Key Takeaways:
Zhipu AI reported H1 revenue of RMB 954 million, up 399.7% year over year, as cloud API revenue surged 2,735.7% to RMB 825 million.
The company said its commercialization path is shifting from one-time local deployments to usage-based cloud services, describing the evolution as "selling models, selling calls, selling subscriptions, and selling end-to-end task results."
Open platform and API revenue rose from RMB 29.1 million to RMB 825 million, lifting its share of total revenue from 15.2% to 86.5%. Enterprise general LLM revenue fell 54.6% to RMB 67.04 million. MaaS platform paid daily active users grew 603% since the start of the year, while token call volume increased more than 40 times.
The company raised HK$31.4 billion in a July placement at HK$1,588 per share to fund model training and compute infrastructure. Net loss narrowed 12.1% to RMB 2.07 billion, with R&D spending of RMB 2.13 billion — about 2.2 times revenue — reflecting the capital intensity of the AI race.
Overall gross margin fell to 26.4% from 50.0% as lower-margin cloud services replaced higher-margin local deployments. But the cloud business itself turned profitable: open platform and API gross margin improved from negative 0.4% to positive 24.6%, driven by a 101% rise in average API selling prices and an 80% reduction in unit token inference costs since January.
Enterprise agent business revenue grew 304.4% to RMB 55.56 million, though it remains a small share of total revenue. Sales and marketing expenses fell 14.8%, while general and administrative costs dropped 44.2%.
Model Iteration Shifts From Parameters to Post-Training
Zhipu AI released GLM-5, GLM-5.2, GLM-5.3, and GLM-5.3-Flash during the period. The company said GLM-5.3 achieved a 50% improvement in end-to-end coding completion rates over GLM-5.2 while maintaining identical architecture and parameter counts, pointing to a strategy focused on post-training and reinforcement learning rather than raw model size.
GLM-5.3-Flash, with 320 billion total parameters and 18 billion active parameters, uses sparse and linear attention architectures trained on 30 trillion multimodal tokens. Priced at roughly one-tenth of GLM-5.2, the model runs on a 100,000-chip domestic cluster with three times the end-to-end service performance of its initial baseline.
The company is also expanding beyond coding into agentic and collaborative work scenarios. Zhipu AI scored 84.5% on the CyberGym benchmark and completed 130 tasks on ExploitGym, with its security teams identifying 2,436 vulnerabilities, including 1,097 medium-to-high severity findings.
The results come as Chinese AI model providers including Alibaba's Qwen and Baidu's Ernie race to convert model capabilities into commercial revenue. Zhipu AI's cloud MaaS model is gaining traction, with API pricing power and cost efficiency improving in tandem. Investors will watch whether the company can sustain this growth while balancing R&D intensity against the path to profitability.
This article is for informational purposes only and does not constitute investment advice.
