Going public is not the finish line; Zhipu’s real test has just begun.

Among China’s six rising stars in AI large models, we’ve already analyzed three—StepFun, MiniMax, and Kimi. Following a comprehensive approach, this piece focuses on Zhipu. Let’s start with Zhipu’s latest move: In mid-March, it released GLM-5-Turbo, deeply optimized for OpenClaw “Lobster” scenarios, while simultaneously raising the API price for this model by 20%.

To be honest, I was taken aback when I saw this move. After all, the dominant trend in China’s large model industry over the past two years has been price wars and competing over free usage quotas. If you cut your API prices by half, I’ll launch a permanently free basic version—companies have been eager to push prices to rock bottom. This marks the second time in six months that Zhipu has raised prices against the trend. The previous hike was a 30% increase for the Coding Plan package in February, and now another rise—yet surprisingly, many developers and enterprise clients are still on board.

Some in my social circle have remarked that for many, the first large model they ever ran on their own computers was ChatGLM-6B. In the blink of an eye, this company has become the 'World’s First Large Model IPO,' standing at a critical crossroads that will determine its future.

That’s no exaggeration. From ChatGLM-6B’s debut in 2023 to its Hong Kong Stock Exchange listing in January 2026, Zhipu transformed from a research project in a Tsinghua University lab into a top-tier player among China’s independent large model vendors in just three years. However, less than two months after the celebratory listing, the industry’s rules have changed. The frenzy of the 'Hundred Models Battle' has faded, and the large model industry has shifted from the first half, focused on 'can it be done,' to the second half, centered on 'can it make money.'

For Zhipu, going public was never the endpoint but rather a true turning point. Looking ahead lies the imperative of commercialization and scalable profitability. Looking around, it faces dimensionality-reducing pressure from internet giants and fierce competition from peers in the same tier. Looking back, the path it carved out over the past three years—relying on technological first-mover advantage and capital support—can no longer sustain growth in the second half.

01

From Lab to Listing: Zhipu’s First Half Story

Many people’s first impression of Zhipu comes from ChatGLM-6B.

Back then, ChatGPT had just taken the world by storm. In China’s developer circles, half were scrambling to find tutorials to register for OpenAI accounts, while the other half were figuring out how to run a functional large model on their own computers. However, the open-source models available at the time were either English-dominated with poor Chinese support or required massive computational resources that ordinary developers couldn’t access without professional hardware, leaving most unable to participate.

At that critical moment, ChatGLM-6B emerged. With 6 billion parameters, it could run smoothly on a consumer-grade graphics card with 8GB of VRAM and demonstrated far superior understanding of Chinese contexts compared to contemporaneous open-source models. I still remember that night—several tech groups I’m in were flooded with screenshots of local deployments. Some said they no longer had to stare at walls of English error messages, while others remarked that this was the first time they’d truly 'touched' a large model.

This open-source model built Zhipu’s initial reputation among developers and positioned the company—with its strong Tsinghua pedigree—at the first key inflection point of China’s large model wave. Zhipu’s core team hails from Tsinghua University’s Knowledge Engineering Lab, led by Professor Tang Jie of the Computer Science Department. This nearly 30-year-old lab has accumulated deep research achievements in natural language processing and knowledge graphs, and this academic technical foundation became Zhipu’s most critical initial competitive edge.

Over the next three years, Zhipu hit nearly every key milestone in China’s large model industry.

In August 2023, Zhipu launched its consumer product, Zhipu Qingyan, becoming one of the first large model products in China to pass generative AI service filing (regulatory approval). In 2024, it released the video generation model Qingying and the text-to-image model CogView3, filling gaps in multimodal capabilities. The GLM-4 series, launched in 2025, ranked first domestically and globally among open-source models in 12 authoritative benchmarks, with coding abilities tied for first globally with models from OpenAI and Anthropic in blind tests. GLM-5, released in February of this year, achieved native compatibility with fully domestic computing power, reducing deployment costs by 50% in long-sequence processing scenarios.

Continuous technological breakthroughs made Zhipu a focal point for capital. Prior to going public, Zhipu had completed 16 funding rounds, raising over 16 billion yuan in total. Its investors included state-backed industrial capital, top-tier financial capital like Sequoia and Hillhouse, and internet giants like Tencent, Alibaba, and Meituan. At its peak, Zhipu’s valuation reached 40 billion yuan, setting a new record for China’s large model startups.

On January 8 this year, Zhipu officially listed on the Hong Kong Stock Exchange, becoming the 'World’s First Large Model IPO.' This move not only brought substantial capital but also, more importantly, secured a critical ticket to capital markets amid an accelerating industry shakeout. After all, the large model industry is notoriously capital-intensive, and having access to public market financing provides more ammunition to endure competition.

Capital support ultimately translated into Zhipu’s commercial foundation. From its early stages, Zhipu targeted government and enterprise markets in the information technology innovation sector. Leveraging its Tsinghua-backed technical credibility, fully self-developed technology stack, and comprehensive compatibility with domestic chips, it quickly gained traction in sectors like government, finance, and energy, where data security and autonomous controllability are paramount. According to its prospectus, 85% of Zhipu’s revenue in the first half of 2025 came from localized deployment projects—customized private deployments of large models for government and enterprise clients.

Looking back, Zhipu’s first half was steady, precise, and fast. Relying on Tsinghua’s technical pedigree, it seized the first-mover advantage in the large model wave, built developer goodwill through open-source models, rapidly iterated technology with capital support, and strategically positioned itself in government and enterprise markets. Ultimately, it emerged from the 'Hundred Models Battle' as a representative of China’s independent large model vendors.

But industry changes always happen faster than expected. As the large model industry shifts from technological competition in the first half to commercialization in the second half, the advantages that fueled Zhipu’s past success now face entirely new tests.

02

At the Turning Point: Zhipu’s Unavoidable Challenges

I’ve discussed with many AI entrepreneurs, and we all agree: The rules of the large model industry have fundamentally changed in 2026.

In the first half, the core competition was about 'existence.' Could you build a functional large model? Could you match GPT’s capabilities? Could you rank well in authoritative benchmarks? These were key to survival. But in the second half, the focus has shifted to 'profitability.' No matter how strong your model is or how high its benchmark scores are, if it can’t be deployed in real-world scenarios, generate sustained revenue, or cover R&D and computing costs, it will ultimately be eliminated by the market.

This is the turning point Zhipu now faces. The growth logic that worked for the past three years has hit bottlenecks, and several core challenges have become undeniable.

The first challenge is the inherent shortcomings of its commercialization structure. With 85% of revenue coming from government and enterprise private deployments, this project-based model provides stable income and high client barriers but inherently struggles with scalability.

Private deployments aren’t about selling standardized products. Each client has unique needs, requiring dedicated teams for data integration, model fine-tuning, deployment maintenance, and ongoing iteration. The more projects undertaken, the more personnel are needed, and marginal costs remain high. Revenue growth depends on expanding team size, making it difficult to achieve snowballing, internet-style scalability.

More critically, competition in government and enterprise markets is intensifying.

Cloud providers like Baidu, Alibaba, and Tencent are leveraging their complete cloud service ecosystems, nationwide delivery teams, and lower computing costs to rapidly capture government and enterprise markets. In the first half of 2025, the domestic large model bidding market reached 6.4 billion yuan, surpassing the total for 2024, with over 60% of procurement coming from central and state-owned enterprises. Baidu Smart Cloud, with 510 million yuan in bids, became the 'bid king' in the government sector.

These cloud providers often bundle large model services with cloud resources and industry solutions, offering significant advantages in pricing and service capabilities. Zhipu’s foothold in government and enterprise markets is facing increasing pressure.

The second challenge is stagnant growth and traffic bottlenecks in the consumer market. While consumer products may not immediately generate substantial revenue, they provide massive amounts of real user interaction data—essential fuel for model iteration and key to shaping user perception. However, growth for Zhipu Qingyan has stalled. According to AI Product Rankings, from January to July 2025, Zhipu Qingyan’s monthly active users grew only from 7.02 million to 8.38 million, showing virtually no growth over six months.

In contrast, ByteDance’s Doubao reached 227 million monthly active users in the same year, while Alibaba’s Tongyi Qianwen and later entrants like DeepSeek also saw rapid growth. Compared to these leaders, Zhipu Qingyan’s gap has widened significantly.

The core issue is Zhipu’s lack of a native traffic entry point.

Doubao benefits from massive traffic support via Douyin (TikTok) and Jinri Toutiao, with direct access points within those apps. Tongyi Qianwen integrates with Alipay and Taobao’s ecosystems, naturally reaching users in shopping and payment scenarios. Zhipu Qingyan, however, relies solely on organic app store downloads and word-of-mouth, making it difficult to achieve user scale amid rising traffic acquisition costs.

The third challenge is widening losses and the R&D investment dilemma. The prospectus shows clear numbers: From 2022 to 2024, Zhipu’s adjusted net losses were 97 million yuan, 621 million yuan, and 2.466 billion yuan, respectively, totaling over 3.1 billion yuan in three years. By the first half of 2025, net losses further expanded to 1.752 billion yuan, burning nearly 300 million yuan per month.

The primary source of losses is massive R&D investment. Zhipu noted that computing service fees as a share of R&D expenses rapidly rose from 17% to over 70%, meaning a significant portion of R&D costs now goes toward purchasing computing power, synchronized with rapid growth in MaaS platform usage.

This is unavoidable. The large model industry evolves at breakneck speed—pause for a few months, and competitors will quickly surpass you. To maintain technological competitiveness, substantial R&D investment is essential, requiring burning capital on computing power and R&D teams.

But the situation changes post-IPO. In private markets, investors tolerate long-term burning if technological leadership and valuation growth are maintained. However, public market investors demand clear profitability and won’t accept unlimited burning. This creates a dilemma for Zhipu: balancing necessary R&D investment with mounting profitability pressure is the question it must answer at this turning point.

The fourth challenge is dual pressure from industry competition.

China’s large model market now has a clear two-tier structure: The first tier consists of internet giants like Alibaba, ByteDance, Tencent, and Baidu, which have their own computing infrastructure, massive scenario data, and natural traffic entry points, giving them inherent advantages in cost control and scenario landing. The remaining market share is divided among independent large model vendors like Zhipu, MiniMax, and Yuezhi Anmian, with Matthew effects intensifying—the top players dominate most of the independent track (market segment).

Zhipu now faces dimensionality-reducing attacks from internet giants ahead and intense competition from peers behind. Relying on scale advantages, internet giants are driving down large model API prices, putting extreme pressure on Zhipu’s API business. Meanwhile, peers are rapidly gaining ground—MiniMax is steadily climbing market share through global expansion and consumer product capabilities, while Yuezhi Anmian has built technical barriers in long-context scenarios, capturing significant users in professional office settings. In this competitive environment, Zhipu risks falling behind with any misstep.

03

Breaking Through in the Second Half: Zhipu’s Unique Path

Zhipu at the turning point isn’t without opportunities. Recent strategic adjustments clearly show Zhipu seeking its own growth path for the second half.

The most critical change is commercialization structure transformation. Zhipu is gradually reducing reliance on project-based private deployments and shifting toward standardized MaaS services. While private deployments offer stable revenue, they’re difficult to scale. API services, however, are standardized—clients pay based on usage volume without needing self-deployment or maintenance, and Zhipu avoids dedicating custom teams to each client. Marginal costs decline rapidly as usage scales, making this key to achieving scalable profitability.

According to CICC research, by early 2026, Zhipu’s annualized API-related revenue had reached nearly 600 million yuan, achieving tens of times year-over-year growth.

The two price hikes in six months fundamentally reflect this transformation. While the industry competes on low prices, Zhipu focuses on optimizing capabilities for specific scenarios, gaining pricing power through tangible value improvements.

For example, the February price hike for the Coding Plan package targeted code development scenarios with deep optimizations, matching global top models in code generation, debugging, and completion. Many developers and small tech firms are willing to pay more for superior capabilities. Similarly, GLM-5-Turbo is optimized for OpenClaw “Lobster” scenarios, significantly improving core agent capabilities like tool invocation and long-chain execution, attracting enterprise clients with real needs despite the price increase.

Beyond API services, Zhipu has launched “Lobster Packages” for individuals and enterprises, shifting from pure API sales to bundled “digital employee” subscription services that combine model capabilities, tool invocation, and scenario solutions. This not only enhances user lifetime value but also creates a healthier commercialization model.

The second key strategic move is deepening barriers in the information technology innovation market to consolidate (consolidate) its foothold.

In the market of information technology innovation, independence and controllability are core and rigid requirements, which is precisely Zhipu's advantage. From the underlying architecture to training and inference, Zhipu's GLM series models are entirely self-developed, and have achieved deep adaptation with seven major domestic computing power platforms, including Huawei Ascend, Hygon, Cambricon, and Moore Threads, effectively reducing supply chain risks. In industries with the strongest demand for information technology innovation, such as government services, finance, and energy, this ability of full-stack self-development and comprehensive adaptation to domestic chips represents a core competitive edge and a barrier that large corporations find difficult to fully replace.

Meanwhile, Zhipu is also optimizing its government and enterprise service model, shifting from the original single privatized deployment to a platform-based service model of 'foundation + ecosystem'. For example, in the Zhuge Large Model project implemented in Chengdu, Zhipu not only provided a foundational model but also established the Western Industry Application Empowerment Platform, attracting local Chengdu ecosystem enterprises to develop industry-specific solutions based on the Zhuge Large Model, forming a complete ecosystem of 'foundational model - industry application - scenario implementation'. This model not only enhances customer loyalty but also opens up long-term revenue opportunities, eliminating the need to rely on earning hard money from individual projects.

The third important direction is global expansion to open up new growth spaces. Competition in the domestic market has reached a fever pitch, while overseas markets in Southeast Asia and the Middle East still have very low penetration rates for large models, offering tremendous growth potential. In early 2026, Zhipu secured a national-level AI project in Malaysia, becoming a landmark case for Chinese AI companies to go global across the entire chain.

The fourth differentiated strategy is early positioning in response to the trend of intelligent agents. The current large model industry has transitioned from the 1.0 era of general dialogue to the 2.0 era of intelligent agents capable of solving practical problems, with intelligent agents becoming the core direction for large model implementation. Starting from the GLM-4 series, Zhipu has focused on strengthening the model's tool invocation and instruction-following capabilities, natively integrating reasoning, coding, and intelligent agent capabilities.

The newly released GLM-5-Turbo has been specifically optimized for the OpenClaw scenario, becoming the world's first general-purpose large model deeply optimized for the lobster scenario.

I have always believed that the watershed moment Zhipu now faces is actually a common challenge for all independent large model companies in China. In the first half, relying on technological first-mover advantage, capital dividends, and the window of opportunity for domestic substitution, many startups emerged and secured their place in the market. However, in the second half, the competitive logic of the industry has fundamentally changed. It is no longer about who can create a large model, but who can continuously generate profits from it and who can build an ecosystem and scenario barriers that others cannot replicate.

Going public has provided Zhipu with sufficient capital and brand influence, but it has also clearly placed the pressure of profitability on the table. Whether it can adjust its commercialization structure from project-based to standardized services, whether it can find its differentiated positioning amid pressure from large corporations and the impact of open-source models, and whether it can translate its technological advantages into sustained profitability are all questions Zhipu must answer at this watershed moment.

Zhipu's answers will not only determine its own future but also point out a viable path for all independent large model companies in China. After all, in this rapidly changing industry, surviving the fierce competition among hundreds of models is already challenging; forging one's own path under pressure from large corporations is the real test.

— END —