Why US companies are quietly switching to Chinese AI models

Several well-known US companies, including Coinbase, DoorDash and Airbnb, are quietly shifting some of their everyday AI work onto open-weight models built in China, because they are much cheaper than the leading US alternatives.

Open-weight models are AI systems whose underlying settings are published, so anyone can download and run them on their own computers rather than paying to access them through a company's website. Fortune reported on 26 July that three Chinese labs, Moonshot, Z.ai and DeepSeek, are winning US business largely on price, according to Fortune.

The price gap is large. Fortune found that for a million output tokens (roughly a measure of text generated), Anthropic's Claude Fable 5 costs $50, while Moonshot's Kimi K3 costs $15, Z.ai's GLM-5.2 costs $4.40, and DeepSeek-V4-Pro costs about $0.87.

Coinbase has gone furthest. Chief executive Brian Armstrong said in late June that the company nearly halved its internal AI spending by making Zhipu's GLM 5.2 and Moonshot's Kimi K2.7 Code the default models for its engineers, using an internal system that sends each task to whichever model offers the best price for the job. Better caching, storing repeated answers instead of recalculating them, pushed the share of requests served this way from 5% to 60%. Coinbase said 91% of engineers never hit their original usage limits, and the savings came from shifting routine work like code review and document summaries away from top-tier Anthropic and OpenAI models, not from cutting anyone's access.

DoorDash co-founder Andy Fang said the company hands "lower-level work" to Kimi because it gets better quality for less money. In DoorDash's own testing, pairing Kimi K2.6 as a first-pass reviewer with Claude Fable 5 as a final check gave the best results, at $3.81 per code change reviewed, according to The Decoder. Bloomberg has reported that Airbnb, DoorDash and Coinbase run these Chinese models on their own servers rather than through a public website.

Moonshot released its newest model, Kimi K3, on 16 July, reportedly the largest open model yet built, with its open weights due to be published on 27 July.

If you work at a company that uses AI tools for coding, writing or summarising documents, some of that work may already be quietly running on Chinese models behind the scenes, chosen because they are cheaper, not because they are branded differently. This is not something most individual users need to act on, but it shows that the cost gap between US and Chinese AI is now large enough to change how companies build their products.

There is no ban on US companies using these models. Some government agencies have banned DeepSeek specifically, but that restriction does not apply to private companies.

The shift has drawn political attention. The House Homeland Security and China committees have written to DoorDash asking for a list of the Chinese models it uses and details of its security testing, with a response due by 14 August.