Key Takeaways:
Chinese open-source AI models now account for 41% of downloads on Hugging Face, surpassing US models and reshaping the economics of enterprise AI.
Chinese open-weight AI models captured 41% of downloads on Hugging Face this spring, overtaking US models, as enterprises shift toward cheaper, customizable alternatives to frontier systems from OpenAI and Anthropic.
"If you're an AI company or a technology company, you don't want to outsource your core capabilities to another company, to a black box API that you don't control, don't have any visibility on, and don't really have any sort of ownership," Clem Delangue, CEO of Hugging Face, said.
On OpenRouter, the top six most popular models all come from Chinese firms including Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai. Anthropic's Claude Opus 4.7 trails in seventh place. Open models handled nearly a third of AI requests on Vercel's platform in June, while closed models retreated to a premium tier for high-value tasks.
The shift threatens the pricing power of US frontier labs that have poured billions into proprietary AI. With inference costs ranging from $5 to $50 per million output tokens across Anthropic's model tiers, enterprises are routing simple tasks to cheaper open alternatives and reserving premium models for complex reasoning. About 60% of enterprises have already imposed some form of AI spending limits, according to a UBS report published June 23.
The Cost Calculus Behind Model Routing
The price gap is the primary driver. Anthropic's Haiku 4.5 costs $5 per million output tokens, while Opus 4.5-4.8 runs $25 and the highest-end model reaches $50 — a 10x spread from low to high. Enterprises are adopting "model routing" as a standard cost-control technique, assigning different tasks to different models based on complexity.
Major cloud providers are responding. AWS Bedrock now lists MiniMax, Kimi, Qwen, DeepSeek and GLM among its available models. Microsoft Azure AI Foundry offers DeepSeek access. One large global bank has deployed Alibaba's Qwen locally to balance costs from Claude and other premium models, according to the UBS report.
Hugging Face hosts nearly 3 million public models and 1 million public datasets, with a new repository created every seven seconds. Half of all Fortune 500 companies use the platform to deploy private or open-source models, Delangue said.
The Open vs. Closed Debate Intensifies
Microsoft CEO Satya Nadella recently warned against single-provider lock-in, arguing that control of data should be a primary concern. "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself," Nadella said.
Anthropic CEO Dario Amodei has argued that releasing powerful open-weight models carries risks because they become difficult to control once distributed. Delangue counters that restricting access concentrates power in a few companies while reducing transparency.
"The biggest risk in AI is concentration of power," Delangue said. "The way you make the world safer, in my opinion, is by leveling up the playing fields and creating transparency on these models."
Beijing-based Z.ai recently released GLM-5.2, an open-weight model that competes with Anthropic's latest on agentic coding and security vulnerability identification. The model can be downloaded and run on local hardware without the content moderation guardrails that frontier labs impose.
For investors, the trend raises questions about the long-term pricing power of US AI leaders. If open models continue improving while remaining cheaper and more customizable, the premium that frontier labs can charge for general-purpose intelligence may compress. Companies like Anthropic and OpenAI may need to differentiate on specialized capabilities, enterprise service levels, or proprietary data access rather than raw model performance.
This article is for informational purposes only and does not constitute investment advice.
