Companies Turn to Low-Cost Chinese AI Models Amid Skyrocketing Costs

The snowballing costs of artificial intelligence (AI) have emerged as a major hurdle for corporate AI transformation.
Reuters reported on June 29 (local time) that companies, which had been encouraged to adopt AI tools to boost productivity, are now facing massive bills and are turning to cheaper alternatives, including Chinese AI models, after internal deliberations.
According to Reuters, most enterprise AI agents (work assistants) use a pay-as-you-go structure where fees increase in proportion to usage, measured in units called AI tokens.
The problem is that many companies, in a race to boost productivity, have engaged in tokenmaxxing—significantly increasing their internal use of AI agents. As a result, these usage-based costs have soared uncontrollably, putting pressure on corporate financial structures.
While major AI agent developers like OpenAI and Anthropic continue to lower costs per token, another significant issue is that the rapid advancement of AI tools has expanded their range of applications, making it extremely difficult for companies to predict their AI usage and expenses.
For instance, the ride-hailing platform Uber had to implement measures to limit AI token consumption after its AI coding tool usage surged, exhausting its annual AI budget in just four months.
According to the consulting firm Gartner, costs generated by AI coding tools are projected to exceed the annual salary of an average human developer by 2028.
Other survey results conducted by Gartner among businesses show that three-quarters of companies expect an increase in their tech budgets this year.
In particular, nearly half of these companies anticipate double-digit growth rates.
Consequently, companies are rapidly turning to more affordable options, such as Chinese open-source AI.
Routing services like OpenRouter have also gained popularity.
A routing service acts as an AI marketplace, where simple tasks are selectively assigned to cost-effective AI models, while high-difficulty tasks are directed to premium AI tools like Claude, effectively reducing overall costs.
According to an analysis by Citigroup, the processing ratio of tokens for open-source AI on OpenRouter jumped from 34% in January to 65% this month.
This indicates a sharp surge in the usage of open-source AI.
The four most popular AI models on OpenRouter are all made in China, with the top spot held by DeepSeek, a leading high-efficiency AI from China.
The cost of these Chinese AI models is as low as 18 cents per million tokens, which is less than 5% of the $4 average price for premium U.S. AI models.
Harold Byun, CEO of AI startup BlueRock, said in an interview with Reuters, "In the past, open-source models lagged behind major AI by more than a year, but now that gap has narrowed to about four months," adding that "the gap will shrink even faster in the future."
Val Bercovici, Chief AI Officer (CAIO) at Weka, an AI infrastructure efficiency company, pointed out, "Recent open-source models deliver 90% of the performance of big tech products at just 10% of the cost," adding that "there is no need to use expensive premium models for every step of a task."
However, a downside is that companies may be reluctant to adopt Chinese open-source AI on a large scale due to concerns over security risks or data leaks.
For this reason, a multi-platform strategy is being discussed as an alternative in the industry.
Analysts suggest that a flexible hybrid approach may spread, where tasks are categorized so that low-cost AI is used for non-core, simple tasks, while complex, core tasks are entrusted to advanced AI.
(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.