CNBC’s The China Connection newsletter: For Chinese businesses, it’s not about which AI is the smartest

This report is from this week’s CNBC’s The China Connection newsletter, which brings you insights and analysis on what’s driving the world’s second-largest economy. You can subscribe here.

Will the U.S. or China win the artificial intelligence race? That’s the big question for investors wondering where they should put their money.

But frequently, I’ve found that many companies in China are asking a different question: Which AI tools can help me survive in a tough economy?

All this means that investors who choose an AI winner based on smarts alone likely overlook a bigger story.

First, there’s the cost factor.

While OpenAI’s gpt-oss open-weight model and DeepSeek are the cheapest AI models to operate – 30 cents per 1 million tokens – Google’s Gemini 3 Pro charges $4.50, while Anthropic’s top-tier Claude Opus 4.5 costs $10, according to Artificial Analysis, which benchmarks AI models.

That’s a pay-per-use structure that can quickly balloon in price — whether businesses are using a “closed” AI model from OpenAI or remotely accessing an open-source model from DeepSeek or Alibaba.

Instead, many companies in China prefer to control costs by choosing a third option — downloading open-source models, James Tong, CEO of Suzhou-based Movitech, told me. His enterprise software company’s clients include Starbucks and Unilever.

Another key difference with the U.S. is how Chinese businesses are using AI.

Companies in the U.S. tend to use AI more for coding, while in China, they use the technology more for specific situations, according to Tong. For example, since the second half of 2025, Movitech has used AI agents to help manufacturers improve their production processes and assist state-owned enterprises with compliance.

Locally run open-source models again fit better because they don’t require users to upload information to third-party cloud servers.

It’s a bonus productivity tool when domestic demand has been sluggish.

Tong said he’s had to lower product prices due to economic pressures, but AI has boosted internal efficiency. Without those gains, he said, the business would be in a far weaker position.

Industry rankings, such as Artificial Analysis, show that the gold standard in generative AI models still belongs to the U.S. But those tools aren’t officially available in China.

Beijing’s firewall keeps out Google and Facebook and their AI models, and Anthropic cited national security concerns in its ban on China-owned entities from using Claude, widely considered the best AI for coding right now.

Yet discussion of leading U.S. models is prevalent enough on Chinese social media to signal that such technologies are being used by local developers. And my conversations with various businesses confirm that they’re trying different AI models, including U.S ones, to find one that suits their needs.

“When selecting AI solutions, Chinese companies no longer prioritize ‘who is the smartest,'” said Sun Xin, vice president of research at Gartner. Instead, they’re looking for reliability, control over hallucinations and integration capacity.

Productivity gains, in particular, are critical.

Thanks to AI, 10 people can complete the same amount of work as 20 to 30 people, said Wenhao Zhang, CEO of Beijing-based consumer marketing consultancy Doodod Technology.

As a result, the company has roughly the number of its clients in the last year to well over 100 customers.

Low prices and ease of access also mean China-made models are getting more overseas users, especially among those who are jumping in to try out the viral open-source AI agent OpenClaw.

Chinese models also rank highly on OpenRouter, a popular platform that lets companies access multiple AI models through one portal.

While U.S. models from Anthropic and Google were still the top two by usage in the last week, one from DeepSeek ranked third, while Moonshot AI’s Kimi’s K2.5, also from China, came in fourth — just days after its launch.

Moonshot said that since the release of its latest K2.5 model, its revenue from outside China has surpassed that from the domestic market, as new overseas users quadrupled.

Reflecting how investors want a piece of the pie, CNBC reported last month that Moonshot’s valuation surged by at least $500 million in just a few weeks — after its rivals Zhipu and Minimax listed in Hong Kong.

As the Lunar New Year approaches in China, Chinese businesses and tech developers anticipate that DeepSeek may release another breakthrough AI update. Just before the holiday last year, the Chinese company released an open-source model with usage costs far below OpenAI’s ChatGPT.

DeepSeek’s release at the time shocked global investors, who thought China could not innovate on AI without high-end Nvidia chips. Local companies, previously locked out of leading U.S. AI models, started trying out DeepSeek and similar open-source models developed by Alibaba, ByteDance and local startups.

“As long as the private sector of China continues to have the leeway to make money, China’s innovation can be kept up with the U.S., then people are going to be interested in Chinese equities,” said Liqian Ren, director of Modern Alpha at U.S.-based fund manager WisdomTree.

As Chinese AI models try to catch up to their U.S. rivals, innovation isn’t just about becoming more intelligent — it’s also about hitting a price point that makes them usable for business.

Damien Ma, director of Carnegie China, said we are now living in an “à la carte interest-based world” rather than a values-based one, and that he doubts there will be a new “golden age” of UK-China and US-China economic relations.

Zhiwei Zhang, president and chief economist of Pinpoint Asset Management, said markets have been expecting policy changes in the property market since Q4 2025, but questions remain about whether the reported relaxation of borrowing rules for developers is enough to shift the sector’s momentum.

China stocks were little changed on Wednesday amid mixed trading in the region, but software stocks took a hit as investors weighed how AI is threatening to automate workflows, squeeze pricing and lower the barriers for competitors to enter the market.

Hong Kong’s Hang Seng index and CSI 300 were little changed as of 12:45 a.m. ET. Year to date, both indexes are up roughly 4.6% and 0.7% respectively.

However, the Hang Seng Tech Index was down roughly 1.8%. The offshore yuan was little changed at 6.933 against the greenback.

Chinese software companies saw a selloff, amongst other regional peers. Shares of China’s Kingdee International Software plunged more than15%, while cloud major Tencent fell 3.27%. Alibaba lost over 1%, while Baidu was down over 2%.

Feb. 7:  Uruguay President Yamandú Orsi concludes a nearly week-long state visit