U.S. Congress Launches Investigation Into Chinese AI Use in Airbnb and Cursor with 46% Token Share

A CNBC report published this Wednesday revealed that U.S. lawmakers have launched an investigation into the use of Chinese artificial intelligence models within Airbnb and Anysphere, the parent company of the coding platform Cursor. Both companies admitted to integrating open models such as Qwen from Alibaba and Kimi from Moonshot AI into parts of their product pipelines. As of the time of this writing, there had been no public response from either company regarding the scope of the investigation or the position of the office conducting it.

The report coincides with data that has been developing over the past five months and has now gained political weight. The share of tokens generated by American companies using Chinese models via OpenRouter, the multi-model router for developers, has remained above 30% every week since February 8, peaking at 46%. The average over the previous 12 months was 11%, with 4.5% in the first half of 2025.

Why Companies Adopted

The trigger is price. Chinese open models appear in benchmarks to be between 60% and 90% cheaper per million tokens than the main SKUs from OpenAI and Anthropic. In coding workloads and agents, where a single agent can consume millions of tokens per session, cutting costs is the difference between negative gross margin and a viable product.

The second trigger is open licensing. Available weights allow customers to host the model within their own VPC, without sensitive prompts traveling through third-party APIs. For a company like Airbnb, which handles reservation data, check-in, and identity verification, this is not a detail: it is a pathway to meet GDPR requirements in Europe and LGPD in Brazil. For a startup like Anysphere, whose Cursor product competes with Microsoft's Copilot and Anthropic's Claude Code, cutting variable costs per autocomplete line is survival.

The case of Z.ai illustrates the speed. The GLM-5.2 model saw an 80-fold growth in clients and a 27-fold increase in daily token volume in its first complete week within Vercel, marking the fastest adoption of a single model on the platform this year. The gap between technical availability and scalable production has shifted from quarters to days.

The Political Context of the Investigation

The investigation does not occur in a vacuum. On June 10, Anthropic sent a letter to Senators Tim Scott, a Republican from South Carolina, and Elizabeth Warren, a Democrat from Massachusetts, claiming that operators associated with Alibaba's Qwen team used around 25,000 fraudulent accounts to generate over 28.8 million exchanges with Claude between April 22 and June 5, with the aim of extracting capabilities from the American model. Alibaba denies it and accuses Anthropic of embedding tracking code into developer tools, a dispute that has led Chinese companies to instruct employees to avoid Claude Code.

Shortly after the letter, the U.S. Department of Commerce placed Anthropic's Mythos and Fable models under export controls. The chain of events puts the technological foundation of American companies in a vise: American models are becoming more expensive and more regulated, while Chinese models are getting cheaper and more capable, prompting Congress to ask these companies why the adoption curve has followed economic incentives.

Reading for CIOs Across Markets

In the United States, the investigation is expected to produce mandatory disclosures from model suppliers in the coming weeks. CIOs in regulated sectors that are already using Qwen or Kimi in production need to anticipate inventory and swap pipelines. A direct ban on corporate use is a possibility being discussed in Washington since April, and the supporting legislation has bipartisan support.

In Europe, the issue is entering the review of the EU AI Act through the lens of systemic risk and supply chain transparency. The AI Office is expected to require platform operators to declare the use of open models of non-European origin. In Singapore and Japan, authorities have signaled interest in tracking dependency on Chinese-origin AI infrastructure in banks and telecoms.

In Brazil, the debate is still in the mapping stage. Banks like Itaú and Bradesco, as well as consultancies like CI&T and Falconi, have regulatory freedom to test open models, but compliance offices are starting to inquire about the provenance of weights, especially in contracts involving health or financial data. The question that matters to the board is not whether the Chinese model delivers, but whether the exit cost, should the regulator shift, fits within the budget.

What this week has revealed is that AI competition has moved out of the laboratory and into the cost accounting. And that the cost accounting may become the subject of investigation.