"Agents have gone crazy — how can China's token economics be sustained?" Yang Zhilin grills AI leaders on OpenClaw, tokens and the Agent era

OpenClaw lightning and the token shock

A high‑profile roundtable at the Zhongguancun Forum saw Yang Zhilin (杨植麟) moderate a wide‑ranging discussion with Luo Fuli (罗福莉), head of Xiaomi's MiMo large‑model effort; Xia Lixue (夏立雪), CEO of Wuwen Xinqun (无问芯穹); Zhang Peng (张鹏), CEO of Zhipu Huazhang (智谱华章); and Huang Chao (黄超), assistant professor at the University of Hong Kong (香港大学). The group homed in on one disruptive catalyst: OpenClaw — an open, IM‑embedded Agent framework that speakers said is changing how people interact with and extend models. It has been reported that Wuwen Xinqun's token calls have doubled every two weeks since late January and are now roughly ten times higher — a surge Xia likened to the early 3G mobile‑data boom.

From harnesses to 100M‑token context — technological levers

Panelists argued that Agent frameworks plus “harness” and skills systems are unlocking latent model capabilities. Luo said OpenClaw raises the ceiling for many domestic open models, pushing some toward the performance of closed systems like Anthropic’s Claude Code. Huang emphasised the shift in UX — from tool feel to “personal AI” — while speakers outlined the technical necessities for sustainable self‑improving agents: much longer context windows (10M–100M tokens), hybrid sparse/linear attention architectures (DSSA, KSA, Xiaomi’s “Highest Bus”), and complementary memory/harness layers to reduce planning failures. But long context and richer skills mean far higher inference cost. Zhang explained that GLM’s recent price increase reflects this reality: complex planning tasks can consume ten to a hundred times more tokens than simple Q&A, so unit pricing must realign with commercial cost.

Economics, security and geopolitics: can China build a self‑sustaining stack?

If agents are driving explosive token demand, how will China's ecosystem pay for it? The panel suggested a mix of technical optimisation, higher prices and community governance for skills quality — but also flagged risks: low‑quality or malicious skills can break task completion and safety, and the heavy compute appetite amplifies infrastructure bottlenecks. This comes against a geopolitical backdrop: US export controls and trade policy constrain access to the most advanced chips, reportedly accelerating domestic efforts to optimise software‑and‑architecture stacks and push local silicon. So the question is not just technical. Who pays for persistent, agent‑native software — and can a token‑based economy scale without careful market design, better tooling and hardware policy support?

The roundtable concluded on an open note. Agents are no longer a research curiosity; they are a demand engine reshaping UX, models and compute economics. The immediate answers will be technical and community driven — long contexts, smarter harnesses, quality‑controlled skills — but sustainable growth will also require clearer commercial models and policy attention to the compute supply chain.