American Companies Keep Turning to Chinese AI Models

The price gap between American and Chinese AI models has grown too large for many companies to ignore. As the cost of running frontier systems from OpenAI and Anthropic climbs, a report from CNBC on July 7 describes US firms leaning harder on open-weight models from Chinese labs to keep their bills down.

The math is stark. A coding session that runs about $10 on Claude can cost less than 50 cents on DeepSeek, according to reporting by Rest of World. For a startup burning tokens all day, that difference decides which model goes into production.

How deep the adoption runs

Venture investors estimate that roughly 80 percent of US startups now build on Chinese base models, adapting them into their own products. The pattern reaches established companies too. Airbnb has used Alibaba's Qwen to power customer service, citing low cost and quality close to leading American systems.

The shift shows up in usage data. At the developer platform Vercel, DeepSeek's share of tokens processed jumped from under 1 percent to 17 percent in a single month last spring. The research group RAND has tracked a sharp rise in traffic to Chinese models and a climb in their global market share. Cloud providers have followed the demand, with AWS and Microsoft both offering DeepSeek and Qwen through their catalogs.

The complications

Cheap capability comes with questions that boardrooms are only starting to work through. Lawmakers have opened inquiries into companies, among them Airbnb and Cursor owner Anysphere, that disclosed building on Chinese open models. The concerns range from data handling to the wisdom of depending on systems shaped by a strategic rival.

None of that has slowed the trend much. When one model costs twenty times more than another that performs nearly as well, cost tends to win. The open-weight strategy from Chinese labs, covered here before with releases like Meituan's LongCat 2.0 and Z.ai's GLM 5.2, is turning free weights into real market share inside American software.