Zhipu's transformation to "China's Anthropic": Pricing power, scale effects, and transformation speed are key challenges

When Zhipu released its first financial report after going public, the market found that this company—once branded as “China’s OpenAI”—was quietly adjusting its strategic direction. The financial report showed that in 2025, the company’s revenue reached 720 million yuan, doubling from the previous year, but net profit recorded a loss of 4.718 billion yuan, meaning a loss of 5 yuan for every 1 yuan earned. Behind these figures lies a key turning point in the large-model industry: moving from a technical contest toward commercial deployment.

Changes in revenue composition reveal the deeper logic behind the strategic shift. The share of revenue from localized deployment fell from a high level to 73.7%, while cloud deployment rose from 15.5% to 26.3%. Although the gross margin of cloud services is only 18.9%, far lower than 48.8% for localized deployment, the company still set up two subsidiaries—Beijing Zhiyuan Chengzhang Technology Co., Ltd. and Hangzhou ZhiFu Qianhang Technology Co., Ltd.—to keep R&D core work in the parent company, while gradually shifting high-value delivery businesses such as enterprise solutions and private deployment to the subsidiaries. In essence, this structural adjustment moves the growth engine away from heavy-asset deployment and toward lightweight Token sales.

The ebb and flow between R&D expenditures and capital expenditures confirms the fine-tuning of cost control. In 2025, R&D expenses reached 3.18 billion yuan, 4.4 times revenue, with compute costs accounting for more than 70%. Meanwhile, capital expenditures dropped sharply from 460 million yuan to 74.70 million yuan, a reduction of more than 80%. The company changed its GPU leasing model, shifting to on-demand procurement of compute power. While this improves flexibility, it also plants concerns about long-term costs. The financial report shows cash on hand of 2.26 billion yuan; combined with 4.5 billion yuan raised through the IPO and 5.2 billion yuan in unused bank financing, at the current monthly burn rate of more than 300 million yuan, the cash reserves can still support the business for several years.

While the industry has largely fallen into price wars, Zhipu became the first domestic large-model company bold enough to raise prices. In the first quarter of 2026, after an 83% increase in API prices, usage volume did not fall—instead, it rose by 400%. Specifically, in February, the initial discount for the GLM Coding Plan was canceled, and package prices increased by more than 30%; in March, the GLM-5-Turbo model launched with another 20% increase in API prices. Although the “Lobster” package offers a monthly card for 39 yuan including 35 million Tokens, and an advanced version for 99 yuan including 100 million Tokens, the market still votes with real money. As of March 2026, registered users exceeded 4 million, and MaaS annual recurring revenue (ARR) reached 1.7 billion yuan, growing 60-fold within 12 months.

The confidence behind this “counter-cyclical price increase” comes from benchmarking Anthropic’s business logic. As a growth benchmark in AI, Anthropic increased its ARR from $1 billion to $19 billion in 15 months, with 80% of revenue coming from enterprise-level APIs, and per-user monetization efficiency that is 8 times that of ChatGPT. At the earnings call, Zhipu CEO Zhang Peng said plainly: “When the model is strong enough, API is the best business model.” But the gap in reality is significant: as of April 2026, Anthropic’s ARR had reached $30 billion, its valuation was $380 billion, and its PS multiple was about 12x; in contrast, Zhipu’s market capitalization is about HKD 350 billion, with 2025 revenue of 724 million yuan and a PS multiple as high as 450x. This valuation premium comes both from the scarcity of AI-related assets among Hong Kong-listed stocks and from the market’s expectations for a “China’s Anthropic” narrative.

To support this narrative, Zhipu needs to overcome three challenges. First is the sustainability of pricing power. Even though demand has not declined after the price hikes, the financial report does not disclose the specific breakdown of API revenue in terms of channel distribution, project delivery, and platformization capabilities. In a context where models iterate about every 3 months, how to respond to competitors such as Alibaba and others that are both customers and rivals still has no clear answer. Second is breaking through scale effects. In Zhipu’s formula—“intelligent upper limit × Token consumption scale”—the intelligent upper limit depends on spending on R&D, and whether domestic chips can approach the efficiency of NVIDIA directly affects the cost curve. If revenue maintains compound growth of over 130% while R&D expense growth slows down, theoretically the R&D expense ratio could be reduced to a reasonable level within 2–3 years. Finally is the pace of transformation. Localized deployment still accounts for 70% of revenue, but that business has declining gross margins and heavy delivery requirements. Zhipu needs to quickly increase the proportion of cloud API to truly realize a lightweight model of “collecting fees through Tokens.”

Internationalization has become a new growth driver. Over the past year, Zhipu expanded cooperation with Middle Eastern and Southeast Asian countries by “taking Tokens overseas.” Enterprise API services cover more than 50 countries and regions, and the share of overseas revenue rose from 5% in 2024 to 35% in 2026. This cross-border output model essentially expands Token consumption scale through technology licensing, providing incremental inputs to the variables in the business value formula.

Currently, the Agent framework track has become the focal point of competition. Zhipu, DeepSeek, Kimi, Alibaba, and other companies all have initiatives in this area, trying to define industry standards and gain an early advantage. Historical experience shows that once the application layer is proven, scale effects can be unleashed at the infrastructure layer. For large-model companies, the current stage is not only about verifying model capabilities, but also about cultivating future revenue demand through application scenarios. And the prerequisite for all of this is sustained technological leadership and engineering capability. Behind it, the amount of capital investment required is still a core question Zhipu must answer.