AI has become one of the fastest-growing line items on many companies’ budgets, and a growing number of them are responding the same way businesses always respond to rising costs: shopping around.
Data from OpenRouter, a platform developers use to access AI models from multiple providers, shows the share of tokens US companies are running on Chinese AI models has sat above 30% every week since February 8, spiking as high as 46%, up from a 12-month average of just 11% and a mere 4.5% in the first half of 2025.
This isn’t a political statement companies are making. It’s a spreadsheet decision.
The Companies Making the Switch, and What They’re Saving
Lindy, a San Francisco AI assistant startup, moved 100% of its traffic off Anthropic’s Claude models onto China’s DeepSeek-V4 in June.
Founder Flo Crivello said the move was roughly ten times cheaper and would save the company millions of dollars within months, calling it, in his words, a very simple business decision.
Individual developers are making the same calculation at smaller scale: San Diego developer Stu Clott said an hour of coding that used to cost about $10 on Claude runs him less than 50 cents on DeepSeek.
Shopify found an even sharper gap after replacing a GPT-5-based pipeline for merchant data extraction with a self-hosted, fine-tuned version of Alibaba’s Qwen3, reporting a 75-fold reduction in per-unit costs alongside higher output quality. Not every company is going all-in, though.
Coinbase CEO Brian Armstrong has described routing lower-stakes internal tasks to cheaper open-weight models like GLM and Kimi while keeping frontier US models for harder problems, treating model selection more like choosing cloud infrastructure than picking a single vendor.
Airbnb CEO Brian Chesky has said the company relied on Alibaba’s Qwen for parts of its stack, calling it fast, cheap, and good enough for the job.
Why the Capability Gap is Narrowing Enough to Make This Workable
The switching only makes sense because Chinese models have closed most of the performance gap while staying dramatically cheaper.
Brookings fellow Kyle Chan estimates leading Chinese models now trail top US frontier systems by roughly six to nine months, close enough that they handle the overwhelming majority of everyday tasks, coding, customer support, data extraction, without a noticeable quality hit.
The clearest data point: Z.ai’s GLM 5.2, released in June, landed within a single percentage point of Anthropic’s Opus 4.8 on one closely watched agentic benchmark, at roughly a fifth of the cost, and saw the fastest first-week adoption of any model tracked by infrastructure platform Vercel in 2026, with daily token volume growing 27-fold and its customer count growing 80-fold in seven days.
That kind of adoption curve is happening despite real friction.
US lawmakers have opened investigations into Airbnb and Cursor-maker Anysphere after both disclosed using Chinese open models, and most companies using them route traffic through US-based intermediaries like OpenRouter or Featherless specifically to keep user data within US borders rather than dealing directly with Chinese providers.
That friction also explains why Chinese AI companies are struggling to convert usage into revenue: DeepSeek’s share of Vercel’s gateway token traffic jumped to 17% in May, but its share of platform revenue stayed near 1%, a gap that reflects both aggressive pricing and lingering caution among enterprise buyers about paying a Chinese company directly.
The pattern emerging isn’t wholesale replacement of US models so much as tiered routing, expensive frontier models reserved for the hardest problems, cheaper Chinese alternatives absorbing the high-volume routine work that used to run through the same premium pipeline by default.
Source: The Times of India, "One of Biggest AI Problems Is Forcing American and European Companies to Use Chinese Models"
