AI Industry Leaders Discuss OpenClaw, Model Evolution, and Computing Power Challenges

Industry leaders Yang Zhilin, Zhang Peng, Luo Fuli, and Huang Chao convened at the 2026 Zhongguancun Forum to discuss the evolution of AI, focusing on the OpenClaw framework and the challenges of scaling artificial intelligence. The forum, held in Haidian, featured a dense agenda including the establishment of an Open Source Alliance and the release of a Sovereign Large Model White Paper. Participants represented a broad spectrum of the AI industry, from foundational model developers to embodied AI companies.

Digital scaffolding supporting a glowing AI model, symbolizing foundational frameworks.

OpenClaw's Impact and Technical Evolution

The discussion centered on OpenClaw, a framework that enables broader access to advanced AI capabilities. Zhang Peng of Zhipu highlighted OpenClaw's role as a "scaffolding" that allows non-programmers to leverage top-tier models for complex tasks, transforming AI from a niche tool to a widely accessible resource. Xia Lixue from Wuwenshinqiong noted the shift from AI as a chatbot to an agent capable of completing extensive tasks, which has led to a tenfold increase in token usage for his company since January.

Luo Fuli of Xiaomi MiMo described OpenClaw as a "revolutionary and disruptive event" in the Agent framework, praising its open-source nature and its ability to elevate the performance of Chinese models to near parity with leading closed-source alternatives. He emphasized how OpenClaw fosters community engagement and expands the imaginative scope of AI beyond traditional research circles. Huang Chao of the University of Hong Kong pointed out OpenClaw's "human-like" interaction model, which he believes has contributed to its popularity by making AI feel more like a personal assistant. He also underscored the effectiveness of the Agent Loop framework and the potential for a lightweight operating system to integrate various tools and skills within the AI ecosystem.

Abstract graph showing sharply rising data points, symbolizing increasing AI model pricing and token usage.

Model Pricing and Inference Era Challenges

Zhang Peng addressed Zhipu's recent price increase for its GLM5 Turbo model, explaining it as a reflection of AI's transition from simple chat to complex task execution. He stated that completing tasks can consume significantly more tokens than basic Q&A, justifying the price adjustment as a return to commercial value and a move towards a sustainable industry model.

Xia Lixue discussed the implications of the "inference era" for infrastructure providers like Wuwenshinqiong. He highlighted the exponential growth in token usage, which demands greater system efficiency and optimization. Wuwenshinqiong integrates diverse computing chips and clusters to maximize resource utilization and conversion efficiency. Xia emphasized the need for "Agentic Infra capabilities" to support AI agents that can initiate tasks at millisecond speeds, a departure from traditional cloud infrastructure designed for human-initiated, minute-level tasks. He envisioned a future where infrastructure itself becomes an intelligent, self-evolving agent.

Intricate, glowing microchips and circuits symbolizing optimized computing power and efficiency.

Chinese Innovation and Agent Development

Luo Fuli elaborated on the unique advantages of Chinese teams in large model development, particularly in innovating model structures under limited computing power. He cited examples like DeepSeek's fine-grained Mixture of Experts (MoE) as breakthroughs that optimize intelligence per unit of computing power. Luo stressed the importance of long-context architectures for model self-iteration, noting that achieving low-cost, high-speed inference with millions of tokens is crucial for enabling complex, high-productivity tasks.

Huang Chao outlined key technical directions for agent development, focusing on planning, memory, and tool use. He identified challenges in planning for long-chain tasks due to models' lack of implicit knowledge in specialized domains. For memory, he noted issues with inaccurate information compression and recall, advocating for layered memory designs that can handle diverse data modalities. Huang also highlighted the emerging concept of "Agent Swarms" and the immense pressure this places on memory management. In tool use, he pointed out the persistent problems of quality assurance and security risks in skill ecosystems, calling for community collaboration to develop high-quality, secure skills.

Abstract digital tendrils and glowing data forms symbolizing an evolving AI ecosystem.

Future Outlook: Ecosystem, Self-Evolution, and Computing Power

The panelists offered one-word predictions for the next 12 months in AI. Huang Chao chose "ecosystem," emphasizing the need for collaborative development of agent-native software and data to enable agents to transition from personal assistants to integral co-workers. Luo Fuli selected "self-evolution," describing how powerful models, combined with agent frameworks, are enabling models to learn and optimize tasks autonomously. He suggested that this self-evolution could dramatically accelerate scientific research.

Xia Lixue's keyword was "sustainable tokens," reflecting the challenge of providing continuous, stable, and large-scale token availability. He envisioned China becoming a "token factory," exporting high-quality tokens globally, akin to the "Made in China" phenomenon in manufacturing. Zhang Peng concluded with "computing power," stating that it remains the most fundamental bottleneck. He highlighted the tenfold to hundredfold explosion in inference demand, which, if unmet, will hinder AI's progress and accessibility.