Zhongguancun Forum | Top AI influencers Yang Zhilin, Zhang Peng, Xia Lixue, and Luo Fuli discuss what?

Jing Li, a reporter for China Business Network, Beijing

On March 27, at the 2026 Zhongguancun Forum, top-tier talent from China’s large-scale model and AI infrastructure sectors gathered for a roundtable discussion.

The roundtable was hosted by Yang Zhilin, founder of Moonshot AI, with Zhang Peng, CEO of Zhipu Zhizhang; Xia Lixue, co-founder and CEO of AskQiong; Luofe Li, head of the MiMo large model at Xiaomi; and Huang Chao, assistant professor at the University of Hong Kong, doctoral supervisor, and leader of the Nanobot team, all taking part.

A reporter from China Business Network noticed that several technical talents from the AI industry didn’t exchange pleasantries during the conversation. The discussion covered everything from the industry shock triggered by the “Lobster (OpenClaw)” to the undertow of an open-source ecosystem; from anxiety over soaring inference costs to how compute power is being reshaped in the Agent era.

In this collision of ideas, they not only revealed the evolution path for large models over the next 12 months—from “chatting for fun” to “getting work done”—but also raised the question that excites and worries everyone: when Token usage sees a hundredfold explosion, are we ready?

Talking about the “Lobster”: Agent reshapes human-machine interaction

At the start of the roundtable, the discussion focused on the recently viral “Lobster.” This is not just a product update; it is also seen as a watershed moment for AI industry applications.

Regarding this wave of Agent-driven momentum, Zhang Peng believes its core lies in evolving AI from “simple conversation” into “truly getting work done.”

He pointed out that “Lobster” puts extremely high demands on model capabilities, requiring the model to be able to think and plan over long periods and handle massive context. This has also directly led Zhipu Zhizhang’s recent pricing strategy for its GLM inference models.

“Because doing complicated work means inference costs rise exponentially, and the normal logic of returning to commercial value is a necessary path for healthy development in the industry,” Zhang Peng responded.

Xia Lixue revealed that since “Lobster” took off earlier this year, the number of Tokens his company processes has nearly doubled every two weeks, reaching a cumulative 10x growth—this kind of growth speed was last seen during the 3G phone data-usage popularization period.

In Xia Lixue’s view, the past cloud infrastructure was designed for “human engineers,” whereas in the Agent era, what’s needed is infrastructure designed for “AI.” He even used a “minute-level vs. millisecond-level” comparison to describe this mismatch: humans initiate tasks at the minute level, while an Agent’s thinking and task initiation happen at the millisecond level.

“Therefore, in the future, infrastructure must be as efficient as a Token factory— and may even evolve into an intelligent agent that can self-evolve and self-iterate,” Xia Lixue predicted.

Luofe Li believes the greatest value of OpenClaw lies in “open source,” which benefits deeper community participation. It raises the ceiling of domestic second-tier closed-source models to a very high level. In most scenarios, task completion is already extremely close to the latest models, and at the same time it ensures a floor through the Skill system. Meanwhile, open source has also ignited the community’s enthusiasm for exploring Agent layers beyond large models, bringing more non-researchers into the transformation toward AGI.

Huang Chao added that with interactions similar to those of an instant messaging software, OpenClaw gives agents a stronger “human feel,” lowers the barrier to creating and using Agents, and in the future may be able to help unlock the entire tools ecosystem.

Challenges of compute and cost

As AI moves from the training era toward the inference era, model technology iteration, the return to commercial value, and compute infrastructure support have become key topics for industry development.

Yang Zhilin compared open source to “scaffolding,” saying it lowers the barrier for ordinary people to use the capabilities of top models, so that AI is no longer the exclusive domain of programmers. But he also said directly that the open-source community currently faces a huge gap in inference compute.

The “compute anxiety” mentioned repeatedly in the roundtable discussion. Zhang Peng admitted that because task complexity has increased, the number of calls required to complete one task may be 10 times—and even 100 times—that of simple Q&A, making compute a bottleneck restricting industry development.

Xia Lixue said the current problem is that the demand explosion brought by AI requires optimization of system efficiency. “We solve this by integrating software and hardware. We connect to almost all kinds of computing chips, link dozens of chips and compute clusters in China, improve conversion efficiency, and turn China into the world’s Token factory.”

He also described a grand vision—“AI Made in China,” meaning using China’s energy and manufacturing advantages to convert energy into high-quality Token outputs for the global market through efficient infrastructure.

However, existing cloud computing infrastructure constrains the development of Agents, and it’s necessary to build a smarter AI infrastructure that can adapt to Agents’ high-frequency demands. “In our view, the infrastructure itself should also be an intelligent agent—able to self-evolve and self-iterate, forming an autonomous organization,” Xia Lixue added.

Luofe Li pointed out that two years ago, when Chinese teams were constrained by compute power—especially constrained by interconnection bandwidth—they achieved breakthroughs in model architecture (such as DeepSeek).

“Although domestic chips are no longer constrained now, this exploration of high efficiency and low inference cost is still important,” Luofe Li said. “The premise for OpenClaw getting smarter the more you use it is inference Context. The real challenge now is: how do we keep costs low enough and speed fast enough in long contexts of 1M or 10M? Only by achieving efficient inference for Long Context can we support an Agent’s self-iterative capabilities in complex environments—this will be the key battleground for competition in the future.”

In addition, Luofe Li predicted that as inference demands driven by Agents explode, this year’s demand for Tokens may grow by 100 times, and the competition dimensions may go down to the compute layer, inference chips, and even the energy layer.

Huang Chao believes that in terms of Memory, Agents today still have issues: information compression isn’t accurate, and they can’t remember reliably. In long-horizon tasks and complex scenarios, Memory will surge, bringing enormous pressure. “I think in the future, Memory should move toward a layered design, making Memory more general.”

Looking ahead to the next 12 months

If you could describe the trend of large-model development in the next twelve months with one word, what would you choose?

Huang Chao’s keyword is “ecosystem.” He foresees that in the future, software may be natively designed for Agents, and the whole technology stack needs to evolve toward an “Agent Native” mode. This requires the open-source community, model developers, and application vendors to build it together, so that Agents can truly solidify from “fun new things” into reliable “digital coworkers” (Co-worker).

Luofe Li’s keyword is “self-evolution.” She believes that a powerful Agent framework activates the upper limits of large models that were previously untapped. In verifiable loops, models can already autonomously optimize goals for days, such as exploring better model architectures in scientific research. She expects that this kind of “self-evolution” capability will, within 1–2 years, raise the efficiency of creative work like scientific research by multiple orders of magnitude.

Xia Lixue’s keyword is “sustainable tokens.” In his view, AI development as a whole is still in a long-term process of continuous advancement, but a very real problem is that resources are ultimately limited.

“Can we provide tokens in a sustainable, stable, and large-scale manner, so that top models can truly serve more downstream scenarios over the long term—that’s a very critical question.” Xia Lixue said. “If we can take China’s advantages in areas like energy, transform them into high-quality Tokens through Token factories, and output them globally, then China would have a chance to become the world’s ‘Token factory.’”

Zhang Peng chose “compute power.” The Agent framework indeed unlocks many people’s creativity and can improve efficiency by 10x, but the prerequisite is that everyone can afford it—and that it’s usable.

“Two years ago, Academician Zhang Yaqin said something at the Zhongguancun Forum—roughly meaning, ‘No chip, no feelings; talking about chips makes people sad.’ I feel that today we’re kind of back to that state,” Zhang Peng said.

(Editor: Zhang Jingchao; Reviewer: Li Zhenghao; Proofread by: Yan Jingning)

A wealth of information and precise interpretation—All on the Sina Finance app