Z.ai/ Zhipu: one of the first major LLM start-ups to go public. Competition with giants and aims for AGI

In this episode, I sit down with Zixuan Li, who leads the chat API and global partnerships at Z.ai, one of China’s leading LLM labs (one of the four tigers) and now one of the first to head toward an IPO.

Z.ai started as THUDM, a Tsinghua data-mining lab best known in open-source circles for GLM and CogVideo, and has since grown into a model-as-a-service platform powering millions of devices and thousands of enterprises in China and beyond.

We talk about what it actually means to be an “independent” lab in a market dominated by platform giants like Alibaba, ByteDance, and Tencent, why Z.ai pivoted from SOE-heavy infrastructure projects to a product-led GLM stack, and how they landed on a different business model, and the creation of the GLM Coding Plan, instead of charging by tokens. Zixuan is very candid about pricing (“If Anthropic charges $200, we charge 200 yuan”), the realities of on-prem-first China vs cloud-first West, and what it’s like to race against Minimax and Moonshot with fewer GPUs and less cash.

We also zoom out and look at China’s AI talent pipeline (and the meme that the AI race is “Chinese in China vs Chinese in the US”), how he thinks about AGI as self-learning agents that live on your phone, why he’s comfortable being a white-label backbone in the Global South, and where he sees China’s AI landscape in the next 6–12 months. If you want a ground-level view of how a Tsinghua spinout is trying to survive, and maybe win, in the LLM wars, this one’s for you.

Newly launched (Dec. 22) GLM 4.7:

In today’s world, there’s no shortage of information. Knowledge is abundant, perspectives are everywhere. But true insight doesn’t come from access alone—it comes from differentiated understanding. It’s the ability to piece together scattered signals, cut through the noise and clutter, and form a clear, original perspective on a situation, a trend, a business, or a person. That’s what makes understanding powerful.

Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.

For more information on the podcast series, see here.

01:20 – From THUDM to Z.ai: rebrand, Tsinghua roots, and model-as-a-service

03:30 – Quiet period & IPO: pride, pressure, and the business challenge of LLMs

06:33 – Pivoting from SOEs: infra projects, agentic models, and why strategy followed capability

07:25 – Competing with Minimax, Moonshot & DeepSeek: focus, compute, and capital constraints

08:34 – Chasing benchmarks vs real-world IQ: math, humanities, and alignment trade-offs

11:05 – On-prem vs cloud: why Chinese SOEs still won’t touch APIs

13:43 – Zero-retention and trust: can China’s culture around data ever shift?

14:07 – Inventing the GLM Coding Plan: subscriptions, stickiness, and “pay by value, not tokens”

16:00 – “If Anthropic charges $200, we charge 200 yuan”: pricing strategy and margins and GLM’s open-source flywheel

19:41 – Who really pays: sticky indie devs, big tech customers, and bargaining power

23:32 – GLM Coding Plan vs Cursor/Qwen/Claude: plans, agents, and avoiding lock-in

25:57 – Z.ai’s AGI ladder: AutoGLM, self-learning, and personalized weights

27:03 – Independent labs vs platforms in China: speed, resources, and “dirty work”

29:34 – Moonshot vs Z.ai: chasing the moon vs being “down to earth”

30:53 – Will China’s LLM market consolidate?: 5–10 players, Doubao, and video-generation winners

31:44 – Doubao phone & Honor partnership: bargaining power with OEMs

34:11 – Beyond China–US: Global South strategy and being a white-label backbone

35:29 – Being comfortable as infrastructure: letting others own the brand

38:05 – Who joins Z.ai and AI talent: thriving with scarce resources

40:07 – Culture, 007 hours, and survival: what it takes to be infrastructure

42:33 – Social welfare, AI safety, and cheap tools in India & Indonesia

44:38 – How China actually talks about AI safety (or doesn’t)

47:29 – Differentiated view: why Zixuan believes you should “enjoy lacking resources”

AI-Generated Transcript

Grace Shao:Hey Zixuan, thank you so much for joining us today. Really excited to have you on. Walk us through your journey and what led you to Z.ai to start off with.

Zixuan Li:Yeah, so currently I’m the head of Zhipu AI’s chat API services and also head of global partnerships. I collaborate with LMSys Chatbot Arena, OpenRouter, Vercel, these large companies, and ship our products through their platforms.

The reason why I joined Zhipu is it’s one of the leading AI labs in China and I can do overseas businesses, because I have a background at MIT’s Schwarzman College of Computing. So that brings my knowledge into real-world practice.

Grace:I see. Was there any incentive for you to move back to China versus stay in the US?

Zixuan:I think it’s more personal, because my wife’s based in China and she’s used to her work, so there’s no way she can move to the US.

Grace:Fair enough.

So let’s talk about the company’s mission and origins, because I think it does seem a bit mysterious, especially to people outside of China. From the outside, people know Zhipu, Z.ai as one of the leading Chinese LLMs. But that doesn’t really capture everything you guys do, right?

In your recent prospectus, you describe yourself as a MaaS — model-as-a-service — company first. So tell us about that.

Zixuan:Okay, so before Zhipu AI, we were called Zhipu or THUDM, because we named ourselves by the AI lab’s name. We originated from Tsinghua University’s data mining group — THUDM. But I think it’s hard to pronounce, and also “Zhipu” is also very hard to pronounce. So this year we bought the Z.ai domain and finally changed our name to Z.ai.

When we were called THUDM, we were very famous inside the open-source community because we had a lot of repos, a lot of models under the THUDM name. And we open-sourced not only text models, also CogVideo, CogView, these models. I think they were sold at that time.

But with the launch of VEO, Hailuo, and also a lot of current top models, we began to be more focused — basically more focused on text models, visual understanding, and so on. So I think that’s the origination of the lab.

But as you said, there’s this terminology called model-as-a-service. From our side, when we compete with large companies like Alibaba and ByteDance, we need to be more focused. They have their inference level, they have their cloud services, but we don’t. So we try to let the model itself provide the service — like the API, or technologies like visual understanding — and try to use the model itself to be the selling point.

Grace:I definitely want to double-click on how you position yourself compared to peers — a few of them you just mentioned, whether it’s Minimax and Moonshot, and then you also mentioned the BATs.

But to start off with, you’re currently in your quiet period as your prospectus just hit the public. And if successful, you will become one of the first major LLM startups globally to be listed on a stock exchange. How does that feel?

Zixuan:I think we are proud of it, but things are very challenging, because it’s really hard to do LLM inference. Both OpenAI and Anthropic have very high revenue, but a lot of loss on their income statements. So we have to figure out how to make money from large language models and also provide cheaper service to the customer.

So I think it’s only a starting point for us.

Grace:Definitely. I think right now only the big tech companies in many ways are essentially seeing ROI, and the model companies and the model labs themselves are really finding it hard to make a profit.

I want to ask you about the branding. You did say you guys changed your company’s name to Z.ai this year, partially because Zhipu is just hard to pronounce. But was that also related to the fact that you guys seem to have made a pivot into really focusing on going global? Z.ai seems to be a lot more non-Chinese-native-speaker friendly, right? So is that the push right now?

Zixuan:I think that played an important role, because we have observed the success of DeepSeek, Qwen — they got famous globally and Chinese people will think that they are the “SOTA” in the domain and their models are the best. They are recognized by NVIDIA and other large company CEOs. So I think that’s one factor.

But the other factor is when we changed the name to Z.ai, the dot also plays an important role. We want people to enter that URL into their browser and try to visit our website. Yeah, two factors.

Grace:And tell me about your origin story, actually. You mentioned earlier you started off from the Tsinghua data mining group. Maybe provide some context to people outside of China. What does Tsinghua represent? I mean, it’s an institution, it’s a university, but why are so many of these LLM companies or even deep tech companies coming out of Tsinghua right now?

Zixuan:I think it’s kind of a combination of Stanford and MIT. So talents are everywhere and there’s a lot of funding from internally and also externally. And also people are chasing the highest IQ there. So it will be very natural to pursue AI in Tsinghua University.

Grace:So I have a question on that, because a lot of tech companies, even the previous generation internet companies that came out of Tsinghua, had some kind of connection with Beijing city. And my understanding is Zhipu’s original business model was also very focused on SOEs and local government work, both in China and even across Southeast Asia.

Before the more recent pivot leaning into tools and APIs, what were the reasons for the pivot from the heavy AI infrastructure focus and SOE projects to a much more product-led tools and API strategy?

Zixuan:I think it depends on the capabilities of the model, because nowadays the model can perform agentic tasks, use tools, use coding to perform tasks. But before that, we could only do customer service, data processing — these “dirty work.” I think it’s better for SOEs or other scenarios.

But with the change of Cloud Code, GLM 4.5, these agentic stuff, people can really use the model in other areas like Manna, Gainsburg, Lobe. So I think it’s not only our strategy, but also the capabilities of the model have changed a lot.

Grace:Yeah, and I think to put you on the spot, where do you see yourself compared to your peers — like the DeepSeeks and the Minimaxes and the Moonshots of the world?

Zixuan:I think compared to Minimax and Moonshot, we are close competitors. We are startups, but DeepSeek is like another kind of enterprise because they have Qwen. So I think they’re very unique, and also ByteDance, Alibaba, they’re sitting at the same table. So they’re from large enterprises.

We are all chasing somehow the same direction, but we lack compute, we lack money compared to these giant enterprises. So we need to stay very focused.

Like Moonshot, they focus on the Kimi K2 series. They only release Kimi K2, another K2 and K2 Thinking this year. And also Minimax — they’ve become more focused and they kind of shift away from multimodal to text models. I think it will be very fierce. The competition will be very fierce in the coming months.

Grace:And for yourself, when you say you’re chasing the same direction, what does that direction look like in layman’s language?

Zixuan:In layman’s language, I think… more practical. Because the reason why we do coding and agentic is that we see people using it. We see people using Codex, Manna, Claude Code. So that represents high token usage.

And also we are chasing AGI, or the IQ. So we want the model to solve very hard math problems, to memorize a lot of hard stuff. As you can see from a lot of benchmarks like ARC-AGI, HLE, we’re also chasing in that way.

So we balance the two: figure out how to balance the performance on benchmarks and in real-world development.

Grace:I actually have a question on that that’s a bit off-track from the business strategy side of things, but I wonder how you view this.

So you’re saying you’re chasing benchmarks on math problems, IQ, advanced physics, etc. But what about the humanity side of things? I think people are still questioning whether AI can be used to replace humans in a more humanities-focused industry or sector.

Zixuan:So that’s a very big issue. But for now, I think it’s still not there yet, because we see hallucinations happen inside the model and instruction following is not very good.

So we test the status of that harm and try to assess what we can do with this model and try to synthesize a lot of data to make it more aligned to human judgment or other things. I studied alignment at MIT. I know a lot of stuff, but when I came back to China, I thought we were not there yet.

So capabilities, I think, are still more important than alignment at this stage. But we need to focus on the future and try to prevent something really bad happening. I’ve learned a lot of news like suicide or emotional feelings, depression. But somehow I think it’s still not that harmful yet.

We try to incorporate as much human judgment or human alignment into the model as we can. But as I said, it’s kind of a balance between different aspects.

Grace:Yeah, it’s always a balance between setting up the guardrails and actually still allowing the technology and innovation to continue, right?

I want to reshift the focus back on business model, pricing, deployment. Reading the prospectus, what stood out to me was how much you support both on-prem and cloud.

What are the main product lines today of Z.ai or Zhipu, and how do you map those onto on-prem versus cloud deployments in terms of how customers actually adopt GLM? Because I do believe I read that in China there’s a very different preference. In China, it seems like more people prefer on-prem, right? Whereas in the US it’s more cloud — or did I understand that incorrectly? Please explain.

Zixuan:Yeah, I think you have a very good understanding of the current status, because large SOEs, large enterprises in China, prefer on-premise or more private deployment. So it’s hard to do API services with them.

But currently a lot of tech companies accept API services. So we collaborate with nine out of ten of the largest Chinese tech firms or social media firms with our API services. So it depends on their needs.

We try to sell API, but actually some people have privacy concerns. They have policies not accepting API services. They don’t want any data to go away from their servers. So basically it depends on the users’ needs.

Grace:This is actually kind of a reflection of what happened during the SaaS era too, right? Chinese SOEs and big companies would rather build their own app — maybe not even be as good — but they just don’t want to give their data out to anyone and have that potential security risk, right?

So do you think that will change in terms of company culture as we see AI continue to develop, or do you think that will continue to be the trend in China — that this would be the differentiating point between the Chinese market and maybe the Western markets?

Zixuan:I think it will continue to be the trend. As you said, we had that pattern in the era of SaaS. And when we go to the AI era, nothing changed.

But somehow, we can figure out a way to balance, because there is more “private host on cloud” service. And we’re trying to store user data in a more secure way, with a zero-data-retention policy. That will mitigate the risk and the issues and try to let them feel more comfortable with it.

Grace:I see.

A lot of Western developer tool companies now go pure usage-based, but you guys also have a GLM Coding Plan — basically for developers with very low entry points. Why did you choose a subscription approach versus going with other pricing models? I guess this part, I just want to understand how you guys are making money right now, especially as you’ve just had your prospectus go public.

Zixuan:Yes, I think we are the company that invented this coding-plan business model, because we found out that API users are not sticky. One day they use Claude, they can switch to Gemini or GPT another day. It’s the same with Chinese models.

So we remembered: why do we pay for subscriptions — Spotify or YouTube service? Maybe we just listen once during the whole month, but we don’t regret it, right? So we don’t want our users to pay by tokens. We want them to pay by value or by the product itself.

So if they just use it once or twice within a month, I think it’s totally fine. If they want to subscribe the other month, we’ll try to provide better service. We have GLM 4.5, GLM 4.6, GLM 4.7, trying to ship better models. But if they decide to quit, I think it’s still good for us because they paid for one month, not just several tokens.

So the users may be very sticky here, and we have our branding — not only the model, not only GLM, but also the subscription, GLM Coding Plan. So when we use Cursor…

Grace:But if they were to use it a lot, would it be loss-making for you guys then?

Zixuan:I think it’s still an issue for Claude Code and also Codex. You have to balance the rate limit and also the service level. So for us, we are very generous, but we’re trying to operate globally, because that will make our traffic more stable — not receiving very high demand at one time and no demand during the nighttime.

Grace:I see.

I know that you’ve been quite active in a lot of podcasts recently. You were on ChinaTalk, you were on Steven Hsu’s. And one of the lines you said, I think it was on ChinaTalk, you said, “If Anthropic charges $200, we charge you 100 yuan.” I thought it was quite funny. It was very memorable.

So how did you make that kind of decision, and how does it work in practice? Does that mean you have a long-term structural advantage, or does that mean you charge less and therefore have smaller margins?

Zixuan:I think we serve different customer needs. For example, someone sells Rolls-Royce to people, but we sell Benz to people. Both are good cars, but Rolls-Royce charges way more than Benz. The performance, I think, is very close.

But like I said, Anthropic deserves that premium. But by selling Benz, we can still earn a lot of money. Maybe the profit margin is very thin currently, but we can lower the inference cost. We can change our infrastructure to make it more profitable. So it’s a long-term strategy, not focused on the current cost structure.

We’re trying to make people more sticky to the brand, more sticky to the service. I think it’s essential at this time.

Grace:Essentially, you’re saying the utility purpose of having a car — getting from point A to point B — is the same, but maybe you’re selling a Toyota then or a Honda, right? Not even Mercedes, which still charges a pretty premium margin.

I remember in the same interview, you were kind of challenged, saying: look, you only really take up about 5–6% market share in China for general-purpose models. But you said, “Wait, 5% is enough.” What exactly are you thinking when you say 5% is enough?

You serve — I think from public disclosures — 123 large enterprise clients on-premise deployments, plus around 5,500 customers using cloud services. How does that actually stack up to your peers? Because it doesn’t look like huge numbers, to be honest. And that already is 5–6% of China’s market share.

Zixuan:I believe that the 5% refers to the percentage of all the GLM services.

Grace:Yes, sorry, GLM.

Zixuan:GLM services, because we open-source our models. And it’s hard to get revenue when you open-source your model because you have to compete on speed and stability.

But I think our model is good enough. Maybe it’s not like Toyota — it’s kind of a Benz. And we let more people adopt GLM, like what Qwen did in the past. They open-sourced their reflection models and more people tried out Qwen. They got famous, so people believed they got better service from Alibaba.

It’s the same underlying methodology from our side. So if GLM gets really famous, even 5% is enough for us. But if it’s not famous, 5% is totally not okay. We’re trying to make our model more influential, like DeepSeek, like Qwen.

Grace:I see. I do want to go into GLM and your tools later as well. But one last question on the business side of things. We kind of touched on this: you said a lot of your customers are the big tech companies, but in the beginning, they were the SOEs, right?

So right now, is there any pattern you’re seeing in terms of who becomes the most valuable users and who becomes the most sticky users and who are actually willing to pay the big bucks for your product or for your service?

Zixuan:So from my department, I think two types of customers. One is individual developers, because we have the GLM Coding Plan. Someone bought a yearly max plan. A lot of users bought yearly plans. They are very sticky.

And the other type is large tech companies, because we are still leading the open-source models. So we have bargaining power. Maybe they want to shift away from our model and choose other models, but we keep evolving from 4.5 to 4.6 and 4.7. Every time they try to change the model, they find that we can ship better models.

So these customers are very sticky. And they care more about performance because we are leading in performance. They care less about cost or relationship.

Grace:Nice.

Let’s actually double-click on GLM. You mentioned GLM 4.5 and 4.6. They’ve been positioned as highly competitive on coding and reasoning, and you’ve often been the highest-ranked Chinese model on public leaderboards.

When you compare the GLM series to US and Chinese peers, what dimensions matter most to you beyond the leaderboard scores right now? And where do you think GLM actually genuinely stands out compared to other peers, whether it’s American ones or Chinese peers?

Zixuan:I think real-world development, real-world practices, and general chat — these real practices — are more important than benchmarks. And in terms of real-world experience, we are tier two, because I believe Anthropic, DeepMind, and OpenAI have better user experience compared to us.

But I think we are enough compared to other open-source models, because we understand user needs. We have better quality in data — pre-training data and post-training data — and we’ve figured out ways to synthesize agentic tool-use trajectories and very hard problems. That makes us stand out in solving these really tough problems.

Because when you look at the benchmarks, they are not for real-world practices. Some are very tough, but it doesn’t mean they stand for human practices. Because we have a lot of customers… yeah.

Grace:Yeah, I’m going to challenge you on that actually. What about the Alibabas of the world? Because when I speak to Alibaba or Tencent, they also say their biggest differentiating point is real use-case data. And frankly, they have all the existing touch points with their users, whether it’s getting data through helping with businesses, enabling businesses, or consumer use. They probably have the best data, right? So how do you compete with that?

Zixuan:I think that’s their advantage in 2024, but not 2025. Because in 2025, most of the high-quality data we need, we have never met in real-world use cases.

When you want to create a slide, you first do search and then come back and do another round of thinking, and then choose a design tool or something like that. Nobody interacts with Alibaba’s product like that. So you have to fully understand Cursor, Claude Code, Manna — how these tools interact with people.

So ByteDance and Alibaba’s customer data cannot play a role in today’s agentic era. We have understanding of maybe Claude Code or Codex — we try to understand how a top-performing agent manipulates tools and how our model can be integrated in that system.

Grace:I was actually going to ask you about the GLM Coding Plan. So for context for listeners, it’s essentially their tool, like a Cursor tool.

So how does the GLM Coding Plan actually compare with Cursor or Alibaba’s coder, as you mentioned, or Claude, in terms of coding experience? For a developer who already knows these tools, how would you explain the distinction — or, you can be frank, is it mainly a pricing advantage here?

Zixuan:Okay, so I want to compare Cursor with GLM Coding Plan, not the model. Within Cursor, you have one coding agent and you can switch between different models. But with GLM Coding Plan, you first select the model and then you can switch between different tools.

You can integrate GLM into Claude Code, Kimi Code. You can even use GLM in Cursor with GLM Coding Plan. That made our product or model widely accepted or widely integrated into these systems — not just for Claude Code, but also it can be integrated into Cursor or Kimi Code.

We understand different coding agents and try to synthesize data that best fits these coding agents’ needs. And there’s no lock-in for our users.

Grace:So your GLM Coding Plan is not only your proprietary model, right? You actually are open to multimodal?

Zixuan:Yes, it’s a model. GLM Coding Plan is called a plan, not an agent, not something like Claude Code. You subscribe to an API, you’re not subscribing to a product. You use that API maybe in Claude Code, maybe in Kimi Code. So you can choose the mode.

Grace:Yeah. Okay.

Okay, thanks for explaining that to me. That’s helpful, I was getting a bit confused there.

Now I wanted to ask: in your prospectus, you laid out five stages of progression into AGI. We talked about your vision of AGI earlier. You said it’s about real-life implications, real-life practicality, usage of AI.

When you look at where you guys are at right now, what does crossing the next stage look like in terms of concrete capabilities or products or tools? Or maybe a more straightforward way of asking this is: what should we be expecting from you guys in 2026 to help you progress on your so-called AGI pursuit?

Zixuan:Maybe self-learning. Because currently when we do reinforcement learning, we synthesize all the data, we prepare the data beforehand, but the weights of the model won’t change during the interaction.

For example, we have this AutoGLM. It’s a model that can be deployed on your phone and can manipulate different apps for you. It can order food or order an Uber for you, but it’s the same model for everyone.

To chase AGI, we might have AutoGLM for everyone. When you interact with the model, the weights of the model may change. Currently, we have a memory engineering package that’s more on the engineering side — handling this memory stuff.

But for AGI, it needs to be very personalized. Every model needs to be personalized. The model learns from the environment, from the interaction. We also call it on-policy reinforcement learning.

Grace:I see.

Let’s take a step back and look at China’s overall LLM landscape and competition. You kind of alluded to this earlier — you guys are in the same pool as the Minimaxes and Moonshots of the world. Then there are the big techs like Alibaba, ByteDance, Baidu, Tencent, even Huawei these days, right? There’s so many. Everyone’s producing their own LLMs now.

From inside the ecosystem, what do you see as the structural differences between independent labs versus the big tech platforms — in terms of commercialization of their models as well as their incentives and objectives in the coming year or two?

Zixuan:Strategy and objectives. Because we lack resources, we need to be very focused. And when we are very focused, we need to move very fast.

For talents, our team is very small. I lead a team…

Grace:It’s not that small — a couple hundred, right? You guys have like 800 people now?

Zixuan:But for every team, there are just a bunch of people. We have sales, we have product solution, but for the product team, product solution, or training team, sometimes you need to be very lean. You don’t have to hire a lot of people, because they chase different directions.

Sometimes you have to hire people that can do “dirty work.” Maybe one person is enough to do all the training on this side, and you have a bunch of people preparing data or understanding customer needs for you.

Like I said, you have to understand Claude Code, you have to understand these coding agents. So there will be people studying all the products, looking inside these products to see why they are performing so well.

But for large enterprises, they can hire a lot of researchers. They have enough resources to do a lot of experiments. They have compute, so they worry less. Maybe they can find some scientific breakthrough from those experiments.

But in terms of model performance, I think our competitive advantage is we are closer to users and customers, because we move faster together with our users.

Grace:So that’s how you position the startups versus incumbents. But what about just within the startups yourselves? How do you differentiate yourselves between one another?

Zixuan:I think compared to Moonshot — because we both originated from Tsinghua University, we know each other pretty well — I think we are more down to earth. We are the ones that care more about real-world usage or practices.

Moonshot is kind of… they have this “AGI plan,” chasing the moon or landing on the moon, and they have more imagination on the surface. We’re also chasing AGI, but when we train the model, we care more about real-world practice and usage.

Grace:You’re taking a more pragmatic approach. And they’re definitely, I think, a very eccentric bunch, right? Even the name — how it came about — was quite interesting.

So do you think eventually in the Chinese LLM space it’s going to be winner-takes-most? Maybe not winner-takes-all, but winner-takes-most? Or is it going to be able to support multiple strong players?

Because there’s been rumors about consolidation for a while. There are quite a few players for how big the market is, and like you said, it’s extremely capital intensive. Not everyone has this much money to keep burning through it. So where do you see the direction of this fragmented landscape right now?

Zixuan:I think the market is enough to include 5 to 10 players. I think it’s enough. And like I said, the large enterprises only accept on-premise deployments, so there’s no way a winner can take it all, because there are thousands of large enterprises. You don’t have the team to deploy models for every single enterprise.

But in terms of applications, maybe Doubao will take more than half of the consumer side. And also for video generation, there will be a winner. But I think the market is still very large to have all these players, and they will compete for a long time. I can guarantee that they will compete for a long time.

Grace:What do you think of the Doubao phone situation? This is completely random. This is not relevant to our LLM conversation, but I’m quite curious to hear your thoughts on it, because I think it’s making a lot of noise outside of China. People are quite curious to see where that will lead to.

Zixuan:So we are the first company to launch this phone use agent. But I think the issue is bargaining power. We also collaborate with a phone company, and instead of using something like a “GLM phone,” we finally used their name. Their phone, powered by our model.

Grace:Which phone is this?

Zixuan:Rongyao.

Grace:Okay — Honor. I think it’s called Honor, yes.

Interesting. You know what? I really haven’t heard about it, but I should look into it. Is it actually already available to the mass market or no?

Zixuan:I think the phone was launched last year, not this year.

Grace:Okay, super interesting. I’ll look into it.

Zixuan:Yeah, so a lot of phones at that time were powered by AutoGLM’s capabilities. But we don’t have the same bargaining power as ByteDance, so we cannot name the phone by our name. We just power their scenarios.

So it’s about bargaining power, I think. Because like there’s the ByteDance vs Tencent issue, also with WeChat — it really depends on how you split the revenue, the value, how you make sure that you won’t influence other people’s business.

So finally, you have a line: maybe this app will collaborate with you, and that app rejects your endpoint.

Grace:Yeah. For context for you listeners, WeChat rejected Doubao phone’s direct access, and there was a huge headline war on this like two weeks ago.

Okay, I want to pivot a little bit. Right now there’s a lot of focus on the China–US lens. And you yourself spent time in China and the US as well.

But I did notice in the beginning days of Zhipu you guys were actually really focused on the so-called Global South — for lack of better words — Southeast Asia, Latin America, maybe even Africa. Is that still a strategy you guys are pursuing? Looking to sell or actually embrace markets that go beyond just China and the US?

Zixuan:Yes, definitely. Because I think in GLM 2, GLM 3, we only had Chinese and English capabilities, but now we have more than 100 languages. So that can support us going beyond English-speaking countries. Maybe in Brazil, maybe in Malaysia, we have opportunities to showcase our model or showcase our product solutions to people and finally compete with those large enterprises.

But I think things are really different in those countries, because they also want their data as private as possible. They accept on-premise, and maybe they want white label — they fine-tune the model and they want to ship it to their citizens under their name, not GLM or Zhipu’s name. So we have to meet their needs and see what we can offer.

Doing business in the US, I think it’s much simpler because you have this API, you have products, you can do a coding agent, you can earn money. But when you do business in other countries, you have to go really deep, twist a lot of things, and try to make it happen.

Grace:It’s also interesting — I think you touched on something. You’re quite comfortable being that white-label provider, versus I think a lot of other companies, whether it’s ego or belief, are not as comfortable. They definitely want their name on it.

So it seems like you guys are actually the backbone supporting a lot of technology or clients without really having your name attached to it.

I want to ask you about talent. This is a question we touched on in the beginning — you said yourself you came back to China for personal reasons, because your wife is in China. But I assume that’s not the case for everyone.

There’s this interesting and funny joke going around saying right now in the AI war or AI race, it’s really between the Chinese in China and the Chinese in the US. It’s just funny — there does seem to be a high percentage of ethnic Chinese or Chinese nationals or Chinese-naturalized Americans or ABCs. If we’re being non-PC, people who look Chinese in the field.

Why is that? I don’t understand. Did Chinese people just get a tip-off saying AI is gonna be really big early on and they went into this field earlier, or what happened?

Zixuan:I think I cannot explain it, because doing math problems is simple for us. I’m not sure why other people won’t pursue this business.

Because when I did internships and research at MIT, I saw a lot of talented people beyond Chinese — they’re still talented. They finally went to Anthropic, OpenAI. But somehow people only care about Chinese because they are co-launching products with Sam Altman or Elon Musk.

I think people overrated the influence of Chinese people in the large language model area, because still there are a lot of enterprises not relying on Chinese.

Grace:It’s quite funny — it’s kind of like the last generation, where every Chinese student in the US is either studying to be a lawyer or a banker, and now everyone switched over.

Actually on a more serious note, how does the talent competition play out then? Do you see yourself at Zhipu having to really convince people to join you compared to a US peer?

Or do you think there’s certain tendencies for certain researchers that would prefer to work for a Chinese lab or return to China? How do you see that play out?

Zixuan:I think we finally choose the people that best match our environment. Like I said, we lack resources, but some people really enjoy the lack of resources — like me. Because I think it’s good to have a small team competing with a very large team, and you have better enjoyment when you conquer a puzzle or problem, or you finally win at the end.

So people who enjoy this feeling, we try to hire them. And like I said, we want to move really fast. We want people — both the product team and the training team — to understand the user scenarios, to understand the data itself, not just theory or the algorithm.

So we try to find those people, and they will finally choose us because they don’t care about compute or resources, or they find it too toxic competing with other teams doing the same experiments and the same thing. Because that happens a lot in large enterprises — a lot of teams doing the same thing.

Grace:Yeah, for sure. I think even when I speak to the BATs in China, there’s so much internal competition that drives people crazy. It’s internal politics that drives people crazy. But that also becomes an incentive for people to really push.

On that note, you guys are about what, 800 to 1,000 people altogether, roughly around 100 to 200 in R&D — something around that rough figure. It’s essentially not really a startup company anymore — it’s just small compared to how big the big tech incumbents are.

So at this size, and as you guys head into becoming a publicly listed company, do you see the culture changing? And what are the ways you keep your researchers, scientists, and engineers motivated? Are we seeing crazy salary numbers as well, like the ones coming out of Meta? How do you keep people motivated?

Zixuan:I think we are more lean, more entrepreneurial. Especially in our team, because I only slept 50 minutes for the past 24 hours. So we want to move really fast, faster than everyone else. Yeah… beyond that.

Grace:You’re going beyond 996. This is not 996, this is 007.

Zixuan:Because the competition is really fierce. Moonshot, Minimax — they’re doing an excellent job. And we also have DeepSeek, Qwen — not to mention the frontier AI labs in the United States. So we have to keep pushing. I think there’s no other choice.

Because when we try to do AI, we want to survive. Frankly speaking, survival is a very high standard for the tech industry. When we look at operating systems: Windows, macOS, Linux — I think that’s enough. And when we look at phones — only Android and iOS.

So the competition must be fierce when you want to be the infrastructure for the industry.

Grace:Yeah, I agree on that. Okay, well, I hope you get some rest soon after this call. I really appreciate you jumping on the call after 50 minutes of sleep today.

Looking at your long-term vision and where you guys are headed now, especially with an imminent IPO: in your public materials, you talk about AGI integration with the physical world and social welfare as a long-term vision. I think this is something not many AI companies frankly are really thinking about.

Even within our conversation, you’ve talked about the balance between tech acceleration and actually putting up safety guardrails, essentially to prevent more sad, tragic happenings caused by AI psychosis, etc.

When we look at this, how do you personally reconcile the social-welfare North Star with the commercial realities and the pressure you just talked about? Where are the areas where you guys are frankly more okay to let go a little bit for business gains? What areas are definitely your red lines that you cannot cross, where you really want to hold people accountable and ensure there are no AI-caused tragedies?

Zixuan:Yeah, I want to answer this by giving an example. We have this GLM Coding Plan — it’s very cheap, three dollars a month. A lot of people in India, Indonesia, or even in the United States use GLM Coding Plan to do their side projects or even their startup.

I just talked to a person today. He’s doing a startup that uses GLM Coding Plan to write a program that can collect recyclable bottles. They scan the bottle and recognize it, try to differentiate trash from recyclable products, and make it a real business. So we truly use AI to empower these businesses. You can see there is a lot of social welfare behind this.

People just use the coding, but you can use coding to do a lot of stuff. We provide the service, but we let people decide whether they try to contribute more or only care for themselves. So I think it’s a starting point for us.

With more powerful products — maybe next year — we can empower larger scenarios. Maybe we can empower robots.

Grace:And in terms of this topic, I think in the US there’s a very dominant voice and discussion about the potential risk of AI or the negative impact it might have on society. AI Proem and Differentiated Understanding, frankly, very much focus on the business strategy of technology. So a lot of my guests and myself, we focus on capital deployment and feasible business models.

But I do want to ask: you’re plugged in, you are in China, in the LLM space, in the AI space. Is there a discussion about AI safety, or are people really just quite focused on acceleration and pragmatic deployment and diffusion?

Zixuan:I think compared to the United States, not that much. It’s more pragmatic. But that’s still on people’s minds, because AI safety is still an issue for us.

We can see the ceiling, the threshold of all the current capabilities, and understand what’s the top priority for our model or our scenario, and try to fix those before going to the next step. But we always keep the security and safety issue in our head. And when that day finally comes, we can be fully prepared.

I’m engaged in a lot of these conversations in the US and I’m also part of Concordia AI. It’s an organization focused on AI safety. I’m part of it in Beijing. But when I left that company, I saw them — and for anyone talking about this — it’s not because people don’t care. We can train a model with better capabilities and also a safer system.

So there’s no trade-off at the current stage because we don’t have to balance performance with safety concerns. We can improve them at the same time.

Grace:I see, it’s more like taking a mindful approach.

I want to end on two quick questions. One is: usually people ask, “Where do you see yourself in the next five years?” Right. But I think for AI we can’t really ask that right now — no one will know what five years looks like.

But for our listeners: where do you think China’s AI space will look like in, let’s say, six to twelve months? Where do you think the focus will be, or the potential breakthrough?

Zixuan:Potential breakthrough may be integration with the physical world. When we see a lot of robotics companies and we see a lot of smart glasses, people are shifting focus from AI companies to these, we call it broader intelligence companies. So that might be a shift.

And also DeepMind — I think they’re doing the same path. When you look at Gemini, it’s not just a large language model but also it can perform world knowledge or integrate with real-world use cases.

When we look at Gemini 3 Pro use cases, someone is controlling the camera or trying to integrate with the computer. So there are a lot of things we can do with large language models.

Grace:Okay, I think the last question I have for you is a question I ask every single guest, which is: what is one differentiated view you have — a non-consensus view? It could be about anything: about the industry, about how you see the world.

Zixuan:I think for me, I’ll just share my thought: you should enjoy lacking resources — lacking people, lacking everything. In the AI world, that pushes you to the boundary. That pushes DeepSeek to change their architecture, to really do something innovative.

For me, I don’t train models, but I build products and do marketing. I had this GLM Coding Plan thought because we don’t have very loyal customers. When they use API, they try to shift from GLM to someone else and then come back one day. So I noticed these difficulties. That’s what we aim for: to try to solve these really tough difficulties.

Grace:Yeah, I think to your point — when you lack resources, it also means you have the agility and the flexibility to change things, because there’s no bureaucracies, there’s no chain of command, and it’s much faster.

I appreciate that in itself too. I was talking to a friend about that recently as well. Since I left big tech and left traditional media to do this myself, you have so much more flexibility, and sometimes you’re upset that you don’t have the access or the resources you used to have. But it does help you build faster and connect with your community faster.

Thank you so much anyway, Zixuan. I really, really appreciate your time. Please get some sleep after this.

Zixuan:Thank you too. Yeah, I truly agree that you have your competitive advantage, because those large media companies — their journalists won’t reach out to me. So it’s my honor being here, but also a good opportunity for you to understand the Chinese market.

Grace:Yeah, for sure. And I really appreciate you giving me your insights during this time. And for all the other Chinese AI labs out there, if you’re listening to this, please reach out. I would love to have a conversation. Thanks again.

Zixuan:Yeah, thanks.

AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Get full access to AI Proem at aiproem.substack.com/subscribe