- Net loss widened to 4.72 billion yuan in 2025, primarily due to a sharp increase in research and development costs.
- The startup's revenue more than doubled, reaching 724 million yuan in 2025, with cloud-based deployment business surging nearly threefold.
Zhipu, one of China's most prominent AI startups, reported a wider annual loss as it ramped up spending in the AI race and leaped toward "Agentic Engineering."
The company's net loss for 2025 stood at 4.72 billion yuan ($683 million), up 59.5% from a net loss of 2.958 billion yuan in 2024, underscoring the hefty price tag of the generative AI arms race.
However, excluding share-based compensation expenses, listing expenses, and other items, its adjusted net loss under Non-IFRS measures was 3.18 billion yuan, representing a year-on-year expansion of 29.1%.
The significant increase in net loss was mainly driven by the company's continuous and massive increase in R&D investments, according to its earnings report released on Tuesday.
Zhipu's R&D expenses in 2025 reached 3.18 billion yuan, an increase of 44.9% compared to 2024.
This was primarily due to increased staff costs resulting from the expansion of the R&D team and share-based compensation, as well as an increase in computing service fees paid to third-party computing power suppliers for the iteration of foundational models.
Despite the deteriorating bottom line, the AI startup's top-line revenue growth remained exceptionally strong, with remarkable achievements in commercialization.
Zhipu's total revenue for the full year of 2025 reached 724 million yuan, representing a staggering year-on-year growth rate of 131.9%.
The robust revenue growth was mainly driven by the enhanced capabilities of its foundational models (from GLM-4.5 to GLM-5) and the subsequent increase in model invocations.
In terms of revenue structure, Zhipu's business is undergoing significant structural changes:
- Cloud-based Deployment has become the strongest growth engine, with revenue surging 292.6% from 48.48 million yuan in 2024 to 190 million yuan, and its share of total revenue rising from 15.5% to 26.3%.
On-premise Deployment remains the largest revenue source, reaching 534 million yuan, a year-on-year increase of 102.3%, accounting for 73.7% of total revenue.
Broken down by core product lines, Enterprise-level general-purpose large models contributed 366 million yuan (50.4%), Open platform and API contributed 190 million yuan (26.3%), while the emerging Enterprise-level Agents business also reached 166 million yuan, a massive year-on-year increase of 248.8%.
However, the company's gross margin dropped from 56.3% in 2024 to 41.0% in 2025.
This was primarily due to the increased proportion of the cloud-based deployment business, as well as the investment of more delivery resources to meet customer needs, which led to a phased decline in the gross profit margin of the on-premise deployment business (from 66.0% to 48.8%).
Regarding business ecosystem and technological progress, Zhipu disclosed several stellar figures in the financial report:
- Developer Ecosystem: Relying on its native high-quality engineering reasoning capabilities, its GLM Coding Plan rapidly expanded globally, with paying developers exceeding 242,000.
- Bolstered by its technological leadership, Zhipu raised prices by 30% and eliminated first-purchase discounts in February 2026, yet the market still showed a trend of "rising volume and price."
- MaaS Platform: Relying on BigModel.cn, the platform surpassed 4 million registered users as of March 2026. Nine of the top ten internet companies in China have now deeply integrated GLM.
- Agent Explosion: Zhipu launched AutoGLM, the world's first mobile phone agent, and AutoClaw, China's first one-click installation agent. The "Claw Plan" launched in March 2026 surpassed 100,000 subscribers within just two days and topped 400,000 within 20 days.
- Computing Power and Domestic Substitution: Faced with the "computing power panic" where demand has outstripped supply since February 2026, Zhipu has increased its investment in "Day 0" adaptation for domestic chips.
- Through deep tuning at the underlying kernel level (such as Lightning Indexer and FlashComm technologies), the inference efficiency of the GLM series models on domestic chips is now comparable to that of top-tier global chips.
Zhipu went public on the Hong Kong Stock Exchange on January 8 this year, trading under the name Knowledge Atlas Technology JSC Ltd (Stock Code: 2513).
As of December 31, 2025, the company held cash and cash equivalents of 2.26 billion yuan.
The startup is attempting to maintain its technological lead in China's fiercely competitive AI market.
Its management emphasized in the financial report that the company's moat lies not in the accumulation of computing power, but in the foundational deconstruction of the essence of intelligence, aiming to become the infrastructure for enhancing the "Token Architecture Capability (TAC)" of the whole society.
($1 = 6.9068 yuan)
