Western AI companies are increasingly dependent on Chinese open-weight models as training bases, creating a structural advantage for Chinese labs that could reshape the entire AI industry. According to analysis by Sequoia Capital, Qwen's share of open-model fine-tunes jumped from 1% to 69% between 2024 and 2026, with many American firms leveraging Chinese model weights to build their own products. This trend reveals a paradox at the heart of American AI policy: while the U.S. government considers restricting Chinese models, Silicon Valley's most innovative companies are building on top of them.

Why Are American Companies Using Chinese Models?

The answer lies in a legal loophole and a practical constraint. American frontier models like GPT and Claude are subject to strict usage restrictions that prohibit using their outputs to train competing systems. Chinese models, by contrast, are released as open weights, meaning anyone can download them, modify them, and build on top of them indefinitely. This creates a legal path for Western companies to learn from cutting-edge AI systems without violating intellectual property agreements.

A concrete example illustrates the advantage. Thinking Machines, a startup building the open-weight model Inkling, bootstrapped its fine-tuning process using synthetic data generated by Kimi K2.5, a Chinese model from Moonshot AI. The company had a legal path to learn from a Chinese open model, while equivalent use of GPT or Claude outputs would have been prohibited. This is not a fringe practice; it reflects how the economics of AI development have shifted.

Bridgewater Associates, one of the world's largest investment firms, demonstrated the practical power of this approach. Its AIA Labs team fine-tuned the open-weight Qwen3-235B model on financial documents labeled by Bridgewater experts. The custom model reached 84.7% accuracy on six financial tasks, outperforming the best frontier models Bridgewater tested, which reached only 78.2% accuracy. Crucially, the custom model cost 13.8 times less per task. The advantage came not from Qwen being smarter than GPT or Claude, but from Bridgewater's ability to teach the model what its own investors considered useful.

What Does This Mean for the Future of AI Competition?

The structural advantage is significant. Each time Qwen, Kimi, or DeepSeek releases a new model, Western companies can download it and use it as a foundation for their own work. But if China stops releasing its best models, firms that depend on them risk falling behind. This creates a dependency that could shift competitive advantage eastward, even as American companies invest billions in their own frontier models.

The White House has taken notice. The Trump administration is reportedly considering procurement bans, Entity List designations, and new liability for American companies that host Chinese models. These restrictions would aim to protect American AI companies from competition, but they would also eliminate the legal foundation that allows startups and enterprises to build specialized AI systems at a fraction of the cost of proprietary APIs.

How Companies Are Building Custom AI Systems on Open Models

• Fine-tuning on proprietary data: Companies download open-weight models and train them on their own labeled data, creating specialized systems that outperform general-purpose frontier models on specific tasks.
• Synthetic data generation: Startups use Chinese open models to generate training data for their own systems, creating a legal pathway to learn from frontier-class models without violating usage restrictions.
• Cost optimization: Custom models built on open weights can cost 10 to 15 times less per inference than equivalent proprietary APIs, making AI economically viable for use cases that would otherwise be too expensive.

The business model emerging around this practice is fundamentally different from the token-maximization approach of OpenAI and Anthropic. Instead of charging the highest possible price for every word processed, companies like Microsoft, Palantir, and ServiceNow are making money from training tools, infrastructure, customization, and deployment. Microsoft Foundry lets companies train and deploy open and custom models on Azure. Palantir AIP lets customers run open models on their own infrastructure and connect them to company data and workflows. These companies do not need one model to win; they make money from the layer around the model.

This explains why Nvidia, Microsoft, Meta, Palantir, ServiceNow, Box, Hugging Face, Y Combinator, and dozens of other companies signed a joint letter defending open-weight models. Nvidia sells the chips. Microsoft sells the cloud. Palantir, ServiceNow, and Box sell the software and data layer. Startups get a cheaper base on which to build. The coalition is not defending Chinese models out of altruism; they are defending a business model that depends on open weights remaining available.

If Washington restricts Chinese models, it would create an obvious market for American open models to fill the gap. Many more startups would be incentivized to build smaller, hyper-specialized open models backed by the same companies that signed the letter. Most companies do not need one model with general intelligence for every task. The legal team needs a model that understands contracts and company policy. Finance needs one that understands internal reporting. Operations needs one that knows the supply chain. These models do not need to beat the latest frontier model on every benchmark; they need to perform one job reliably, protect company data, and cost less.

The irony is that a ban on Chinese open models might not benefit OpenAI or Anthropic at all. It might instead accelerate the shift toward a fragmented AI ecosystem where dozens of specialized models replace the dream of a single general-purpose system. That shift is already underway, and it is being powered by the very Chinese models that Washington is considering restricting.