AI Trends

Chinese Large Models Must Prove Their Commercial Value

EastmoneyUpdated Nov 19 05:48

AI summary

Usage of Chinese large models is growing rapidly, but this has yet to translate into corresponding commercial value. Their technical capabilities and industrial adoption must be converted into stable revenue and measurable delivery value. OpenRouter data show that from November 9 to 15, Chinese large models processed 119.1 trillion model tokens, accounting for about 70% of all model-token usage on the platform, yet their share of paid spending was only around 30%. The article recommends continuing to reduce inference costs, focusing on high-value use cases, and improving ecosystem infrastructure such as tools, deployment solutions, and talent.

Why read this

AI score 63

Starting with the gap between usage volume and the share of paid spending on OpenRouter, the article provides an analytical framework for evaluating Chinese large models’ shift from usage scale to commercial value.

Topics

Tags

Body · AI translation

Article Body

■ Yuan Chuanxi

The latest data released by OpenRouter show that from November 9 to 15, Chinese large models on the rankings recorded 119.1 trillion tokens in weekly usage, up 2.51% from the previous week; during the same period, U.S. large models recorded 14.83 trillion tokens in weekly usage. Moreover, all four of the world’s most-used models that week came from China.

China’s homegrown large models are now accelerating the formation of a virtuous cycle in which technology, ecosystems, and application scenarios reinforce one another. Models such as DeepSeek, Tencent Hunyuan, Zhipu GLM, and Alibaba Cloud’s Qwen continue to evolve, achieving breakthroughs in coding, long-context processing, agents, multimodal capabilities, and other areas. Meanwhile, the open-source ecosystem is expanding rapidly, with open weights for Alibaba Cloud’s Qwen series, Zhipu’s GLM series, and other models being released continuously. China has become one of the world’s most active suppliers of open-source large models.

More importantly, the continued rise in usage shows that homegrown models are no longer confined to demonstrations and trials but are steadily entering real-world settings such as industry, finance, energy, and manufacturing. For example, Zhongguancun Science City recently released 67 representative cases showcasing AI empowerment across a broad range of fields, while Guangzhou now has a cumulative total of 114 large models registered with the Cyberspace Administration of China. The more numerous the application scenarios, the richer the corpora and tasks available for model iteration, accelerating the shift from technology that is merely usable to technology that works well for industry.

However, leading in usage does not equate to leading in commercial value, and the contradiction between “high usage and low revenue” remains unresolved. On the OpenRouter platform, Chinese models account for about 70% of token usage but only around 30% of paid spending. Their lead in usage has yet to translate into a lead in commercial value. This disparity is not difficult to understand: much of the platform’s usage comes from low-cost plans, free trials, and developer testing, where the amount paid per call is limited; enterprise orders that truly pay according to value have yet to expand at the same pace.

For precisely this reason, China’s homegrown large models must quickly answer the “value question”: converting high usage into stable revenue, technological capabilities into measurable delivery value, and industrial adoption into sustainable returns on investment.

Going forward, efforts can focus on three areas. First, inference costs should continue to be reduced, using supply-side technological advances to cut costs and create the conditions for a virtuous cycle in business models. Second, companies should cultivate high-value application scenarios. Customers are purchasing not only inference tokens but also assurances of code correctness, traceable logic, controllable compliance, and stable operation during long-running tasks. Third, ecosystem infrastructure should be improved. The more abundant the tools, deployment solutions, and talent pools surrounding models become, the lower the onboarding costs for new users will be and the smoother the conversion of usage into revenue.

Leading in usage is only the starting point. China’s homegrown large models have already demonstrated their technological strength and ecosystem vitality; the next step is to advance more steadily and further along the path toward commercialization. By remaining focused on the real economy, real-world application scenarios, and sustainable value, China’s homegrown large models can turn their usage lead into outstanding value-creation capabilities and provide an enduring source of momentum for China’s technological self-reliance and high-quality development.

(Source: Securities Daily) Mid-article embedded advertisement

Source

Eastmoney