Summary of Key Points
Zhipu’s performance in the first half of 2026 is characterized by “high growth, narrowing losses, and structural optimization”: revenue surged nearly fourfold year-on-year (exceeding the total revenue of last year), although the company is still in the red (however, the loss is less than last year). The revenue structure has shifted from “selling model deployments” to “selling API services,” and the API business has begun to generate profits. The company addressed concerns about the “small size of model parameters,” emphasizing that model effectiveness is the result of multiple factors. Zhipu has achieved large-scale inference using domestically produced chips and is now focusing more on “effective computing power.” In the competitive landscape, the battle is not about short-term rankings but about the ability to deliver high levels of intelligence at low costs over the long term.
1. Revenue Soaring, but Still in the Red: Structural Change is More Important than Growth Rate
Zhipu’s revenue for the first half of the year was 954 million yuan, a year-on-year increase of 399.7% (more than double the annual revenue of last year), but the loss was 207.2 million yuan (12% less than the same period last year). Why is there a loss despite such substantial revenue growth? The key lies in the source of the revenue:
- Previously, the company mainly relied on “localized deployments” (installing models on customers’ servers for a one-time fee); now, 86.5% of the revenue comes from “open platforms and APIs” (making models available as interfaces for enterprises/developers to pay based on the number of calls).
- The gross profit margin of the API business has increased from -0.4% (a loss per sale) last year to 24.6% (starting to generate profits), indicating that this business model is successful. The company’s secretary to the board stated that the shift in revenue composition is more noteworthy, as APIs provide stable, recurring revenue, which will become increasingly profitable in the future.
2. Model Parameters Are Not the Whole Story: How Zhipu Responds to Criticisms of Small Parameters
Some have criticized Zhipu for only focusing on post-training and using small model parameter sizes. Founder Tang Jie responded pragmatically:
- The quality of a model cannot be judged solely by the number of parameters (for example, the number of parameters in large models); it also depends on the amount of training data, the stage of computing power used (pre-training, mid-training, post-training), and how the model is applied.
- For example, GLM-5.3 and GLM-5.2 have the same number of parameters, but GLM-5.3 performed better after a month of long-term training and reinforcement learning. He compared model optimization to adjusting multiple knobs; choosing post-training optimization does not mean other stages are unimportant, and further adjustments will be made in the future.
- With domestic training data amounting to approximately 30-50 trillion tokens and limited computing power, blindly increasing the number of parameters yields low marginal benefits (for instance, doubling the parameters may only improve performance by 5%). It’s better to invest in more effective areas.
3. The Art of Computing Power: Domestic Chips Are Enough, but “Effective” Power Is Key
Zhipu previously mentioned achieving large-scale inference with domestically produced chips. This time, two key points were added:
- The company obtains computing power from various sources: its own clusters, rentals, and purchases, covering the entire process from training to inference.
- The industry’s focus has shifted: the question is no longer whether domestic chips can run models, but whether using them is cost-effective.
- The importance of “effective computing power” was emphasized: it’s not just about the number of chips purchased, but whether they are properly installed, can be used stably, and can be effectively converted into data processing capabilities (e.g., generating a certain number of tokens). Simply stacking chips is useless; only effective computing power matters.
4. What Really Matters in Industry Competition: Not Short-Term Rankings, but “Sustainable Cost-Effectiveness and New Task Capabilities”
Chairman Liu Debing pointed out that models are updated rapidly, and what ranks first today can be surpassed tomorrow. True competitiveness lies in:
- Continuously delivering higher levels of intelligence at lower costs: for example, providing the same AI capabilities at a lower price or achieving better results for the same cost.
- Future pricing trends: the prices for common tasks (such as simple question-answering) will likely decrease, while models capable of handling new tasks (such as complex scientific research and industrial design) will retain high prices.
- Pricing power lies not in the number of tokens (e.g., the fee per call) but in what tasks the tokens can accomplish. If a model can solve problems that others cannot, customers are willing to pay a higher price.
Conclusion
Zhipu’s semi-annual report highlights that the growth of AI companies depends not only on revenue growth or model parameters but also on the health of the business structure, the practicality of the technology, the efficiency of computing power, and the ability to continuously create “useful value.” For consumers, this means that AI services will become more affordable in the future, but AI that can solve complex problems will remain valuable.