90% cheaper, U.S. enterprises are secretly switching to Chinese large models

In July, the plunge of U.S. chip stocks rapidly spread to capital markets including South Korea and Japan. Large deals worth nearly a trillion dollars signed by South Korea's Samsung Electronics, SK Group and U.S. companies such as NVIDIA failed to boost the market, and U.S. and South Korean chip stocks have been hit by panic selling over the past week.

This round of plunge left AI investors, who were still celebrating enthusiastically just last month, dead silent.

From Wall Street banks such as Goldman Sachs to Michael Burry, the investor famous for *The Big Short*, people have warned for months that the transactions in the artificial intelligence industrial chain are of a "circular" nature. NVIDIA provides funding and holds shares for the upstream and downstream of its chips. Investors worry that such "circular" transactions may distort the market.

According to research from Goldman Sachs, the market value of AI-related companies has increased by about 27 trillion US dollars since November 2022. The International Monetary Fund warned that if the profitability of artificial intelligence fails to meet expectations, the market may shrink.

From AlphaGo's dominance in the Go world to the launch of ChatGPT, the United States has accumulated significant first-mover advantages in the field of large algorithm models. However, the rise of Chinese large models in recent years has impacted the AI landscape.

Multiple media outlets including the U.S. *Wall Street Journal* mentioned that Kimi K3, the new generation open-source large model from Chinese large model unicorn Moonshot AI, has shaken the capital market, and its impact is just like the market panic caused by DeepSeek when it was released in 2025, which even Elon Musk praised as "impressive".

Amid the praise, threats are creeping in.

On July 22, Michael Kratsios, Director of the White House Office of Science and Technology Policy, stated that Moonshot AI distilled U.S. artificial intelligence company Anthropic's technology to develop Kimi K3. U.S. Treasury Secretary Bessant threatened to use sanctions and the "Entity List" to respond to the distillation practices of Chinese enterprises.

"The United States should not ban such large models," said Jensen Huang, CEO of NVIDIA. He noted that Chinese large models are very excellent, and artificial intelligence technology will penetrate into more industries.

A number of U.S. technology enterprises and institutions issued a joint statement on July 24, supporting open-weight artificial intelligence models. This is regarded as a response to the attempt of some people in the U.S. government to restrict open-weight models.

In fact, if Chinese artificial intelligence enterprises are sanctioned, U.S. companies will not have an easy time either.

According to the data platform OpenRouter, the overall usage of Chinese AI models has exceeded that of similar U.S. products this year.

"As the cost of artificial intelligence soars, Chinese artificial intelligence models are highly attractive to U.S. companies," said Kyle Chan, a researcher at the John Thornton China Center of the Brookings Institution. U.S. companies are now paying more attention to cost control.

According to the U.S. Consumer News and Business Channel (CNBC) citing expert opinions, open-source Chinese models are 60% to 90% cheaper than Anthropic and OpenAI models, but their performance is close to that of top-tier models.

In June, artificial intelligence startup Lindy moved its business from Anthropic to DeepSeek. Flo Crivello, CEO of the company, said that the cost curve has dropped off a cliff, and the company will save millions of dollars in a few months.

"More than 90% of us do not need the large model from Anthropic," said Meinhold, founder of the digital heritage platform LilyList. The company's operations only require models that are sufficient for its needs.

Yacine Jernite, Head of Machine Learning at Hugging Face, said that more and more companies tend to choose more affordable and self-controllable AI technologies, which usually adopt Chinese solutions.

Some large U.S. companies have also begun to turn to Chinese large models.

According to the Associated Press, cryptocurrency exchange Coinbase said it is turning to Chinese artificial intelligence models to reduce costs. Airbnb has adopted Alibaba's Qwen model, praising it as "fast and cheap".

After the threats against Chinese artificial intelligence enterprises were issued, the Silicon Valley startup circle fell into panic. Nearly 200 U.S. technology startups urged the U.S. government not to cut off access to Chinese open-source models, arguing that this would weaken the competitiveness of U.S. enterprises.

"Hundreds of companies will go bankrupt instantly," said Sohail Doshi, founder of Particle, an artificial intelligence infrastructure startup. Banning Chinese AI models will force small businesses to turn to the expensive models from Anthropic.

At present, the U.S. artificial intelligence industry is haunted by concerns about the "AI bubble". The U.S. *Time* magazine pointed out bluntly that there is a fundamental mismatch between trillions of dollars of investment in developing artificial intelligence infrastructure and the billions of dollars of artificial intelligence spending by users.

At this critical decision-making juncture, the U.S. government has more options available.

A spokesperson for China's Ministry of Commerce recently stated that it hopes the U.S. side will work in the same direction as China, strengthen artificial intelligence dialogue and communication, jointly promote the sound and positive development of artificial intelligence, and benefit the two countries and the world by advancing and improving artificial intelligence governance.

The views in this article are for reference only and do not constitute investment advice. Investment is risky, so be cautious when entering the market.

This article is from the WeChat official account "China News Network (ID: jwview)", written by Chen Junming, edited by Li Xiaoxuan, with Jiang Ying and Jia Yifu as the responsible editors, and published by 36Kr with authorization.