US companies are increasingly adopting Chinese-developed AI models to reduce operational costs, a trend that accelerated since February 2026 with Congressional security investigations launched in April 2026. AI service provider Lindy switched all traffic from Anthropic's Claude to DeepSeek in June 2026, expecting millions in cost savings, while cryptocurrency exchange Coinbase announced in late June 2026 a switch to Zhipu AI's GLM-5.2 and Moonshot AI's Kimi K2.7 Code, cutting AI expenses nearly in half. The shift is driven by extreme cost differentials: running identical tests costs approximately $4,811 with Claude versus $1,071 with DeepSeek, $948 with Kimi, and $544 with GLM, according to Artificial Analysis benchmarks. However, the House Homeland Security Committee and House Select Committee on CCP launched a joint investigation in April 2026, citing national security risks after Anthropic reported in February 2026 that DeepSeek, Moonshot AI, and MiniMax used approximately 24,000 fake accounts for over 16 million interactions to improperly extract Claude's capabilities. This trend occurs amid tightening US export controls, with Zhipu AI placed on the Commerce Department's entity list since January 2025 and Anthropic's advanced models Mythos 5 and Fable 5 suspended from June 12 to July 1, 2026.
US Companies Switch to Chinese AI Models for Cost Savings
Developers and enterprises are directing more workloads toward open-source and open-weight models developed by Chinese companies including DeepSeek and Z.ai. Lindy CEO Flo Crivello stated that after AI usage costs exceeded personnel expenses, the company transferred all traffic from Anthropic's Claude model to DeepSeek in June 2026, expecting to save millions of dollars. Coinbase CEO Brian Armstrong announced in late June 2026 that the company replaced the standard model used by over 1,200 internal AI agents with Zhipu AI's GLM-5.2 and Moonshot AI's Kimi K2.7 Code, successfully cutting AI-related expenditures nearly in half. Ride-sharing platform Uber exhausted its 2026 AI budget by April 2026 and subsequently imposed a $1,500 monthly AI usage cap per engineer to control spending.
Chinese AI Model Usage Surges on OpenRouter Platform Since February 2026
Data from OpenRouter platform shows that the proportion of token usage (data units processed) by US companies for Chinese-made AI models increased significantly from approximately 4.5% in the first half of 2025. Since February 2026, this proportion has exceeded 30% every week, with peaks reaching 46%, compared to an average of only 11% over the past year. Artificial Analysis benchmark testing indicates that running the same test costs approximately $4,811 with Anthropic's Claude model, $1,071 with DeepSeek, $948 with Moonshot AI's Kimi, and only $544 with Zhipu AI's GLM model, demonstrating the competitive pricing of Chinese-made models.
Congress Launches Security Investigation in April 2026
In April 2026, the House Homeland Security Committee and the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party jointly sent letters to Anysphere (operator of AI code editor Cursor) and Airbnb, requesting explanations regarding their use of Chinese-made AI models, and launched a joint investigation. Records show that Cursor's Composer 2 model is built on the foundation of Moonshot AI's Kimi. House Homeland Security Committee Chairman Andrew Garbarino stated that reports indicating Chinese open-weight models can match US mainstream models in "specific vulnerability detection and cybersecurity tasks" are "extremely concerning."
Anthropic Reports Distillation Attempts in February 2026
Anthropic published investigation results in February 2026, alleging that DeepSeek, Moonshot AI, and MiniMax used approximately 24,000 fake accounts to conduct over 16 million interactions with its Claude model, attempting to improperly extract Claude's capabilities through "distillation" techniques. The White House Office of Science and Technology Policy characterized this action as an "industrial-scale organized imitation operation" in April 2026. Notably, Zhipu AI has been listed on the US Commerce Department's export control entity list since January 2025.
Commerce Department Imposes Export Controls
Booz Allen Hamilton testing conducted in May 2026 found that when given prompts similar to US government agency usage scenarios, three of four Chinese-made coding models produced increased code vulnerabilities, with Qwen3-Coder's vulnerabilities increasing approximately 130%. Due to US Commerce Department export controls, Anthropic's two most advanced models, Mythos 5 and Fable 5, were suspended from June 12, 2026, and did not resume service until July 1, 2026. Cybersecurity consultant Xin Shijie noted that enterprise AI application has shifted from "whether to use" to "which country and governance system's AI model to use," warning that many companies may unknowingly assume supply chain risks. He emphasized that the "price disruption" of open-weight models is an irreversible trend, creating significant challenges for US companies providing top-tier AI model services to balance pricing strategies with security guarantees.
FAQ
Why are US companies switching to Chinese AI models?
US companies are adopting Chinese AI models primarily to reduce operational costs. According to Artificial Analysis benchmarks, running identical tests costs approximately $4,811 with Anthropic's Claude versus $1,071 with DeepSeek, $948 with Moonshot AI's Kimi, and $544 with Zhipu AI's GLM. Lindy switched all traffic from Claude to DeepSeek in June 2026, expecting millions in savings, while Coinbase cut AI expenses nearly in half by switching to GLM-5.2 and Kimi K2.7 Code in late June 2026.
What security concerns exist regarding Chinese AI models?
Anthropic reported in February 2026 that DeepSeek, Moonshot AI, and MiniMax used approximately 24,000 fake accounts for over 16 million interactions to improperly extract Claude's capabilities through distillation techniques. The White House Office of Science and Technology Policy characterized this as an "industrial-scale organized imitation operation" in April 2026. Additionally, Booz Allen Hamilton testing in May 2026 found that three of four Chinese coding models produced increased code vulnerabilities when given government-related prompts, with Qwen3-Coder's vulnerabilities increasing approximately 130%.
What actions has Congress taken regarding Chinese AI model usage?
In April 2026, the House Homeland Security Committee and the House Select Committee on the Strategic Competition Between the United States and the Chinese Communist Party jointly sent letters to Anysphere (operator of Cursor) and Airbnb, requesting explanations about their use of Chinese-made AI models, and launched a joint investigation. House Homeland Security Committee Chairman Andrew Garbarino expressed concerns about reports that Chinese open-weight models can match US models in specific cybersecurity tasks.
