Every month, hundreds of millions of users arrive on Pinterest searching for what’s next—new fashion trends, interior design ideas, recipes, and visual inspiration that might spark their next purchase. For many, the platform is less a social network than a discovery engine, a place where tastes are formed before wallets are opened.
One popular page, titled “the most ridiculous things,” offers a glimpse into the strange and whimsical corners of Pinterest’s universe. There are Crocs repurposed as flower pots. Cheeseburger-shaped eyeshadow palettes. A gingerbread house constructed entirely from vegetables. It’s chaotic, creative, and oddly compelling.
What most users browsing these boards don’t realise, however, is that the technology shaping what they see may not be American-made.
Behind the scenes, Pinterest has been experimenting with Chinese artificial intelligence models to fine-tune its recommendation systems—part of a quiet but growing shift that is rippling across Silicon Valley and corporate America more broadly. The change reflects a deeper transformation in the global AI landscape, one that is challenging long-held assumptions about where the most advanced and influential AI technology comes from.
“We’ve effectively made Pinterest an AI-powered shopping assistant,” chief executive Bill Ready told me.
Pinterest, based in San Francisco and long regarded as a tastemaker for global consumer trends, could have chosen from a wide range of US-based AI providers to power its systems. American companies still dominate headlines, investment figures, and public attention in the AI race.
But since the release of China’s DeepSeek R-1 model in January 2025, Chinese-developed AI has increasingly become part of Pinterest’s technology stack.
Ready describes what some in the industry now call the “DeepSeek moment” as a genuine breakthrough.
“They chose to open source it,” he said, “and that sparked a wave of open-source models.”
That decision proved pivotal. Unlike many high-profile US models that are locked behind proprietary systems and paid APIs, DeepSeek R-1 was made freely available to developers. It could be downloaded, modified, and adapted for specific business needs—a powerful incentive for companies seeking flexibility, performance, and lower costs.
DeepSeek is not alone. Other Chinese tech giants have followed a similar path. Alibaba’s Qwen models and Moonshot AI’s Kimi have gained traction, while TikTok owner ByteDance is developing comparable large language models of its own. Together, they form a growing ecosystem of Chinese AI tools that are increasingly competitive with, and in some cases preferred over, their American counterparts.
Pinterest’s chief technology officer, Matt Madrigal, said the appeal of these models lies in how easily they can be customised.
“The strength of open-source models is that we can download them and adapt them to our own use cases,” he said. “That’s simply not the case with most US models.”
Companies like OpenAI, which developed ChatGPT, typically offer access to their most advanced systems through paid services. While powerful, those models are largely closed off from deep modification.
Madrigal said Pinterest’s in-house models—trained using open-source techniques and building on foundations like DeepSeek—are delivering significant performance gains.
“The open-source techniques that we use to train our own models are about 30% more accurate than the leading off-the-shelf models,” he said.
There is also a financial incentive. According to Madrigal, the cost savings can be dramatic—sometimes as much as 90% less than relying on proprietary AI services from US developers.
Fast, cheap, and good enough
Pinterest is far from alone.
Across corporate America, Chinese AI models are quietly gaining ground. They are being adopted by a growing number of Fortune 500 companies, often not as headline-grabbing replacements for US systems, but as practical tools that deliver strong results at a fraction of the price.
Airbnb is one notable example. In October, chief executive Brian Chesky told Bloomberg that the company relies “a lot” on Alibaba’s Qwen model to power its AI-driven customer service agent.
He offered three simple reasons: it is “very good,” “fast,” and “cheap.”
That combination—performance, speed, and cost—is proving hard for US developers to match, particularly for companies operating at scale.
Further evidence of this trend can be found on Hugging Face, a popular platform where developers download and share ready-made AI models. The site hosts tools from major global players, including Meta and Alibaba, and serves as a bellwether for what developers are actually using.
Jeff Boudier, who helps build products at Hugging Face, said cost is often the decisive factor, especially for start-ups and smaller firms.
“If you look at the top trending models on Hugging Face—the ones that are most downloaded and liked by the community—Chinese models from Chinese labs occupy many of the top 10 spots,” he told me.
“There are weeks where four out of five of the top training models on Hugging Face are from Chinese labs.”
In September, Alibaba’s Qwen overtook Meta’s Llama to become the most downloaded family of large language models on the platform—a symbolic milestone.
When Meta released its open-source Llama models in 2023, they were widely regarded as the default choice for developers building custom AI applications. For a time, they dominated the open-source ecosystem.
But momentum has shifted. The release of Llama 4 last year left many developers underwhelmed, and Meta has reportedly turned to external open-source models—including those from Alibaba, Google, and OpenAI—to help train its next generation of systems, expected later this spring.
Airbnb, for its part, uses a mix of AI models from different providers, including US-based ones. The company says all models are hosted securely within its own infrastructure, and that no customer data is shared with the developers of the underlying AI.
Rethinking who’s ahead
At the start of 2025, the prevailing view was that despite massive investment by US tech giants, Chinese AI companies were at risk of falling behind—constrained by export controls, limited access to cutting-edge chips, and geopolitical pressure.
“That’s not really the story anymore,” Boudier said. “Now, the best model is an open-source model.”
A report published last month by Stanford University supports that assessment. It found that Chinese AI models “seem to have caught up or even pulled ahead” of their global counterparts, both in terms of technical capability and user adoption.
The findings challenge the idea that leadership in AI is defined solely by who builds the biggest or most expensive models. Instead, they point to the growing importance of accessibility, efficiency, and real-world deployment.
In a recent BBC interview, Sir Nick Clegg, the former UK deputy prime minister, argued that US companies may be overly focused on the pursuit of artificial general intelligence—systems that could one day surpass human intelligence—at the expense of more practical goals.
Sir Nick stepped down last year as head of global affairs at Meta, whose chief executive Mark Zuckerberg has pledged billions of dollars toward achieving what he calls “superintelligence.”
Some experts now argue that these ambitions are vague and ill-defined, consuming resources while leaving room for competitors to dominate the more immediately useful layers of the AI stack—particularly open-source tools that developers can build on today.
“Here’s the irony,” Sir Nick said. In the ideological contest between “the world’s great autocracy” and “the world’s greatest democracy,” China and the United States, he argued that China is “doing more to democratise the technology they’re competing over.”
The Stanford report also suggests that China’s success in open-source AI may be partly explained by state support, which allows companies to prioritise adoption and influence over short-term profitability.
By contrast, US firms face intense commercial pressure. Companies like OpenAI are under growing scrutiny from investors to generate revenue and chart a clear path to profitability. OpenAI has begun exploring advertising and enterprise services as part of that effort.
Although OpenAI released two open-source models last summer—its first in years—it has concentrated most of its resources on proprietary systems that can be monetised.
Chief executive Sam Altman told me in October that the company is investing aggressively in computing power and infrastructure.
“Revenue will grow super fast,” he said, “but you should expect us to invest a ton in training, in the next model and the next and the next and the next.”
A quieter race
None of this means the United States has lost its edge in AI. American firms still lead in funding, cutting-edge research, and the scale of their infrastructure. But the race is no longer playing out solely at the frontier of raw capability.
Instead, it is unfolding more quietly, in places like Pinterest’s recommendation engine or Airbnb’s customer service tools—where open-source, adaptable, and affordable AI models are increasingly winning out.
For users scrolling past vegetable gingerbread houses and cheeseburger makeup palettes, the origins of the technology shaping their feed may seem irrelevant. But collectively, these choices reveal a broader shift: one in which China is becoming an indispensable player in the global AI ecosystem, not through grand announcements, but through tools that work—and that anyone can use.
