Chinese large language models are steadily gaining market share, primarily due to their structural cost advantages rather than price discounts. This edge is becoming more pronounced over time, according to an internet analyst at a prominent securities firm in China.
The recent launch of Kimi K3 by a Beijing-based startup on July 16 reinforced this perspective. The analyst explained that leading Chinese AI models cost about 10% of the training expenses compared to their top international counterparts, and their average API pricing is roughly 10% to 20% of the rates charged by overseas rivals, based on data from a major financial research firm.
Although Kimi K3 is on the higher price spectrum among Chinese models, it still costs only around 30% of comparable models from advanced foreign markets, the analyst added. Domestic API services for these models generate gross margins between 20% and 40%. While this is somewhat lower than the margins reported by US competitors, it remains a healthy profit margin.
A significant part of the cost difference stems from the generally smaller parameter sizes in Chinese models. However, a larger factor is innovation in algorithms and engineering. Unlike US developers that heavily prioritize performance improvements and invest heavily in cutting-edge research, Chinese companies aim to strike a balance between performance and cost-efficiency. This approach reflects the country’s limited compute resources and a broader push to make AI more accessible.
China also boasts a vibrant open-source community where top labs regularly publish research with each new model release. This collaborative environment accelerates technological advancements and fosters shared momentum, contrasting with the more closed and siloed development approach seen in the US.
The launch of Kimi K3 is viewed as a positive development that will enhance confidence in China’s open-source ecosystem rather than weaken it. The new model is seen as a beneficial addition that will support the broader industry rather than just capturing market share from competitors.
In terms of inference costs, Chinese models hold advantages across three key areas. First, their model architecture, especially with mixture-of-experts designs, enables activation of only a small portion of the parameters during inference—approximately 10% to 15%—compared to 15% to 30% in US models, reducing computing requirements.
Second, efficiencies in engineering and scheduling mean Chinese industry leaders are utilizing GPU resources more effectively—over 70% utilization—thanks to improved scheduling, compared to industry averages around 40% to 50%. Third, infrastructure costs, including electricity and data-center expenses, are lower in China, creating additional long-term cost savings, especially as more inference tasks shift to domestically produced chips.
Demand trends are also increasingly favorable for Chinese providers. Earlier this year, a trend called “token maxxing” encouraged users to consume the maximum tokens possible, but as costs mount, companies are moving toward more budget-conscious strategies that prioritize return on investment.
As enterprise needs diversify, the cost-effectiveness of Chinese models is becoming a key factor in choosing AI solutions. Some foreign companies are already beginning to transition parts of their workflows to Chinese open-source models.
Looking ahead to the second half of the year, three developments are likely to be particularly influential: first, the growth of coding-focused AI tools will help dominant developers consolidate their market edge; second, Chinese firms’ strengths in multimodal and video-generation technologies position them well as AI applications expand beyond traditional software development; third, the pricing power gap between top-tier and less advanced models may widen due to increasing market stratification.
Finally, the main obstacle to revenue growth for Chinese model providers no longer stems from demand but from limited compute capacity. Many companies lack sufficient inference hardware, and future advancements in domestic chip manufacturing will determine how much of the realized demand can be converted into revenue.