AI & TECHNOLOGY — Global Strategy

China's Open-Weight AI Strategy Is Winning — While the West Looks Away

For two years the AI race was framed as a capability gap: American frontier labs led on benchmarks, Chinese models trailed by six to eighteen months. A different story has since emerged. China is now shipping frontier-class models as downloadable open weights, pricing intelligence at a fraction of what the US giants charge, and the rest of the world is quietly adopting them.

• DeepSeek, Alibaba's Qwen, and Moonshot's Kimi have moved from lagging followers to credible challengers on real developer benchmarks for coding, reasoning, and agentic tasks.
• Instead of hoarding the most powerful systems behind paid APIs, Chinese labs publish their model weights publicly, letting third parties fine-tune, deploy on private hardware, and embed the models in products worldwide.
• Analysts at BCG and Brookings describe an emerging split: the US holds the frontier on raw capability, while China dominates the cost-optimised, deployable-model layer that most organisations actually use.

The shift is structural rather than technical. Big-tech AI has spent its first decade selling intelligence as a service: access a model through an API, pay per token, never touch the weights. That model works for flagship research but creates dependency. A Chinese developer, a European hospital, or an Indian startup that builds on an open-weight model owns its own stack — the model can run on local hardware, be fine-tuned on private data, and operate without a US cloud provider, a terms-of-service change, or an export-control restriction. Sovereignty has become a product feature.

The economics reinforce the strategic logic. Training frontier models now costs hundreds of millions of dollars and consumes specialised silicon that is itself subject to export controls. Chinese labs work around those constraints with more efficient architectures, larger scale where they can obtain chips, and rapid open release. The result is a model family that benchmarks close to the leading US systems on the tasks that matter in production, at a per-query cost that is often an order of magnitude lower.

Most of the global developer community has responded with its feet. Adoption surveys across Europe and beyond show Chinese open-weight models — DeepSeek in particular, followed by the Qwen series — as the most-deployed open models in production environments. The pattern mirrors software history: free and low-cost eventually captures the broad market, while premium closed products are squeezed into narrower niches. PCs overtook minicomputers; Linux overtook proprietary UNIX; and now open-weight intelligence is doing the same to the API economy.

For Western labs the implication is uncomfortable. A capability lead on paper means little if developers choose cheaper, deployable alternatives for their actual workloads. The race is no longer purely about who has the biggest model — it is about who controls the stack that runs inside companies, hospitals, phones, and satellites. China's strategy has been to give that stack away.

Knowledge takeaway: Chinese AI labs are releasing frontier-class models as open weights rather than closed APIs, letting the global developer community deploy them on private hardware; analysts now describe a split where the US leads on raw capability while China dominates cost-optimised deployable models; the economics of free, downloadable intelligence are reshaping the AI market the same way Linux reshaped operating systems.