On March 27 at the ZGC Forum, Xia Lixue, Co-founder and CEO of Infinigence, joined Yang Zhilin (moderator), CEO of Moonshot AI; Zhang Peng, CEO of Zhipu; Luo Fuli, Head of Xiaomi’s MiMo LLM team; and Huang Chao, Assistant Professor at the University of Hong Kong. Together, they engaged in a deep discussion on the future of open-source large language models and AI agents.
Representing the infrastructure layer, Xia Lixue shared his perspectives on the current wave of AI transformation, focusing on token supply dynamics and the long-term sustainability of the industry. Key points include:
· From “conversation” to “getting work done,” AI agents have driven a more than tenfold surge in token demand. The growth curve resembles the early breakout phase of mobile data traffic in the 3G era, signaling both the start of rapid industry expansion and a broader structural shift.
· Infrastructure must evolve for the agent era. Future “token factories” will need to be not only more efficient, but also significantly more intelligent.
· A China-specific “token economics” framework should establish a full chain linking energy, tokens, and GDP, enabling a sustainable token supply system that supports "AI-Made in China".
The following is a transcript of the roundtable discussion, excerpted from Intelligent Emergence:
Yang Zhilin: When you use OpenClaw or similar products in your daily life, what do you find most imaginative or impressive? From a technical perspective, how do you view OpenClaw today and the evolution of related agents?
Zhang Peng: I started experimenting with OpenClaw quite early—back when it wasn’t even called OpenClaw, but Clawdbot. As someone with a programming background, I’ve had quite a few hands-on experiences with it.
The biggest breakthrough—or rather, the most interesting shift—is that this is no longer something exclusive to programmers or hardcore geeks. Today, even non-technical users can relatively easily access the capabilities of state-of-the-art models, especially in coding and agent-related tasks.
That’s why I tend to describe OpenClaw as a kind of “scaffold.” It provides a layer on top of foundation models—a framework that is both robust and convenient, yet flexible enough for users to assemble and extend capabilities as they need. People can freely tap into underlying models and the new capabilities they enable.
In the past, many people had ideas but couldn’t execute them simply because they didn’t know how to code. Now, those ideas can often be realized through natural language interaction alone. That shift has had a profound impact on me—it has fundamentally reshaped how I think about what’s possible.
Xia Lixue: At first, I wasn’t quite used to OpenClaw, because I was more accustomed to chatting with large language models. Compared to that, OpenClaw felt noticeably slower in response.
But later I realized a key difference: unlike traditional chatbots, OpenClaw is designed to help you complete large, complex tasks. Once I started assigning it more substantial workloads, it actually performed very well.
That really struck me. The shift from models that simply generate token-by-token responses to agent-based systems—what people jokingly call a “Lobster,” a DIY AI agent-building system—that can actually complete tasks has significantly expanded the imagination space of AI.
At the same time, this also places much higher demands on system capabilities, which explains why I initially found OpenClaw somewhat sluggish.
As an infrastructure provider, I see that OpenClaw brings both more opportunities and challenges to the entire large-scale AI systems and the broader ecosystem. The resources we can currently mobilize are still not sufficient to fully support such rapid growth.
For example, since late January, our company’s token usage has roughly doubled every two weeks, and has now increased by about tenfold.
The last time I saw this kind of growth curve was during the 3G era, when mobile data traffic first took off. Today’s token consumption feels similar to when we were only using around 100 MB of mobile data per month.
In this context, we need to better optimize and integrate all our resources.
And this is not just about AI—it’s about society as a whole. Eventually, every individual will be able to use systems like OpenClaw.
So from the perspective of someone working on infrastructure, I feel both excited and deeply reflective. There is still a vast amount of optimization space ahead of us, and it is something truly worth exploring.
Luo Fuli: OpenClaw has been a genuinely transformative development for the agent ecosystem.
While most of the people around me who do serious coding still default to Claude Code, I believe that anyone who has actually spent time using OpenClaw can appreciate how far ahead it is in terms of agent framework design. In fact, many of Claude Code's recent updates appear to be moving in a similar direction.
My experience with OpenClaw is that the possibilities it unlocks keep expanding.
OpenClaw delivers value on two fronts. First, it's open source. Open source is Openness is a critical prerequisite for deep community participation and sustained investment in the agent ecosystem. A major advantage of agent frameworks such as OpenClaw and Claude Code is their ability to significantly raise the performance ceiling of models that, while still trailing the leading closed-source models, are already competitive in their own right. In most scenarios, these models can achieve task completion rates that are very close to those of Claude’s latest models. At the same time, features such as the Harness system, Cache-to-Cache mechanisms, and the Skills framework help establish a strong baseline for performance, ensuring both task completion and accuracy.
So from the perspective of foundational LLMs, OpenClaw not only raises the performance ceiling but also provides a strong performance floor.
Beyond that, OpenClaw's broader contribution to the community is that it has inspired people to see how much room there is for innovation at the agent layer, beyond the underlying models themselves.
This is also why I've noticed more people—not just researchers—becoming involved in this wave of AGI development. By leveraging more capable agent frameworks, such as Harness and Scaffolds, people are able to automate parts of their work and free up more time to focus on higher-value and more creative tasks.
Huang Chao: OpenClaw's rapid rise can be understood from two perspectives.
The first is the interaction model. We've been building agents for the past couple of years, but tools such as Cursor and Claude Code have largely felt like tools in the traditional sense. OpenClaw is the first to adopt an IM-style interface, making interactions feel much more natural and human. In many ways, it comes closer to the personal Jarvis experience that people have long imagined.
The second is the inspiration it provides at both the architecture and ecosystem levels. On the one hand, its simple and efficient Agent Loop architecture once again demonstrates the effectiveness of this approach.
On the other hand, it prompts us to rethink a fundamental question: what do we actually need—a fully integrated super-agent, or a lightweight operating system and orchestration layer that acts more like a personal assistant?
Through its "Lobster" super-system or ecosystem, OpenClaw enables the community to tap into and orchestrate the full range of tools across the ecosystem in a much more accessible way.
As capabilities such as Skills and Harness continue to emerge, more people can build applications on top of these systems and apply them across different industries. This also makes OpenClaw a natural fit for the open-source ecosystem.
These are the two biggest takeaways OpenClaw has brought us.
Yang Zhilin: Building on the OpenClaw discussion, I'd like to turn to Zhang Peng. Zhipu recently launched the new GLM-5 Turbo model, which comes with stronger agent capabilities.
Can you tell us what differentiates this model from the alternatives on the market? And given the price increase, what signal do you think that sends about the current state of the market?
Zhang Peng: We did push out an update a couple of days ago. It was actually a milestone that had already been on our roadmap—we just decided to move it forward. The primary goal was to bridge the gap between conversation and execution.
I completely agree with a point that came up earlier: OpenClaw has made people realize that large models are no longer just chatbots—they can actually get work done. But once you move from conversation to execution, the requirements on the model become much more demanding. It needs to plan over long horizons, retry when things fail, manage and compress context, debug issues, and often work with multimodal inputs. These requirements are fundamentally different from those of traditional dialogue-oriented models. GLM-5 Turbo was specifically strengthened in these areas. In particular, as several people mentioned earlier, enabling the model to keep working continuously—even running in a self-directed loop for up to 72 hours—required a significant amount of engineering and optimization on our side.
Another important aspect is token consumption. When you ask a capable model to complete complex tasks, token usage can become enormous. Most users don't experience this directly—they simply see the bill increasing. So improving token efficiency was also a major focus for us. For complex tasks, the model is now able to accomplish more work with the same token budget.
At its core, GLM-5 Turbo is still a general-purpose model designed to handle a wide range of tasks. What we've done is strengthen a number of capabilities that are particularly important for agent-style workloads.
As for the price increase, the explanation is fairly straightforward. We're no longer in a world where users ask a question and receive a single response. Today's agent workflows often involve long reasoning chains, writing code to interact with underlying infrastructure, debugging issues, and continuously correcting mistakes along the way. All of that comes with a substantial computational cost.
The number of tokens required to complete a task can easily be ten or even a hundred times greater than what's needed for a simple question-and-answer interaction. As a result, both costs and pricing have increased accordingly.
The model itself has become more capable, and inference costs have risen as well. We believe pricing should ultimately reflect the value being delivered. Competing solely on low prices over the long term isn't healthy for the industry, and that's one of the factors we've considered.
At the same time, it allows us to build a sustainable business model, continue improving model capabilities, and provide users with better models and more efficient access to compute over time.
Yang Zhilin: Open-source models and inference compute are now forming an ecosystem, where a wide range of open-source models can run on diverse inference compute to deliver greater value to users.
With the rapid increase in token consumption, we are also seeing a shift from the training era to the inference era. Xia Lixue, from an infrastructure perspective, what does this shift to the inference era mean for Infinigence?
Xia Lixue: We are an infrastructure company born in the AI era. Today, we work with teams such as Kimi, Zhipu, and MIMO to help them make more efficient use of our token factory.
At the same time, we also collaborate with universities and research institutions. This has led us to continuously ask a fundamental question: what should the infrastructure for the AGI era actually look like, and how do we gradually build toward it over time? We are actively working toward this, with a clear view of the short-, medium-, and long-term challenges that need to be addressed.
In the near term, the most immediate challenge is the surge in token demand driven by agent frameworks like OpenClaw, as discussed earlier. This places significantly higher demands on system efficiency, and pricing adjustments are one way of responding to that pressure.
From the beginning, we have taken a vertically integrated hardware–software approach. For example, we integrate almost every type of compute chip available to us, connecting dozens of domestic chips and multiple compute clusters into a unified system, in order to address the shortage of compute resources in AI infrastructure.
When resources are constrained, the best strategy is first to utilize everything available, and second to ensure that every unit of compute is used where it has the highest impact.
At the core, our current focus is to build a more efficient token factory. This includes optimizing the tight coupling between models and hardware—such as GPU memory systems—and exploring deeper co-design opportunities between model architectures and hardware architectures.
However, solving efficiency at this level only gets us to a standardized token factory. For the agent era, that is not enough. As mentioned earlier, agents are more like autonomous actors—we can assign them tasks directly.
I strongly believe that much of today’s cloud infrastructure was fundamentally designed for running programs and serving human engineers, not for AI systems.
The interfaces we build today are still human-centric, and adding an additional layer for agents inevitably constrains their capabilities. For instance, agents are capable of reasoning and launching tasks at second- or even millisecond-level latency, while systems like Kubernetes are still optimized around human-scale, minute-level operations. This mismatch means the underlying infrastructure needs to evolve. This is what we refer to as agentic infrastructure: building a more intelligent, adaptive factory. This is exactly what we are working on at Infinigence.
Looking further ahead, in a true AGI era, we believe even infrastructure itself should become an intelligent agent. The system we build should be capable of self-evolution and iteration, forming an autonomous organization with an “agent CEO”—potentially coordinated by a Claw-like system—that manages the infrastructure and continuously evolves it based on the needs of AI customers.
In this setting, AI systems can better interoperate with one another. We are also working on improving agent-to-agent communication, as well as cache-to-cache replication capabilities.
Ultimately, infrastructure and application development should not be isolated layers. They should co-evolve and generate rich synergy. This is what enables true software–hardware collaboration—the alignment between algorithms and infrastructure—which has always been the core mission of Infinigence.
Yang Zhilin: Next, I’d like to turn to Luo Fuli. Xiaomi has recently contributed significantly to the community by releasing new models and open-sourcing parts of its the underlying technology. In your view, what are Xiaomi’s unique advantages when it comes to building large language models?
Luo Fuli: Let me set aside Xiaomi’s specific advantages for now. I’d rather talk about the broader strengths of Chinese teams working on large language models, which I think is a more generally meaningful topic.
Two years ago, foundational LLM teams in China had already made important breakthroughs. Working under limited compute, they introduced a series of architectural innovations that, at the time, could be seen as trade-offs for efficiency—for example, MoE-based designs in the DeepSeek V2 and V3 series. The idea of achieving the highest possible intelligence under fixed compute constraints gave domestic teams a lot of confidence and conviction. Today, with domestic chips—especially inference and training hardware—becoming less of a bottleneck, those constraints are no longer as strict.
At the same time, those earlier constraints also pushed the field toward new architectural directions that improve both training efficiency and inference cost. We are now seeing a range of explorations, including hybrid attention, sparse attention, and linear attention mechanisms—such as DSA, NSA, Kimi’s KDA, and Xiaomi’s next-generation HySparse. These represent a shift away from the previous generation of architectures like MIMO-V2, and reflect how we are rethinking model design for the agent era.
Why is architectural innovation so important? One reason is OpenClaw. If you actually use it, you’ll notice that it becomes more capable and more useful the longer you run it. A key prerequisite for that behavior is inference context.
Long context has been discussed for a long time. The goal is to achieve strong model performance in long-context settings, while keeping inference cost and latency low. Today, many models still cannot handle one million or even ten million tokens of context efficiently, because the cost and latency become prohibitive. If we can make long-context inference efficient at that scale, then we can unlock truly high-productivity workloads, and enable much more complex tasks to be handled within a single context window—potentially even self-iteration within very long contexts.
By model self-iteration, I mean the ability for a model to evolve within a complex environment by leveraging extremely long context—whether that means improving the agent framework itself or even influencing model parameters, since long context can itself be viewed as a form of parameter evolution.
Achieving long-context efficiency—both in architecture design and in inference optimization—is a full-stack competition. On the pre-training side, we began exploring long-context-efficient architectures about a year ago. Today, the challenge has shifted toward post-training: how to push long-context systems to achieve both high upper-bound performance and strong results on long-horizon tasks.
Post-training now focuses on several key problems: designing more effective learning algorithms, collecting real-world data with long-range dependencies at scales of one million, ten million, or even one hundred million tokens, and incorporating trajectories generated in complex environments.
With the rapid progress of large models and the rise of agent frameworks, we’ve seen inference demand grow nearly tenfold over a short period, and total token usage could increase by as much as 100x this year.
This shifts competition into a different dimension: compute infrastructure, inference chips, and ultimately even energy.
Yang Zhilin: That was a very insightful discussion. Next, I'd like to turn to Huang Chao. You've built several influential agent projects, including NanoBOT, and have a strong following in the community. From a Harness perspective, what are the key technical directions you think people should be paying attention to going forward?
Huang Chao: If we abstract agent technology, it really comes down to three key components: planning, memory, and tool use.
Planning today is still mainly focused on long-horizon tasks or very complex procedures—sometimes on the order of hundreds of steps or more. Looking forward, planning may evolve toward embedding more of this complex task knowledge directly into the model itself.
Skills and Harness (control/governance) mechanisms help mitigate errors introduced by planning. High-quality Skills, in essence, allow the model to directly handle certain difficult tasks.
Memory has always suffered from issues such as imperfect compression and inaccurate retrieval. In long-horizon and complex scenarios, memory usage can grow rapidly, putting significant pressure on the system. Current Lobster-style systems typically rely on very simple representations, such as Markdown-based memory. In the future, memory may evolve toward more hierarchical designs and greater generality. At the moment, memory systems still struggle to generalize across domains such as coding, deep research, and multimedia, where data modalities differ significantly. How to index and retrieve memory efficiently across these settings is ultimately a trade-off.
Another issue is that as Lobster significantly lowers the barrier to building agents, we may move toward a future where each person has not just one Lobster, but a fleet of them. Compared to a single agent, multiple agents will dramatically increase context usage, which in turn puts additional pressure on memory systems.
There is currently no robust mechanism for managing a fleet of agents like this. In complex scenarios such as coding and scientific discovery, this creates significant context pressure—both for the underlying models and for the overall agent architecture.
For tool use, the core still comes down to Skills. Issues such as inconsistent MCP quality and security concerns also exist in Skills. Although there appear to be many Skills available today, high-quality ones remain relatively scarce, and low-quality Skills can directly reduce agent task performance.
In addition, Skills are still vulnerable to issues such as malicious injection, which impacts overall system reliability. In terms of tool use, we likely need stronger community-driven efforts to build and improve Skills, as well as mechanisms that allow new Skills to emerge dynamically during execution.
Yang Zhilin: What Luo Fuli and Huang Chao just discussed highlights the same underlying issue from different angles: as task complexity increases, context usage grows rapidly. rom the model side, one direction is to extend native context length. From the agent harness side—including planning, memory, and multi-agent orchestration—we can push task complexity further within the same underlying model capability. Moving forward, these two directions will likely continue to reinforce each other and create more synergy.
Finally, I'd like to end with an open question. If you had to describe the next 12 months of LLM development and your expectations in one word, what would it be?
Huang Chao: Ecosystem.
Lobster has created a lot of engagement today, but in the future, agents need to evolve from being personal assistants into actual “workers.” Right now, a lot of usage is still driven by novelty. Over time, we need these systems to settle into real productivity tools, or even become coworkers.
hat requires effort from the entire ecosystem. Open source is also critical. Once technical explorations and model innovations are open-sourced, the broader community needs to co-build on top of them. Model iteration, the Skills platform, and various tools all need to evolve together into a better ecosystem tailored for Lobster.
My sense is that much of future software may no longer be designed primarily for humans. Across the ecosystem—whether GUIs, MCP, CLIs, software systems, data, or infrastructure—we will need to move toward an agent-native paradigm, so that agents can operate in a much richer environment.
Luo Fuli: Narrowing the horizon to one year is a very meaningful framing. By my definition of AGI, it will be achieved in five years.
If I had to describe the most important theme for the coming year in one word, it would be: self-evolution.
Although the term may sound abstract—and has been mentioned many times before—I’ve only recently developed a more concrete understanding of it, or rather, a more practical way to implement it, enabled by much more capable models.
The previous chat-based paradigm never fully unlocked the potential of pre-trained models. That potential is now being activated by agent frameworks. Once models are allowed to run over longer-horizon tasks, we observe that they can begin to improve and evolve their own behavior.
A simple setup illustrates this: if you add a verifiable constraint within an existing framework, introduce a loop, and let the model continuously iterate toward an objective, we find that it can consistently produce better solutions over time. On domestic models, this form of self-evolution can already run for one to two days, depending on task difficulty.
In scientific research settings—for example, exploring better model architectures where evaluation signals exist, such as lower perplexity—we’ve seen models autonomously optimize over two to three days on well-defined objectives.
From my perspective, self-evolution is the only mechanism that can generate genuinely new knowledge. It is not about replacing human productivity, but about exploring what does not yet exist, in a way similar to top researchers.
A year ago, I expected this capability to take three to five years to mature. More recently, that timeline has shortened to one to two years. With large models combined with strong self-evolving agent frameworks, we may be able to significantly accelerate scientific research.
Within our team, research workflows are highly uncertain and creative. We’ve found that using Claude Code together with top-tier models has improved research efficiency by nearly ten times.
So I’m very excited to see this paradigm extend into more disciplines and domains.
Xia Lixue: My keyword is “Sustainable Tokens.” AI development is a long-term process, and we want it to have lasting momentum and vitality.
From an infrastructure perspective, one of the key challenges we see is that resources are ultimately finite. Similar to the idea of sustainable development, a central question for us as a token factory is whether we can provide tokens that are sustainable, stable, and scalable—so that top-tier models can continue to serve a growing range of downstream applications. We need to zoom out and look at the full ecosystem: from energy to compute, from compute to tokens, and ultimately from tokens to human GDP. This entire chain needs to evolve through continuous economic iteration.
We are not only leveraging domestic compute resources, but also extending these capabilities overseas, with the goal of connecting and integrating global resources.
I want to build what you might call “tokenomics with Chinese characteristics.”
In the past, we talked about “Made in China,” where China’s cost advantages in manufacturing were transformed into high-quality goods exported globally.
What we are aiming for now is something like “AI-Made in China.” By leveraging China’s advantages in energy, we can sustainably convert them into high-quality tokens through token factories, export them globally, and eventually become a global token factory.
That is the value I hope China can bring to the global AI ecosystem this year.
Zhang Peng: While everyone else is looking at the bigger picture, I’ll stay a bit more grounded. Over the next 12 months, the biggest challenge we face is likely compute.
All of these technologies—including agent frameworks—have improved creativity and productivity by an order of magnitude for many people. But the key prerequisite is simple: people actually need to be able to use them. We can’t have a situation where an agent spends a long time thinking without returning an answer simply because there isn’t enough compute available.
Compute constraints have also been a bottleneck for our own research progress. A couple of years ago, Academician Zhang Yaqin once said: “Talking about chips hurts relationships; without chips, there are no relationships.”
Today, we find ourselves in a similar situation again, but under different conditions. We’ve shifted into the inference era, where demand is truly exploding—by ten or even a hundred times. And there is still a significant amount of unmet demand. The question now is: what do we do about it? That’s something we all need to think through together.
