Chinese AI models now handle most of the token usage generated by U.S. companies on OpenRouter, creating a new challenge for Washington.
According to Bloomberg, Chinese AI systems now account for roughly 60% of token usage by U.S. companies on OpenRouter, a popular marketplace where developers route their work across competing models.
GuruFocus estimated the figure was even higher, reaching 63% during the first week of July, compared with less than 10% a year earlier. Models like DeepSeek, Alibaba’s Qwen and Moonshot AI’s Kimi K3 have moved from curiosities to daily infrastructure for U.S. startups, researchers and larger enterprise users. DoorDash and Airbnb have both adopted Chinese models as cheaper alternatives to offerings from OpenAI and Anthropic, per Bloomberg.
Why Chinese AI models are gaining ground
Cost is the simplest explanation. AI-assistant startup Lindy AI said switching from Anthropic’s models to DeepSeek cut its inference costs by 90%, according to GuruFocus.
Airbnb CEO Brian Chesky previously said the company was relying heavily on Qwen, calling it “very good” and “fast and cheap,” per reporting cited by The Hill. Research from Andreessen Horowitz found 80% of developers worldwide using open-source tools are building with Chinese models, largely because of rising costs for private, closed systems.
For many companies, performance and cost remain the primary considerations when selecting AI models. But widespread adoption could create new questions about whether businesses should disclose which models power their products and how customer data is handled.
The implications extend to markets as well. Moonshot’s Kimi K3 release briefly rattled AI and chip stocks, echoing the market reaction that followed last year’s DeepSeek shock.
Washington weighs its options
The trend has caught the attention of top officials. In a report from Fox Business, Treasury Secretary Scott Bessent previously said that the administration could sanction foreign models found to have been built using stolen American technology.
“If we see, especially that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft,” Bessent said.
Meanwhile, U.S. Trade Representative Jamieson Greer said on CNBC that officials are “taking a very close look at how China is propagating its AI development.”
But the tools that worked against Chinese hardware don’t translate neatly to software. In a Bloomberg report, Kristy Loke, a fellow at MATS Research, said, “This is very different from the banning of Huawei 5G in 2018,” adding, “It’s a different world.” Once a model’s weights are downloaded, it can be copied, modified and run locally, with no port of entry to block.
Why Chinese AI models are difficult to restrict
For years, U.S. policy aimed at keeping Chinese developers reliant on American chips. The irony now is that American companies have built their own reliance on Chinese software.
Startups have spent months writing code, prompts and internal tools around specific Chinese models; ripping that out isn’t a policy switch, it’s an engineering project with a bill attached. That asymmetry — hardware restrictions hit a supply chain, while software restrictions may affect codebases already in production — is what makes this fight structurally different from the chip war.
The trade-offs ahead
None of this erases the security concerns. Anthropic has alleged that Alibaba conducted what it described as a large-scale model distillation effort, a claim that has become part of a broader debate over intellectual property and AI training practices.
Any crackdown, though, risks pushing developers toward harder-to-monitor channels rather than eliminating the practice, meaning Washington’s choice isn’t between risk and no risk, but between visible dependence and hidden dependence.
Also read: China launches international AI group as it pushes for a larger role in AI governance.
