US companies turn to Chinese AI models as costs surge

US companies are increasingly turning to Chinese-built artificial intelligence models for better value for money, as token prices for the most advanced American models continue to climb this year.

Recent releases from Chinese companies, including DeepSeek and Z.ai, are seen by many in the industry as highly competitive with leading frontier systems from Anthropic and OpenAI, narrowing the performance gap while remaining significantly cheaper to use.

Models such as Z.ai's GLM and Alibaba's Qwen have become attractive options, offering a combination of performance and cost. Z.ai's GLM-5.2, released in June, narrowly trailed leading US models on public benchmarks and ranked high on several leaderboards while costing dramatically less than comparable closed US frontier models.

OpenRouter, a platform that lets developers access a range of AI models, found Chinese AI model usage by US companies has risen significantly this year. The share of tokens used by US companies on Chinese AI models has sat above 30 percent every week since Feb 8, at times rising as high as 46 percent, CNBC reported, citing the platform's figures. That compares with an average of just 11 percent over the previous 12 months, and a mere 4.5 percent in the first half of 2025.

In June, the AI startup Lindy moved 100 percent of its traffic from Anthropic's Claude models to DeepSeek. "We did it, and you could see that cost curve go down, like, crash to the ground," Lindy CEO Flo Crivello told CNBC, adding that the move would save the company millions of dollars within months.

"Cost has to be part of the equation, but it's common in technology that cost per unit changes dramatically with time and volume. That was true of semiconductors, electronics and pretty much everything else," Ker Gibbs, former president of the American Chamber of Commerce in Shanghai, told China Daily.

"The first semiconductor coming out of a fab was enormously expensive to make, but once the fixed costs were absorbed every other chip was not expensive at all. What is new is the speed and scale at which the price differential has emerged in AI," said Gibbs, who teaches at the University of San Francisco and is the author of The Fragile Dragon.

He noted that the savings coming out of Chinese models are rooted less in fixed-cost recovery than in architectural efficiency, though the underlying logic is the same one that Chinese firms have applied across other industries.

"China's business people, as I write about in The Fragile Dragon, are pragmatic. They build products that are not just good – they are good enough. That means they don't overbuild or overdesign. Cost is always a factor," he said. "This is a familiar playbook applied to a new domain: achieve near-parity on performance, undercut dramatically on price, capture volume and let the market do the rest."

Rising costs

The Ramp AI Index, which tracks how American businesses use AI, found that the companies most committed to the technology are spending around $7,500 per employee every month on it. The index also found that DeepSeek is one of the fastest-growing vendors on Ramp's platform, which is evidence, the firm said, that companies are increasingly opting for cheaper AI models.

Palo Alto Networks CEO Nikesh Arora told CNBC that token costs need to drop by as much as 20 percent over the next 12 months and 90 percent by the following year, to enable large-scale AI adoption. Rising token costs have become a major pain point for businesses and a strain on AI budgets, he said, with current pricing making the tools increasingly difficult to implement.

Asked why companies are paying more attention to the cost of AI models now, Gibbs said, "In short, the check arrived. The bills companies are paying for AI usage have come due and suddenly realize what they have been paying," he said. "In some sense we're reaching the end of the beginning, which could be characterized as a bit of irrational exuberance and unrealistic expectations about the results AI would deliver."

In the earliest days of the boom, he said, companies raced to deploy AI as fast as possible simply because competitors might be doing the same. "The costs didn't matter, and the CFO was left out of the room. The fear of getting left behind was stronger than any concerns about costs," he said. "Now that the bills are due, questions are being asked about the results and how to measure them."

Per-token prices have actually fallen, he noted, but usage has climbed even faster and gone largely unrestrained. The productivity gains are real in specific applications such as fraud detection, coding, customer service and supply chain optimization, Gibbs said, but they can be difficult to measure and are far from uniform across the board.

Chinese efficiency

As companies look to deploy AI to build new products and find internal efficiencies, engineers are increasingly experimenting with cheaper open-source and open-weight models, the most capable of which are now made by Chinese companies.

Open-source and open-weight models make different parts of a system available for developers to inspect, use and sometimes modify, in contrast to closed proprietary systems such as those from OpenAI, Anthropic and Google.

The performance of Chinese models has been rising in step with their growing popularity. Stanford's 2026 AI Index report found that the performance gap between US and Chinese AI models has effectively closed, with the two sides trading the lead multiple times since early 2025, starting when DeepSeek released its R1 model, which matched the top US model at the time.

The cost advantage is showing up in how companies allocate work between models. DoorDash Co-founder and Chief Technology Officer Andy Fang said in a post on the social media platform X on July 6 that the company was saving significant money by routing "lower-level work" to Kimi, a model built by the Chinese startup Moonshot AI.

For enterprise customers, Gibbs said, the rational response is the hybrid model — use a cheaper tool for the heavy lifting and more routine tasks, saving budget for the more expensive tools when they are required.

"People need to understand that China is now a source of new ideas and innovations, whether it's in pharma, automotive or artificial intelligence," he said. "It's not an accident that you can walk across the Google campus in Silicon Valley and hear Chinese spoken as well as English and many other languages. Ideas don't have borders."

That does not mean the calculation is purely technical, he cautioned. "Businesses need to be aware of the geopolitical implications as they build out their tech infrastructure," Gibbs said. "The US and China are in a very tense period, and while we hope a better path forward can be found for cooperation and peaceful coexistence, we're not there yet."