(Yicai) Sept. 3 -- Chinese large language model developers are offering application programming interface services at much lower prices than their overseas counterparts, but the discounts reflect efficiency gains rather than a loss-making strategy aimed at gaining market share, according to Xiong Wei, a China internet sector analyst at UBS Securities.

“Chinese model developers aren't subsidizing costs to drive adoption,” Xiong said at a media briefing on Sept. 1. “They’re improving profitability through technology, while maintaining relatively healthy gross margins.”

Research and development costs at some Chinese LLM developers may be less than 10 percent of those at their international peers, Xiong said, citing calculations based on publicly available data. While API pricing for mainstream domestic models averages 10 percent to 20 percent of global rates, most maintain gross margins in the 20 percent to 40 percent range, she said.

Beyond the lower costs for electricity and land in China, Xiong identified model architecture as a more critical factor, offering as an example the design of the Mixture-of-Experts architecture. The share of parameters activated when a model answers specific questions is directly linked to the computational cost of inference, with Chinese models having cut it to single digits, she said, noting that Moonshot AI’s Kimi K3, released on July 27, has a total of 2.8 trillion parameters but activates only 16 out of 896 experts during each inference run.

Because leading US AI models are mostly closed-source and their technical details are undisclosed, UBS estimates -- based on industry research -- that their activation ratios are much higher than those of Chinese models, Xiong pointed out.

The design of attention mechanisms also reduces computational power consumption during inference, not just during training, Xiong said. In addition, innovations in computational resource scheduling, combined with the inherent advantages of domestic infrastructure, make up the remaining cost differences, she stressed.

From ‘Token-Maxxing’ to ‘Token Optimization’

Earlier this year, companies encouraged high internal usage to deepen AI adoption, a practice referred to in the industry as “token-maxxing.” This proved not to be an optimal strategy, according to Xiong. On the one hand, AI bills remain high and sometimes exceed budgets, and on the other, the economic value derived from such token usage is difficult to quantify.

By mid-year, the conversation among companies, developers, and users had shifted toward “token optimization,” seeking a balance between absolute performance and cost-effectiveness, Xiong said, presenting two sets of data that point to this timeframe.

Two sets of data she presented point to the same turning point. The industry-average index for model token costs halted its steep rise during May and June, while usage growth rates diverged between open-source models (represented by China) and closed-source ones (represented by the United States) on third-party model aggregation platforms, she said. This indicates that businesses have begun factoring cost-effectiveness into their model procurement decisions, with the global model market likely to become more stratified, Xiong said.

The most valuable, sensitive tasks with the highest quality requirements will still need top global models, but "in a company's real working environment, there’s plenty of relatively lower-value, lower-risk, and highly repetitive work” that can be handled by more cost-effective versions, she pointed out.

Some leading US platforms and app developers have begun to integrate Chinese open-source models into their real workflows, Xiong said. This is still in the very early stages, but is enough to support UBS's optimistic view of Chinese AI models gaining a larger global market share, she added.

“In the past, the metric for measuring model competitiveness were relatively simple, focusing strictly on absolute model capability," she noted. But as leading models become more expensive and the industry’s average capabilities improve, companies have started incorporating more metrics, with cost-effectiveness being the first to be added, she said.

In light of this change, “we believe that Chinese models will demonstrate stronger competitiveness in the global AI market than before,” Xiong said.

Editor: Martin Kadiev