For weeks this summer the industry focused on new frontier models and on policy fights over who should get access. Meanwhile developers kept building and deploying alternatives without waiting for permission from big providers such as Anthropic or OpenAI.
Signals in usage and downloads
- According to Hugging Face, Chinese open-weight models accounted for 41% of downloads on the platform this spring, surpassing U.S. models.
- On OpenRouter, the six most popular models at the time of reporting were all open models from Chinese firms, including Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai; Anthropic’s Claude Opus 4.7 ranked seventh.
- Vercel’s data shows open-weight models are taking a large share of infrastructure-heavy AI traffic: on that platform open models handled nearly one third of AI requests in June, while closed models function more as a higher-cost, premium layer.
These platforms do not capture the entire ecosystem — notably sessions hosted directly by major labs such as OpenAI or Anthropic likely account for a substantial share of usage — but the growing market share of open-source models raises a question: if most production AI runs on cheaper, customizable alternatives, how central will frontier models remain?
Firms push to own and customize models
Clem Delangue, CEO of Hugging Face, says customers increasingly prefer owning their models instead of renting access to a black-box API they do not control. The costs associated with scaling closed frontier models have made that preference more apparent.
Delangue notes that a new repository is created on Hugging Face every seven seconds; the platform hosts nearly three million public models and one million public datasets. He argues this reality contradicts the idea of a single, dominant model: in practice companies will use many specialized or customized models. He also says half of all Fortune 500 firms use Hugging Face to deploy private and open models.
Chinese labs and the economics of open weights
Every few months another Chinese AI company releases a powerful open-weight model that is less expensive to deploy and easier to customize than closed competitors, shifting the economics of proprietary AI. Most recently, Beijing-based Z.ai released an open-weight model called GLM-5.2 that performs strongly on agentic coding and competes with Anthropic’s models on identifying security vulnerabilities.
Safety, transparency and concentration of power
The rise of open models has intensified debate about whether increasingly capable models should be widely available. Anthropic CEO Dario Amodei has warned that scaling powerful open model weights could be dangerous because once released they are hard to control. Others have noted that open models may be easier for bad actors to access for disinformation, cyber or biological misuse.
Delangue frames the trade-off differently: he sees the largest risk as concentration of power. In his view, openness and transparency make the ecosystem safer because defenders can better patch security risks that are visible in open models. He also argues that keeping models closed does not eliminate risk: frontier-model API guardrails can be bypassed and model weights can be stolen and distributed publicly, concentrating capabilities and reducing transparency.
Views on lock-in and data control
Microsoft CEO Satya Nadella has also warned against single-provider lock-in and argued enterprises should prioritize control over their data. Nadella said it is ironic that model providers benefit from training on public data, yet often impose restrictive terms on distillation and reserve rights to learn from customer interactions — a one-way flow that funnels economic value to infrastructure owners rather than to the creators of knowledge.
What to watch next
Current trends suggest a potential division of labor in the AI market: closed frontier models may remain tools for experimentation and highly specialized, high-value tasks, while mass production workloads migrate to cheaper, customizable open models. That shift has technological, economic and regulatory implications: who controls learning infrastructure, who owns data, and what safeguards are needed to prevent misuse will shape the next phase of the market.
In the coming months the pace of Chinese open-model releases, enterprise decisions to run their own models, and policy or industry responses will determine how pronounced this reorientation becomes.
