Beijing's AI Labs Are Now Cheap Enough for Fortune 500 Budgets
The numbers came fast and hard. When Moonshot AI debuted Kimi K3 on July 16, Nvidia lost almost $600 billion in market value and briefly surrendered its position as the world's most valuable company to Apple. The chip-focused Philadelphia Semiconductor Index fell 1.6%. Markets across Asia sold off.
The trigger: a Beijing-based startup released what it described as the largest open-source AI model ever, benchmarked close to Anthropic's Fable 5 — widely regarded as the most powerful publicly available model on the market today — at a fraction of the cost. One independent benchmark from Arena.AI ranked Kimi K3 the best model currently available, ahead of Anthropic.
Moonshot AI was founded by Yang Zhilin, a 34-year-old Tsinghua and Carnegie Mellon alumnus. Its official benchmarks consistently place K3 among the top three AI models globally.
The export-control thesis took a hit.
Washington began restricting China's access to top-tier AI processors — including those made by Nvidia — in 2022, with the explicit goal of preserving America's AI lead. The strategy assumed that without advanced chips, Chinese labs could not train competitive frontier models.
DeepSeek, a Hangzhou-based lab attached to a Chinese hedge fund, challenged that assumption in early 2025 when it released its V3 and R1 models, which matched U.S. counterparts in performance. DeepSeek claimed it trained the models on a small budget by engineering efficiency through programming and mathematical techniques.
Kimi K3 pushed the timeline further. Anthropic CEO Dario Amodei had not expected a Chinese lab to reach this level for at least another six months. Tesla CEO Elon Musk had suggested it might happen by the first quarter of next year. Both forecasts proved too conservative.
Even a food-delivery app is in the frontier race.
Meituan — best known as a food-delivery platform — released its LongCat-2.0 model, which the company said matched OpenAI and Anthropic releases from February and was trained entirely on Chinese-made processors rather than U.S. chips. The model encompassed as much data as DeepSeek's V4.
'The idea that Meituan could train a 1.6 trillion-parameter model on domestic hardware would have been inconceivable in October 2022,' said Paul Triolo, a partner at DGA–Albright Stonebridge Group, referring to the month U.S. export controls launched.
In June, AI startup Z.ai had already stolen headlines with its GLM-5.2 model, which showed particular strength in coding and creative design. Z.ai's stock surged more than 1,100% through mid-July, at times breaking 1 trillion Hong Kong dollars ($127.6 billion) in market capitalization — valuing a company with just $106 million in revenue last year the same as BYD and Starbucks. Kimi K3's launch then sent Z.ai shares tumbling 40% in two days.
Chinese models now dominate much of the activity on OpenRouter, a marketplace where developers access multiple AI models through a single interface.
The market has already voted.
U.S. startups and Fortune 500 companies are quietly integrating Chinese AI models into their operations to control spiraling AI budgets, according to the Fortune report. The free-market logic is straightforward: when a competitor delivers comparable capability at lower cost, capital follows. That is not a failure of American ingenuity — it is a signal that the regulatory strategy of choking off chip exports produced innovation workarounds rather than a decisive lead.
For Washington policymakers who bet that export controls alone could hold the line, Kimi K3 is a costly data point. The AI race is now a cost race, and Beijing is currently winning that leg of it.
