predicted.
The West built a race around the question of who can train the biggest model. China quietly started running a different race entirely. And by the time most observers noticed, Chinese models were processing 5.16 trillion tokens per week on OpenRouter compared to 2.7 trillion for US models, with four of the five most-used models globally being Chinese.
That is not a story about catching up.
