The Shift No One Saw Coming
In the race for AI dominance, conventional wisdom held that proprietary models would win. Companies like OpenAI and Google invested billions into closed, locked-down systems, while China took a different path: open-weights AI. As of July 2026, the results are in. China's strategy is not just competitive—it's winning.
According to a recent analysis by Ben Werdmuller, published on his blog Source, the American approach of locking down AI models behind APIs and paywalls is causing the U.S. to lose ground. The article argues that by focusing on proprietary systems, American companies are missing the collaborative, rapid-iteration benefits that open-weights models provide.
The Open-Weights Advantage
Open-weights AI refers to models where the trained parameters (the 'weights') are publicly released, allowing anyone to download, fine-tune, and deploy the model on their own hardware. This is distinct from open-source models, which also release training code and data. Open-weights models like Meta's LLaMA series and China's Qwen and DeepSeek families have shown that releasing weights can accelerate adoption and improvement beyond what any single company can achieve.
Real-World Case: DeepSeek's Rise
A concrete example is DeepSeek, a Chinese AI lab. In early 2025, DeepSeek released its V3 model with open weights. Within months, the developer community had adapted it for languages, specialized tasks (like medical diagnosis and legal document analysis), and even hardware optimization for low-cost GPUs. By contrast, a comparable proprietary model from a U.S. company required API access, cost per token, and had usage restrictions. DeepSeek's model was downloaded over 10 million times in its first quarter, according to community estimates.
The result? Chinese open-weights models are now embedded in thousands of small businesses, academic projects, and government systems across Asia and Africa. While the U.S. argues about safety and regulation, China deploys.
The Cost of Lockdown
The article by Werdmuller makes a crucial point: proprietary models are losing the 'network effect.' When a model is open-weights, thousands of developers contribute improvements, find bugs, and create derivatives. This collective intelligence far surpasses the internal teams of even the largest AI companies.
For example, consider the cost of inference. A proprietary model like GPT-4o costs approximately $10 per million tokens for output. An open-weights equivalent like Qwen2.5-72B can run on a single A100 GPU, costing less than $0.50 per million tokens in electricity. For a startup processing 100 million tokens daily, the savings are $950 per day. Over a year, that's over $300,000—the difference between survival and bankruptcy for many small companies.
The Innovation Trap
Proprietary models also create a dependency on a single vendor. If the API goes down, or pricing changes, businesses are stuck. Open-weights models allow companies to switch infrastructure, add custom features, and even fork the model if the original developer abandons it. This flexibility is why many Asian and European governments are mandating open-weights for public sector use.
How China Executed the Strategy
China's success isn't accidental. The Chinese government has invested heavily in AI infrastructure, including national computing centers that provide subsidized GPU access for open-weights model training. Companies like Alibaba (with Qwen) and Baidu (with ERNIE) release their models under permissive licenses, often with no restrictions on commercial use.
The Ecosystem Effect
By releasing weights openly, Chinese AI labs have created a self-sustaining ecosystem. Developers build tools, write tutorials, and create datasets that feed back into the models. For instance, the open-weights community on Hugging Face has over 50,000 Chinese-language models derived from Qwen and DeepSeek. This is a flywheel: more users attract more contributors, which improves the models, which attracts more users.
What the West Is Doing Wrong
The United States, by contrast, has focused on safety through centralization. The Biden administration's 2023 executive order on AI emphasized testing and reporting for large models, but did little to encourage open distribution. Meanwhile, companies like OpenAI and Anthropic have pushed for licensing regimes that would effectively outlaw open-weights models.
This approach has two flaws. First, it assumes that safety can be enforced through gatekeeping, when in reality, open-weights models allow for independent auditing. Second, it cedes global influence to China. When a country in Southeast Asia or Africa adopts an open-weights model, it becomes part of that ecosystem. When it uses a proprietary API, it remains a customer, not a partner.
The Missing Piece: Practical Adoption
The article highlights that American AI companies are losing the 'battle for the real world.' While U.S. models excel at benchmarks, Chinese open-weights models are being used in agriculture (crop disease detection), manufacturing (quality control), and logistics (route optimization). These are not flashy demos—they are daily operations that generate data and feedback, making the models better over time.
Conclusion: The Path Forward
China's open-weights AI strategy is winning because it aligns incentives: developers get freedom, companies get cost savings, and the government gets widespread adoption. The United States can still compete, but it requires a shift in mindset. Instead of locking down models, American companies should embrace open weights, invest in community building, and focus on safety through transparency rather than restriction.
The news from July 2026 is clear: the open-weights approach is not just a philosophical choice—it's a strategic advantage. The question is not whether to open weights, but how quickly the West can catch up.
For businesses looking to integrate AI models into their operations, understanding the infrastructure that supports them is key. ASI Biont supports connection to Hugging Face and other model repositories through API — подробнее на asibiont.com/courses
Comments