Is China quietly winning the AI race?

Is China quietly winning the AI race?

33 minutes ago Share Save Lily Jamali North America Technology correspondent Share Save

Getty Images

Every month, hundreds of millions of users flock to Pinterest looking for the latest styles. One paged titled "the most ridiculous things" is filled with plenty of wacky ideas to inspire creatives - Crocs repurposed as flower pots. Cheeseburger-shaped eyeshadow. A gingerbread house made of vegetables. But what would-be buyers may not know is the tech behind this isn't necessarily US-made. Pinterest is experimenting with Chinese AI models to hone its recommendation engine. "We've effectively made Pinterest an AI-powered shopping assistant," the firm's boss Bill Ready told me. Of course, the San Francisco-based tastemaker could use any number of American AI labs to power things behind-the-scenes. But since the launch of China's DeepSeek R-1 model in January 2025, Chinese AI tech has increasingly been a part of Pinterest. Ready calls the so-called "DeepSeek moment" a breakthrough. "They chose to open source it, and that sparked a wave of open source models," he said. Chinese competitors include Alibaba's Qwen and Moonshot's Kimi, while TikTok owner ByteDance is also working on similar technology. Pinterest Chief Technology Officer Matt Madrigal said the strength of these models is that they can be freely downloaded and customised by companies like his - which is not the case with the majority of models offered by US rivals like OpenAI, which makes ChatGPT. "Open source techniques that we use to train our own in-house models are 30% more accurate than the leading off-the-shelf models," Madrigal said. And those improved recommendations come at a much lower cost, he said, sometimes ninety percent less than using the proprietary models favoured by US AI developers.

'Fast and cheap'

Pinterest is hardly the only US enterprise depending on AI tech from China. These models are gaining traction across an array of Fortune 500 companies. Airbnb boss Brian Chesky told Bloomberg in October his company relied "a lot" on Alibaba's Qwen to power its AI customer service agent. He gave three simple reasons - it's "very good", "fast" and "cheap". Further evidence can be found on Hugging Face, the place people go to download ready-made AI models - including from major developers Meta and Alibaba. Jeff Boudier, who builds products at the platform, said it is the cost factor that leads young start-ups to look at Chinese models over their US counterparts. "If you look at the top trending models on Hugging Face - the ones that are most downloaded and liked by the community - typically, Chinese models from Chinese labs occupy many of the top 10 spots," he told me. "There are weeks where four out of five top training models on Hugging Face are from Chinese labs." In September, Qwen topped Meta's Llama to become the most downloaded family of large language models on the Hugging Face platform. Meta released its open-source Llama AI models in 2023. Up until the release of DeepSeek and Alibaba's models, they were considered the go-to choice for developers working on bespoke applications. But the release of Llama 4 last year left developers underwhelmed, and Meta has reportedly been using open-source models with Alibaba, Google, and OpenAI to train a new model set for release this spring. Airbnb also uses several models, including US-based ones, hosting them securely in the company’s own infrastructure. The data is never provided to the developers of the AI models they use, according to the company.

Chinese success