Token Surges Tenfold in Two Months: Mining Power Anxiety and Pricing Logic on Zhongguancun Forum's OpenClaw Roundtable

According to 1M AI News monitoring, Yang Zhilin, the founder of the Moon's Dark Side, hosted an OpenClaw and AI Open Source-themed roundtable at the 2026 Zhongguancun Forum. The guests included Zhang Peng, CEO of Intelligent Spectrum Huazhang; Xia Lixue, Co-Founder and CEO of InquireChip's AI Computing Infrastructure; Ruo Fuli, Head of Xiaomi's MiMo large model; and Huang Chao, Assistant Professor at the University of Hong Kong, covering the three aspects of models, computing infrastructure, and Agent applications.

Xia Lixue provided an intuitive data point: since the end of January, InquireChip's token usage has doubled every two weeks, accumulating a tenfold increase. "The last time I saw this speed was during the 3G era of mobile data." Zhang Peng, from the perspective of a model vendor, explained the recent logic behind Intelligent Spectrum's GLM5 Turbo price increase: The Agent is no longer just question and answer; completing a task may consume ten to a hundred times more tokens than answering simple questions. This involves long-range task planning, continuous debugging, and multi-modal information processing. "Relying on low-price competition in the long term is not conducive to the overall industry development."

Ruo Fuli believes that the core bottleneck of the Agent era is the cost and speed of long-context reasoning. Only when achieving a low enough cost and a fast enough speed at the scale of millions or even tens of millions of tokens in context can truly high-productivity tasks be assigned to models. She revealed that the concept of model "self-evolution" has moved from theory to practice. In clear scientific research tasks, models can now run autonomously for two to three days, accelerating the team's research efficiency by nearly tenfold. She also posed a question: In the past period, the demand for inference has grown nearly tenfold; will the overall token growth this year reach a hundredfold?

Huang Chao broke down three technical bottlenecks at the Agent application layer: inadequate planning ability for long-link tasks, memory expansion pressure brought by multi-Agent collaboration, and the risk of low-quality tools and malicious injection in the skill ecosystem. He pointed out that the memory management of existing frameworks still remains at the file system and Markdown level, and the future needs to move towards a layered design.

At the end of the roundtable, each of the four guests summarized the trends for the next 12 months in one word:

1. Huang Chao: "Ecology," where software will shift from being human-centric to Agent Native design
2. Ruo Fuli: "Self-evolution," calling it the "only place that can create new things"
3. Xia Lixue: "Sustainable token," hoping that China will become the world's token factory
4. Zhang Peng: "Computing power," stating that behind the tenfold growth, "there is still demand for a hundredfold that has not been met"