Summary of Key Points
Zhipu’s revenue in the first half of 2026 soared by 4 times (to 954 million yuan), although it fell slightly short of expectations; however, the loss narrowed, and investment in research and development continued to increase (2.13 billion yuan). There was a significant change in the revenue structure: the proportion of the open platform/API increased from 15% to 86%, marking the entry into a scaled commercialization phase. The biggest highlight is the shift in its technical approach—instead of blindly increasing the number of model parameters, Zhipu focuses on “post-training” to enhance model capabilities while reducing costs (the price of GLM-5.3-Flash is even lower than that of its industry competitors). To cope with the pressure of high computing costs, Zhipu has taken measures such as building its own domestic chip data center, acquiring infrastructure companies, and optimizing its technical architecture to reduce costs. It also raised 31.4 billion Hong Kong dollars to expand its operations. The stock price initially rose but then fell, reflecting market skepticism about its strategy; however, the focus of industry competition has shifted from “model performance” to the ability to deploy models quickly and at low costs.
Detailed Analysis
1. Strong Revenue Growth, but Not Meeting Expectations? Don’t Worry—Future Growth Is Even Faster
Zhipu’s revenue for the first half of the year was 954 million yuan, a four-fold increase from the previous year, which is impressive. However, it fell short of the 1.3 billion yuan predicted by external institutions. Nevertheless, the management provided reassurance: as of the end of August, the annualized revenue (ARR) had reached 1.6 billion US dollars (approximately 11.2 billion yuan), with a single-month revenue of 133 million US dollars (about 900 million yuan) in August, nearly matching the total for the first half of the year. This indicates that growth will accelerate significantly in the second half of the year, and based on the latest figures, ARR could even exceed 2 billion US dollars. In short, although the initial targets were not met, the company’s momentum is strong, and there’s no need to worry about growth prospects.
2. Moving Away from Parameter Accumulation to “Post-Training”: Zhipu Is Changing the Game Rules
In the past, large model companies competed by having the largest number of parameters (for example, Kimi had 2.8 trillion parameters, and DeepSeek had 1.6 trillion). Zhipu, on the other hand, is taking a different approach. GLM-5.3 uses the same 744 billion parameters as GLM-5.0 from the previous year and enhances model capabilities through “post-training.” What is post-training? Imagine a newly manufactured car; post-training is like fine-tuning it (optimizing the engine, replacing tires, etc.) to make it run faster and more efficiently. Zhipu claims that with the same number of parameters, post-training can improve model performance by 50%. As a result, GLM-5.3-Flash not only outperforms previous versions but also costs one-tenth of GLM-5.2’s price and is even cheaper than DeepSeek’s most competitive models, making it a champion in both cost and efficiency.
This move signals to the industry that simply accumulating more parameters is not the most effective strategy; it’s more profitable to optimize existing models to their full potential.
3. High Computing Costs? Zhipu Reduces Them with Domestic Chips and Its Own Data Center
The biggest expense for large models is computing power (the cost of servers used for training and running models). Zhipu’s gross margin decreased from 50% to 26.4% mainly due to a six-fold increase in computing costs (sales costs rose from 95 million yuan to 700 million yuan). According to Barclays, model companies pay 35-40% of their revenue to cloud giants like Amazon and Microsoft. How does Zhipu address this? By taking control of its own resources:
- It built a 1GW-level data center (equivalent to the computing power of 1 million servers) using all domestic chips.
- It acquired AI infrastructure company Zhongke Jiahe to optimize its technical foundation.
- It adjusted its model architecture to ensure that domestic chips can operate efficiently.
The results are evident: the cost per token (each character processed by the model) has decreased by 80%, and the “computing power multiplier” (the revenue generated per unit of computing power investment) has increased by 14 times—meaning the same amount of investment now yields 14 times the previous revenue.
4. Model Companies Are Becoming “Heavier”: Buying Buildings, Acquiring Companies, Raising 30 Billion Yuan—Why?
Previously, model companies were often lightweight (renting computing power and having small teams). Now, Zhipu is becoming more “heavyweight”:
- It has 981 employees, twice the number of MiniMax’s staff.
- In May, it purchased the Beijing Diamond Building for 360 million yuan, and in June, it acquired Zhongke Jiahe for 290 million yuan.
- In July, it raised 31.4 billion Hong Kong dollars to fund research and development, expand computing capacity, and make additional acquisitions.
Why this shift? Because the competition in the large model industry has moved from the models themselves to the infrastructure. Renting computing power is not only expensive but can also be limiting (for example, if cloud providers run out of capacity). By building its own data center and acquiring infrastructure, Zhipu can ensure a stable supply of computing power and reduce long-term costs. This is a trend in the industry—model companies are transitioning from being lightweight to becoming more resource-intensive, similar to how internet companies build their own data centers.
5. Stock Price Fluctuations: The Focus of Industry Competition Has Shifted from “Performance” to “Deployment Capability”
On September 1, Zhipu’s stock price rose 5% before falling 1.34%, indicating market skepticism about its financial results and strategic direction. Some are optimistic about its growth and technology, while others are concerned about losses and competitive pressures. However, the industry’s focus has changed: instead of comparing model performance (such as test scores), the focus is on who can deploy models quickly and at low costs. Zhipu’s approach aligns with this new trend—instead of aiming for the largest number of parameters, it focuses on creating practical, cost-effective models that businesses and users can utilize.
Conclusion
Zhipu’s semi-annual report highlights two key changes in the large model industry: a shift in technical strategy from parameter accumulation to post-training and cost reduction, and a shift in business models from being lightweight to becoming more resource-intensive (with self-built computing power). Although there are short-term issues such as revenue not meeting expectations and declining gross margins, its approach aligns with industry trends. In the long run, only companies that can generate revenue and effectively deploy their models will be the true winners.