Every month, hundreds of millions of people visit Pinterest to explore new styles and creative ideas. One popular page called “the most ridiculous things” offers deliberately strange inspiration. It shows Crocs turned into flower pots. It features cheeseburger-shaped eyeshadow. It even presents a gingerbread house built from vegetables.
Many users do not realise the technology behind these recommendations is not always American. Pinterest now tests Chinese artificial intelligence models to improve its recommendation engine. The company increasingly relies on this technology to personalise shopping and discovery.
Pinterest chief executive Bill Ready said the platform now functions as an AI-powered shopping assistant. The San Francisco-based company could work with many American AI labs. Instead, it increasingly integrates Chinese-developed models behind the scenes.
A model that changed the conversation
China’s role at Pinterest expanded after the release of DeepSeek R-1 in January 2025. Ready described the launch as a defining moment. He said the developers released the model as open source. That decision triggered a wave of new open-source projects.
The release encouraged fast experimentation across the technology sector. Other Chinese firms quickly followed the same path. Alibaba expanded its Qwen models. Moonshot released its Kimi system. ByteDance also develops similar large language technology.
These models now compete directly with established American systems. They increasingly power digital products used by millions of people worldwide.
Open source as a competitive advantage
Pinterest Chief Technology Officer Matt Madrigal said open-source access makes these models especially attractive. Companies can download and adapt them internally. Many American rivals restrict access to their most advanced systems.
Madrigal said Pinterest trains its own models using open-source techniques. He said these systems outperform many off-the-shelf alternatives. According to him, accuracy improves by roughly 30 percent.
Costs also fall sharply. Madrigal said expenses sometimes drop by as much as ninety percent. Proprietary models from US developers often require significantly higher spending.
Adoption spreads across US companies
Pinterest is not the only American company turning to Chinese AI. Many large firms now depend on these models. Their adoption continues to spread across corporate America.
Airbnb chief executive Brian Chesky said his company relies heavily on Alibaba’s Qwen. The model powers Airbnb’s AI customer service agent. Chesky explained the choice in simple terms. He said it works very well. He said it runs fast. He said it costs less.
More evidence appears on Hugging Face, a major platform for downloading AI models. Developers there access systems from companies like Meta and Alibaba. The platform tracks which models gain the most traction.
Chinese models dominate demand
Jeff Boudier, who builds products at Hugging Face, said cost plays a central role. Young start-ups often choose Chinese models over American ones. Download data clearly reflects that preference.
Boudier said Chinese models frequently occupy top positions in popularity rankings. In some weeks, four of the five leading training models come from Chinese labs. That pattern appears repeatedly.
In September, Alibaba’s Qwen overtook Meta’s Llama. It became the most downloaded family of large language models on the platform. Developers quickly adjusted their strategies.
Silicon Valley responds
Meta released its open-source Llama models in 2023. Developers long considered them the default option for custom applications. That status weakened after DeepSeek and Alibaba entered the market.
Meta released Llama 4 last year. Many developers described the update as underwhelming. Reports suggest Meta now uses open-source models from Alibaba, Google, and OpenAI to train a new system. The company plans to release it this spring.
Airbnb uses several AI models at the same time, including American systems. The company hosts them within its own infrastructure. It says it never shares user data with model developers.
The balance begins to shift
At the start of 2025, many analysts believed Chinese firms threatened to pull ahead. Massive American investment no longer guaranteed leadership. The discussion has since changed.
Boudier said the strongest models now come from open-source communities. A recent Stanford University report supports that view. Researchers found Chinese models have matched or surpassed global competitors.
The study measured technical capability and user adoption. It suggested Chinese developers have closed the gap. In some areas, they have already moved ahead.
Competing priorities shape outcomes
In a recent interview with a British broadcaster, former UK deputy prime minister Sir Nick Clegg criticised American priorities. He said US firms focus too heavily on building AI beyond human intelligence.
Clegg previously led global affairs at Meta. Mark Zuckerberg has committed billions to achieving what he calls superintelligence. Some experts now describe those ambitions as poorly defined.
Clegg said this lack of clarity gives China an opening. He argued China now does more to democratise the technology it competes over.
Commercial pressure in the United States
The Stanford report also pointed to strong government backing inside China. That support may explain part of its success in open-source development.
Meanwhile, American AI companies face growing pressure to increase revenue. Firms like OpenAI must balance research goals with profitability. Some now turn to advertising to support growth.
OpenAI released two open-source models last summer, its first such move in years. Most resources still flow into proprietary systems designed to generate income.
OpenAI chief executive Sam Altman said the company invests aggressively in computing power. He said revenue will grow quickly. He also said spending on future models will remain substantial.
