The State Department has raised alarm about US companies using Chinese-built AI models, telling CNBC they "raise serious concerns" because they are "designed to advance Beijing's narratives, censor dissent and reflect CCP ideology and values." The remarks come as US lawmakers explore ways to slow the growing use of Chinese AI systems within American firms.

The shift is no longer limited to small startups. Coinbase CEO Brian Armstrong disclosed that the company runs two Chinese models—GLM 5.2 from Z.Ai and Kimi 2.7 from Moonshot—a revelation CNBC reported alongside State Department comments. Broader reporting from Nikkei Asia flagged adoption at Airbnb and Uber. Lindy, a startup, switched to DeepSeek in June specifically to reduce soaring inference costs—the computational expense of running an AI model once it's already built.

Two factors explain the rapid shift. First, cost economics: the share of AI queries routed to Chinese models by US companies on the OpenRouter platform has risen significantly through 2026, tracking a wider move toward efficiency as pricing from OpenAI and Anthropic strains corporate budgets. Second, distribution: many leading Chinese models are released as open-weight systems—meaning the underlying code is published—so US firms can download and run them on their own servers, bypassing API access controls entirely. This direct access is central to why Chinese developers are reaching US users at all.

Nikkei Asia's reporting adds context: the surge appears to coincide with Anthropic's suspension of its Mythos 5 and Claude Fable 5 models at the US government's request. Restricting the frontier domestically appears to have nudged some demand toward Chinese alternatives rather than eliminating it. CNBC reported in June that the administration's broader clampdown on AI access could give Chinese labs time to narrow the capability gap with Anthropic and OpenAI.

Washington's policy architecture toward AI has been building for roughly a year. The White House's "America's AI Action Plan," published July 2025, recommended withholding federal AI funding from states it deemed to have burdensome regulations. A December 2025 presidential action directed review of state AI laws under a national framework. March 2026 brought a national AI legislative framework, followed in April by NSTM-4, a national security memorandum addressing model security and what officials term ideological neutrality. Executive Order 14409, "Promoting Advanced Artificial Intelligence Innovation and Security," came on June 2. The US also pressed allies to restrict overseas access to leading American-built models during a June summit in France that excluded Beijing.

Export controls on hardware have tightened in parallel. Reuters reported in May that the US moved to halt Nvidia AI chip shipments to Chinese firms operating outside China, following evidence that advanced chips were reaching subsidiaries of Chinese AI companies through indirect channels. April brought new US restrictions on investment in Chinese AI, semiconductor and quantum sectors, with Beijing preparing matching curbs on US investment in Chinese tech firms.

Beijing's own stance on AI exports has shifted as well. TIME reported that Chinese AI firms grew global market share by offering models for free, but Beijing is now weighing restrictions on overseas access to its most capable systems. Reuters reported similarly that Chinese authorities have opened talks with domestic companies about preventing their AI from being used internationally altogether. Chinese regulators have also made security allegations of their own: the Ministry of Industry and Information Technology announced that Anthropic's Claude Code contains what it calls a backdoor and a "serious threat."

Security threats extend beyond model outputs. China-linked actors have targeted AI startups through cyberattacks, insider threats and espionage as competition in the sector intensifies, CNBC reported.

What is emerging is a simultaneous two-sided restriction effort: each government is working to block the other's models and chips from crossing its borders. Yet commercial incentives on both sides pull in the opposite direction. Cost pressure and open-weight distribution are powerful enough that policy alone has struggled to contain them so far. Coinbase's disclosed use of Chinese models and Lindy's cost-driven switch to DeepSeek both show that enterprise procurement decisions are outpacing regulatory intent on both sides of the Pacific.