U.S. Companies Turn to Chinese AI Models as Costs Rise and Security Tensions Mount

U.S. firms are increasingly routing workloads to Chinese-built large-language models. DoorDash, Airbnb, Siemens and the San Francisco startup Lindy have moved significant portions of their generative-AI usage from OpenAI’s ChatGPT and Anthropic’s Claude to models such as DeepSeek’s V4, Moonshot AI’s offerings and Z.ai’s GLM-5.2. The shift reflects a growing preference for “open-weight” systems that can be customized and inspected, as well as a reaction to soaring token fees for U.S. frontier models.

U.S. Businesses Turn to Chinese AI Models as Cost Pressures Mount

Token pricing has become a decisive factor. One organization reportedly spent $500 million in a single month on Claude usage, while the Ramp AI Index notes that AI-heavy firms average roughly $7,500 per employee each month. Data from the OpenRouter platform shows Chinese models accounting for more than 30 % of weekly token usage by U.S. companies since early February, peaking at 46 %, compared with an 11 % average the prior year and 4.5 % in the first half of 2025. Analysts estimate that leading Chinese open-source models cost 60 %–90 % less than comparable OpenAI and Anthropic offerings.

U.S. Businesses Turn to Chinese AI Models as Cost Pressures Mount

Chinese models are marketed as “open-weight,” meaning their parameters are fully visible to users. This transparency lets enterprises tailor models to proprietary data pipelines and offers greater confidence that sensitive information is not being mishandled. The open-weight design also aligns with corporate cybersecurity strategies that favor in-house control over opaque, black-box services.

U.S. Businesses Turn to Chinese AI Models as Cost Pressures Mount

DoorDash co-founder Andy Fang announced on X that the company is saving “a lot of money” by assigning lower-level tasks to a Moonshot AI model. Lindy’s chief executive Flo Crivello reported that migrating from Claude to DeepSeek caused operating costs to “crash to the ground,” projecting multi-million-dollar savings within months. Georgetown University research fellow Sam Bresnick emphasized that enterprises have a clear incentive to shift workloads to cheaper, workable alternatives.

Futurism

U.S. officials have framed Chinese distillation—training a “student” model on the outputs of a more advanced system without permission—as a national-security threat. A White House Office of Science and Technology memo identified “foreign entities, principally based in China” as engaging in “industrial-scale” theft. Former White House CIO Theresa Payton described the practice as “the ultimate form of industrial homework theft.” House Select Committee Chairman John Moolenaar warned that China’s AI effort is “driven by theft, not innovation.” The administration has discussed the AI Overwatch Act, which would tighten export controls on advanced AI chips and require government permission for sales to “countries of concern.”

New York Post

Anthropic and OpenAI have lodged formal complaints with Congress, accusing Alibaba and other Chinese firms of illicit distillation that harvested millions of Claude and ChatGPT outputs via tens of thousands of fraudulent accounts. Anthropic’s letter to Senators Tim Scott and Elizabeth Warren called the attacks “illicit, systematic and industrial scale.” OpenAI has briefed the Congressional-Executive Commission on China, urging stronger policy tools to curb such activities.

New York Post

• “Enterprises are starting to realize, ‘Hey, we don’t need the best model, we can use the faster, cheaper models.'” — Eugene Cheah, CEO, Featherless AI

Futurism

• “Enterprises have an incentive to shift some of their workload to cheaper models. Why would you pay a premium for Anthropic, OpenAI models when for a lot of the workloads you need, the Chinese models are generally workable?” — Sam Bresnick, Research Fellow, Georgetown University’s Center for Security and Emerging Technology

Futurism

• “The Mythos ban was certainly the most tangible event, and people having their access revoked,” — Aidan Gomez, CEO, Cohere

Futurism

• “It allows foreign adversaries to clone the smartest kid in the class for pennies on the dollar,” — Theresa Payton, former White House CIO

New York Post

• “China’s effort to develop its own AI models is driven by theft, not innovation. This is a threat to our tech companies and our national security,” — John Moolenaar, Chairman, House Select Committee on China

New York Post

• “won’t matter as much for the next era of AI,” — Sara Hooker, CEO, Adaption

Seattle Times

Analysts differ on how far Chinese models trail U.S. frontier systems, with estimates ranging from six to nine months to an undefined “roughly six to nine months.” Legal scholars disagree on whether large-scale distillation violates the Defend Trade Secrets Act, and courts have yet to issue definitive rulings.

U.S. Businesses Turn to Chinese AI Models as Cost Pressures Mount

The White House’s Michael Kratsios signaled forthcoming actions to share intel on distillation attacks and improve coordination with private firms. Congressional proposals such as the AI Overwatch Act and related export-control legislation are slated for debate in the coming months, potentially reshaping the competitive landscape for AI model development and deployment.

New York Post