Original · Unique Research · 2025-11-08
Editor’s note: This complete English edition preserves the original author’s analysis and Wu Weijie’s statements as of November 8, 2025. Revenue shares, comparative model performance, sovereign-model projects and market forecasts remain source-attributed, not independently audited findings. References to ‘today,’ ‘recently’ and ‘this year’ describe the original interview period; ‘last year’ refers to 2024. Q7 retains the source’s broad description of Claude restrictions. Anthropic’s September 4, 2025 announcement expanded existing unsupported-region restrictions to organizations controlled from those regions, including overseas subsidiaries; it was not a first-time blanket withdrawal from all users in China. The first-person narrator is the original author. No publication location is inferred.
Anthropic’s original announcement: https://www.anthropic.com/news/updating-restrictions-of-sales-to-unsupported-regions
Source-photo note: The original article includes a conference photograph with the label ‘China (Guangxi) - ASEAN Artificial Intelligence Enterprise Conference (CAAIEC).’ This text-first edition omits the decorative photograph; it contains no unique business figures or charts.
In this new AI-driven world, whoever can harness the leverage of computing power and algorithms has an opportunity to reshape the industrial order. In this global contest, Chinese AI companies occupy an unusual position: they are not followers, but they may not yet be leaders either. They are more like a “second answer” trying to forge an alternative path through a narrow gap.
A recent in-depth conversation with Wu Weijie, Senior Vice President of Zhipu AI, gave me a more concrete sense of this issue.
1. In the Era of “Models as National Power,” China Begins Exporting Sovereign Large Models
You may not have expected that, in some ASEAN countries, Chinese companies have already begun helping governments train their own “national large models.” Put differently, Zhipu AI is packaging and exporting its algorithms, training, and data-processing capabilities to build the AI foundation for a sovereign nation.
In the original author’s assessment, this goes beyond serving a single customer: it means helping a country construct a new form of digital sovereignty—something the author said OpenAI and Anthropic had not yet done.
What makes this possible is not just model capability, but the highly credible backing of a nation. It resembles an extension of the Belt and Road Initiative, except that infrastructure, roads, and ports have been replaced by GPU, API, embedding, and token capabilities.
This is a signal: once you possess your own large-model production line, you are no longer merely a country that uses models. You begin to acquire your own voice in AI.
2. “Alternative” Versus “Native”: Moving Beyond Distillation Anxiety to Refine China’s Own GLM
When discussing the gap between Chinese models and leading overseas models, Wu Weijie made a painfully direct observation:
“If you keep distilling other people’s work, you will never catch up.”
That is why Zhipu AI insists on conducting the complete pretraining process, even though it is more expensive and slower. The company does use distillation, but it refuses to stop there.
This commitment to a “native model” is not technical purism. It is a strategic choice. In an era when SOTA, or State-of-the-Art, leadership changes every day, only those with their own foundational accumulation earn the right to compete in the next round.
At the time of the original report, GLM-4.5 had, according to the report, achieved top-tier results among global open-source models across multiple evaluations of Agent orchestration, code generation, and reasoning. It is no longer merely a “cheap substitute,” but a Chinese solution capable of competing in the high-end market.
The real difficulty, of course, is whether the world will be willing to trust you once you are good enough.
3. Going Global Is Not an Outlet for Traffic, but an Entry Point for Trust
When many people discuss Chinese AI going global, their first thought is value for money. Wu Weijie puts it bluntly:
“If the performance is not good enough, there is no basis for talking about value for money.”
True globalization is not about opening a few offices or attending a few trade shows. It is about earning trust in an unfamiliar market.
Where does that trust come from? Part of it comes from the advantages created by national strategy, such as the Belt and Road Initiative. Part comes from product strength—for example, GLM’s ability to support multilingual communication in a community with a user base on the million scale. Another part comes from whether a company is willing to take responsibility for outcomes.
This is why Zhipu AI’s international expansion relies primarily on API and SaaS packages, rather than competing through a large sales force or emphasizing customized services. Developers care about only two things: results and stability. It is that straightforward.
That is the language Chinese AI must master as it goes global: proving its capabilities in terms developers understand.
4. The Consumer Market Brings Attention; the Enterprise Market Provides the Foundation
Large-model startups have flourished in recent years, but those who see the market clearly understand that consumer products may attract attention, while enterprise customers are where the money is.
Wu Weijie is candid: “More than 90% of our revenue now comes from enterprise customers.”
This is a sober and realistic assessment. In China, consumers have not yet formed a habit of paying for content or tools. The shared expectation that such products should be free has instead become the greatest pricing obstacle. Among government and enterprise customers, buyers will pay once a model can reduce costs and improve efficiency.
Chinese AI companies are therefore charting a path different from that of American companies. Rather than starting with consumer products and then moving into the enterprise market, they first take service quality to the limit and only afterward consider whether there is an opportunity to make applications popular.
This may be the distinctly Chinese character of a “service-oriented AI company.”
5. “Enterprise Software” Will Be Rewritten: The Next Generation Is Agents Paid for Outcomes
The passage at the end of our conversation was the one that gave me the greatest pause:
“China may not have an enterprise-software market; it may only have an enterprise-services market.”
Wu Weijie repeatedly made this point during his time at ByteDance. It means that Chinese enterprises do not buy software; they buy people, services, and responsiveness.
Large models, however, have changed everything.
Once an Agent genuinely matures, interaction will no longer consist of clicking, choosing from dropdown menus, and importing or exporting. Instead, it will become: “Say one sentence and give me the result.” At that point, customers will begin refusing to pay for the process and will shift to paying for outcomes.
This will have two consequences:
Enterprise-software companies will drastically reduce service headcount and turn labor into a “product”;
Service providers in the middle—especially outsourcers and vendors that compete through labor-intensive service delivery—will face structural collapse.
In the future, the enterprise-services market will split into an M-shaped structure. At one end will be “high-touch service companies” serving very large customers. At the other will be “automation platforms” in which teams of ten use an Agent to serve a million small and midsize enterprises.
6. Stop Counting on Middle Management: AI Is Reshaping the Organization
Future organizations will be flatter, built around smaller teams, and place more value on the ability to do the work than on the ability to manage.
This also means that large models are not merely changing applications and interfaces. They are thoroughly rewriting the meaning of the word “work.”
From this conversation, I saw a distinctly Chinese approach to globalization: winning not by telling stories or burning capital, but through service, products, and results.
The question is not whether a Demo looks dazzling, but whether the system can keep running without crashing; not who shouts the loudest, but who can deliver first.
Perhaps this is the real opportunity for Chinese AI: not to become the next OpenAI, but to use a more grounded approach that is better adapted to complex environments and become the world’s second answer—rather than a cheap substitute.
The future is arriving. The question is: are you ready?
Selected Interview Q&A
Zhipu AI’s Business and Strategy
Q1: Mr. Wu Weijie, which businesses are you currently responsible for at Zhipu AI?
Wu Weijie: I am currently responsible for three main areas at Zhipu AI:
General Internet Business Unit: I lead this business unit.
Commercialization of the MaaS business: I am responsible for commercializing the company’s MaaS, or Model-as-a-Service, business, which primarily means standardized API access services.
Management of the Zhejiang company: I concurrently serve as general manager of the Zhejiang company. We recently partnered with Hangzhou Urban Construction Investment Group to build its artificial-intelligence industry large model. The projects include China’s first large model for public buses, the AutoGLM flood-control agent, a multimodal large model for road-and-bridge maintenance, and an autonomous-learning sanitation-dispatch agent.
Q2: What is Zhipu AI’s API-service strategy? Does the company provide only its own GLM models?
Wu Weijie: No. Zhipu AI’s API strategy changed somewhat this year. We no longer provide only our own GLM models. We now offer “scenario-specific APIs” for particular use cases.
Specifically, this means integrating other vendors’ strongest capabilities in particular domains and giving customers a packaged API service with the best possible experience. For example:
Search scenarios: If a customer needs to search content in a particular ecosystem such as Zhihu, Zhipu AI’s API may integrate another provider’s API when that provider has stronger search capabilities in the relevant domain.
Translation scenarios: Although more than 80% of the translation capability comes from Zhipu AI itself, we may also integrate other capabilities for some extremely rare languages to ensure the best customer experience.
Q3: What stage has Zhipu AI’s globalization business reached, and what types of overseas customers does it primarily serve?
Wu Weijie: Zhipu AI’s globalization remains at a relatively early stage. We currently serve two main types of customers:
Foreign government bodies: We work with some countries, combining Zhipu AI’s model-training capability with local languages and content to help them build their own “sovereign large models.” Last year, we undertook a project of this kind with an ASEAN country. Such cooperation has also benefited substantially from China’s Belt and Road national strategy.
Chinese companies with global operations: We serve Chinese companies whose businesses span domestic and overseas markets. For example, our large models provide global translation for a leading internet content-community platform, supporting that platform’s international business.
Q4: When serving Chinese companies with global operations, what does Zhipu AI rely on to win their business?
Wu Weijie: Their choice is very straightforward and comes down to four Chinese characters: “price and results.”
Results first: Model performance must be strong enough to compare with products from peers such as OpenAI and Claude in overseas scenarios, or at least reach a similar level. If performance falls short, there is no basis for discussing value for money.
Price advantage: Once performance has met the required standard, we provide a competitive price.
Technology and Market Competition
Q5: Compared with leading overseas models, such as OpenAI’s models, how does Zhipu AI perform in terms of value for money?
Wu Weijie: Zhipu AI is not simply cheaper.
Cost: In most cases, Zhipu AI’s models cost approximately one-tenth as much as leading overseas models.
Performance: Wu noted that the GLM-4.5 model released some time ago reached global open-source-model SOTA, meaning State-of-the-Art, performance in Agent capabilities, Reasoning, and coding within the first two weeks after release. He believes top model companies are currently engaged in a pattern of alternating advances and alternating SOTA leadership.
Q6: How do you view the present technological gap between Chinese large models and the world’s leading models, such as OpenAI?
Wu Weijie: The actual gap may be longer than the “three to six months” often cited externally. He divides Chinese model companies into two groups:
Companies that rely on “distillation”: Many model companies train primarily by distilling data from leading overseas models. If that is all they do, they may never close the gap, because they are always following.
Companies that insist on proprietary pretraining: Companies such as Zhipu AI may also reference and use some data, but more importantly, they invest substantial cost and effort in foundational pretraining, or pre-train, work. Wu believes that only sustained investment in these fundamental capabilities can create a genuine possibility of surpassing overseas companies.
Q7: Claude recently announced that it would stop serving users in China. How did Zhipu AI respond to this market change?
Wu Weijie: We “received” this traffic rather than “piggybacking” on it. Zhipu AI seized the opportunity:
Becoming an alternative: After evaluation, many technology companies found GLM-4.5 to be an excellent alternative to Claude, especially as a central orchestration model or programming assistant. Wu said that several very well-known companies would soon officially announce that they had switched from Claude to GLM-4.5.
Launching a targeted product: Zhipu AI quickly launched the corresponding “GLM code” product. On the night it went live, it generated an unexpected amount of subscription revenue despite receiving no promotion.
Q8: How does Zhipu AI position a consumer application such as Zhipu Qingyan, and how much revenue does it contribute?
Wu Weijie: Consumer applications currently serve more as a window for showcasing and experiencing model capabilities, and remain in an “incubation and experimentation” stage. He sees the enterprise market as the future priority because improvements in model capabilities—such as better accuracy and lower hallucination rates—can create tangible productivity value for enterprise customers, while consumers perceive those improvements less strongly.
Zhipu AI’s overall revenue still comes primarily from enterprise customers. Consumer products contribute relatively little, within a single-digit percentage range. The main reason is that domestic consumers generally have not developed a habit of paying for AI applications.
The Industry’s Future and Advice
Q9: What advice do you have for individual entrepreneurs or small teams in AI that plan to expand overseas?
Wu Weijie: Small entrepreneurs should not build general-purpose platforms. They should focus on “extremely vertical application scenarios.”
Direction: He believes entertainment, such as AI companionship, and productivity tools are two areas with significant opportunity.
Team structure: One relatively successful model is to keep research and development staff in China to benefit from the country’s industrial advantages, while placing operations and marketing personnel overseas so they can better understand and integrate into the local market.
Path to success: He has observed that Chinese entrepreneurs are more likely to succeed overseas with tool- and platform-based applications, while ventures that require deep integration into local culture face enormous challenges.
Q10: How do you believe AI will change China’s enterprise-services market, and what form will it take in the future?
Wu Weijie: Large models may completely change China’s current situation of having “only an enterprise-services market, not an enterprise-software market.”
Change in payment model: Future enterprise services, especially at the application layer, will most likely shift toward payment based on results and outcomes.
An M-shaped market structure: The future enterprise-services market will polarize.
At one end will be companies serving very large customers. These companies will still need to invest human labor to provide the “emotional value” of companionship and responsiveness.
At the other end will be companies serving small and midsize businesses, or SMB companies. These will be highly automated platform companies with extremely small teams, possibly fewer than 10 people.
Service providers in the middle will disappear.