U.S. companies accelerate shift to Chinese AI models to cut costs, Congress begins to worry and launches investigation.

Chinese open-source AI models are rapidly infiltrating the AI procurement landscape of U.S. enterprises, with performance approaching top-tier U.S. systems at dramatically lower costs: OpenRouter data shows that the share of U.S. enterprises using Chinese models has surged several times compared to a year ago. The House Homeland Security Committee and the Select Committee on the Chinese Communist Party have launched a joint investigation.

(Previous summary: Analyst predicts: By 2029, enterprise AI token spending could be more expensive than engineer salaries)

(Background: U.S.-China AI competition heats up, but scholars from both countries agree: Don't let AI usher in a "Chernobyl moment")

Table of Contents

Toggle

• How much cheaper, who is switching?
• How close is performance, the tension between open source and closed source
• Why is Washington nervous, can it be stopped?

By the numbers. Over the past 12 months, the average token share of Chinese AI models used by U.S. enterprises on the OpenRouter platform was only 11%, and in the first half of 2025, it was just 4.5%. But since February 8 this year, this figure has exceeded 30% every week, peaking at 46%.

Cheap and good enough is rewriting the AI procurement logic of U.S. enterprises, and it’s also making the U.S. Congress restless.

How much cheaper, who is switching?

The most important reason driving enterprise migration is price. In June, AI startup Lindy moved 100% of its usage from Anthropic's Claude to China's DeepSeek. CEO Flo Crivello posted, "We did it, you can see that cost curve plummet straight down," estimating it would save the company millions of dollars within a few months.

Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models.
Saves us millions of $ and we're actually seeing an increase in performance on many core use cases. Transformative for the business.

— Flo Crivello (@Altimor) June 4, 2026

Coinbase CEO Brian Armstrong also publicly recommended using Chinese models to reduce costs. OpenRouter analyst Justin said that open-source Chinese models are 60% to 90% cheaper than the top-tier models from Anthropic and OpenAI, "and can handle all but the most complex tasks."

As AI costs continue to soar, Chinese models are becoming increasingly attractive to U.S. enterprises.

How close is performance, the tension between open source and closed source

Z.ai's GLM 5.2 was released at the end of June and is the fastest-adopted model tracked by the Vercel platform this year: in its first full week after release, daily token usage grew about 27 times, and the number of customers using it grew about 80 times. Vercel's Harpreet Arora said, "When a task doesn't require the strongest model, teams start directing it to the one that's good enough and cheapest, and this wave of Chinese models is winning that trade."

The numbers are more direct: In one benchmark test, GLM 5.2 was less than one percentage point behind Anthropic's Opus 4.8, but at one-fifth the cost. On some security benchmarks, GLM 5.2 has already matched top U.S. labs. Kyle Chan estimates that Chinese models currently lag behind top U.S. rivals by about six to nine months, and the gap is closing rapidly.

Why is Washington nervous, can it be stopped?

In April, the House Homeland Security Committee and the Select Committee on the Chinese Communist Party announced a joint investigation into the prevalence of Chinese models in U.S. enterprises. A State Department spokesperson told CNBC that the growing use of Chinese models by U.S. enterprises is "deeply concerning," saying the models are designed to advance Beijing's narratives and censor dissent.

House Homeland Security Committee Chairman Andrew Garbarino said, "The Chinese Communist Party is no longer just nipping at our heels; it is rapidly closing the gap," noting that a Chinese open-weight model has already matched top U.S. models in security tasks like vulnerability discovery.

Subcommittee Chairman Andy Ogles called for making U.S. models a real alternative: "When the cheap and capable option is Chinese, the whole world will build on top of it." Kyle Chan said the government might consider federal procurement bans, "but ultimately it's impossible to ban Chinese open-source models because open source is freely available online," which touches on First Amendment issues.

CNAS scholar Daniel Remler also believes that Washington will most likely only express displeasure rather than actually imposing a ban.