Token tenfold in two months: The computing power anxiety and pricing logic at the OpenClaw roundtable at Zhongguancun Forum.

According to 1M AI News monitoring, Yang Zhilin, the founder of the Dark Side of the Moon, hosted a roundtable on OpenClaw and AI open source themes at the 2026 Zhongguancun Forum. The guests included Zhang Peng, CEO of Zhicheng Huazhang, Xia Lixue, co-founder and CEO of AI computing power infrastructure company Wuwen Xinqiong, Luo Fuli, head of Xiaomi’s MiMo large model, and Huang Chao, assistant professor at the University of Hong Kong, covering three aspects: models, computing power infrastructure, and Agent applications.

Xia Lixue provided an intuitive data point: since the end of January, the token usage of Wuwen Xinqiong has doubled every two weeks, with a cumulative increase of ten times, “The last time I saw this speed was during the 3G era’s mobile data.” Zhang Peng explained the logic behind the recent price increase of Zhicheng GLM5 Turbo from the perspective of model manufacturers: Agents are no longer one question one answer; the amount of tokens consumed to complete a task may be ten or even a hundred times that of answering simple questions, involving long-term task planning, continuous debugging, and multimodal information processing. “Relying on low-price competition in the long term is not beneficial for the development of the entire industry.”

Luo Fuli believes that the core bottleneck of the Agent era is the reasoning cost and speed of long-context tasks. Only by achieving low costs and fast speeds at the scale of millions or even tens of millions of tokens can truly high productivity value tasks be assigned to the models. She revealed that the model’s “self-evolution” has moved from concept to practice. In clearly defined scientific research tasks, the model can now operate autonomously for two to three days, and the team’s own research efficiency has thus accelerated nearly tenfold. She also raised a question: the demand for reasoning has increased nearly tenfold in recent times; will the overall token growth this year reach a hundredfold?

Huang Chao dissected three technical bottlenecks from the Agent application layer: insufficient planning capacity for long-chain tasks, memory inflation pressure caused by multi-Agent collaboration, and the risks of low-quality tools and malicious injections in the skill ecosystem. He pointed out that the memory management of existing frameworks is still at the level of file systems and Markdown formats, and in the future, it needs to move towards a hierarchical design.

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

1. Huang Chao: “Ecosystem,” software will shift from human-oriented to Agent Native design.
2. Luo Fuli: “Self-evolution,” stating this is “the only place that can create new things.”
3. Xia Lixue: “Sustainable tokens,” hoping China will become the world’s token factory.
4. Zhang Peng: “Computing power,” indicating that behind the tenfold growth “there is still a hundredfold demand that has not been met.”