Zhipu AI는 ChatGLM 프론티어 모델의 운영사로서 Series C 펀딩을 통해 50억 달러를 조달했으며, 이를 이정표 축하보다는 경쟁을 위한 지속적 원동력으로 표현했다.
On September 13, Zhipu (HK: 02513) announced the completion of approximately $5 billion in financing, comprising approximately $2 billion in share placement and approximately $3 billion in convertible bond issuance. The fixed-price subsequent offering was priced at HK$714 per H-share, representing a discount of approximately 10.0% from the closing price of HK$793 per H-share on the previous trading day, raising approximately $2 billion. The convertible bonds were issued at 100.5% of par value, bearing zero coupon, with a fixed conversion premium of 25% over the placement price, raising approximately RMB 20.14 billion (approximately $3 billion). The key significance of this financing is to strengthen Zhipu's computational capacity to meet the demands of sustained R&D and rapid growth.
More intriguing than the amount itself is where the money will go. According to the announcement, this financing round will be deployed for next-generation GLM foundation models, a fully self-developed training system, and computational infrastructure development. In other words, this $5 billion is not champagne at a celebration—it's more like a supply station for the next leg of a long-distance race.
**Rapid Iteration**
An AI company releases a major model upgrade every two months. From GLM-5 in February 2026 to GLM-5.1 in April, GLM-5.2 in June, and GLM-5.3 in August—roughly one major upgrade every two months—this cadence is uncommon among Chinese large model competitors. What's even more striking is how GLM-5.3 was released: its foundation model is identical to GLM-5.2, without expansion in parameter scale or changes to the underlying architecture. All improvements came from the post-training phase via reinforcement learning.
Zhipu stated in a technical report: "We may still be far from unlocking the full capability ceiling of this foundation model." But the reality is that there is far more room to "extract" additional capabilities from an already-trained foundation model than the outside world imagines. Post-training has become the critical differentiator among models. In the training environment that Zhipu built for GLM-5.3, some tasks involve computational workloads equivalent to several days of continuous work by an engineer.
This technical path demands continuous, not one-time, consumption of capital and compute. Foundation model training is like a big bang; post-training resembles a long-running engine requiring sustained operation. Each iteration demands reconstructing task environments, running extensive reinforcement learning experiments, and validating model performance in real-world scenarios. Without sufficient compute reserves and experimental capacity, this path becomes impassable.
**How to Spend $5 Billion?**
$5 billion may seem substantial, but on the road to achieving AGI, it may still fall far short. Breaking down the $5 billion, approximately 60% comes from zero-coupon convertible bonds and approximately 40% from share placement. The convertible bonds feature a zero-coupon structure, issued at 100.5% of par value, with an initial conversion price of HK$892.50 per share, representing a 25% premium over the placement price of HK$714 per share; using the closing price before the announcement as reference, the conversion price carries a premium of approximately 12.55%. These two measures need separate interpretation—the former reflects the additional cost paid by investors to participate in the placement, while the latter measures the conversion price premium relative to the market price.
In plain terms, this means investors are not only willing to lend money to Zhipu interest-free but are also willing to wait for future conversion at a price significantly above market value. They're not betting on next quarter's earnings—they're betting on the company's long-term value years from now. The zero-coupon provision is equally favorable for the company; in an industry like model R&D where "investment comes first and returns come later," every fraction of foregone interest is a resource that can be directly deployed to training clusters. Moreover, the delayed conversion structure also controls the immediate dilution of share capital.
So where specifically will this money be spent? According to the announcement, the financing will cover research and development of next-generation GLM models, construction of a fully self-developed training system, and simultaneous enhancement of training and production inference infrastructure—including domestic chip adaptation, operator development, and compute scheduling to improve the effective output of computational resources.
Behind this somewhat technical description lies a simple truth: identical compute resources used well versus poorly can yield model capabilities differing by orders of magnitude. Chip adaptation determines whether hardware can run at full capacity; operator development determines how fast each computational step executes; compute scheduling determines whether a cluster of thousands of GPUs sits idle. When training costs are measured in hundreds of millions of dollars, improving efficiency in these engineering components by even a few percentage points yields astronomical savings. This is why "fully self-developed training system" and "computational infrastructure" are placed under the same budget line—the former sets the ceiling on model capability, the latter determines how efficiently you approach that ceiling.
Control over R&D priorities is another key to understanding how this capital will be deployed. One of the harshest realities in the large model industry is the scarcity of experimental windows: an important training experiment often requires occupying the entire cluster for weeks on end. Companies with abundant compute can deliberately schedule experiments and run multiple parallel tracks to test different technical hypotheses; compute-constrained companies can only wager on a single direction within limited windows, and one failure means losing an entire iteration cycle. This financing round expands Zhipu's bandwidth to allocate resources across multiple iteration cycles, transforming frontier exploration from a high-stakes bet where you "can only win, not lose."
Against a backdrop of weakened sentiment in Hong Kong's AI sector and continued volatility in global capital markets, this offering still achieved high-quality execution. During the pre-marketing phase, intended orders already covered the full offering size; following the opening of bookbuilding, the offering quickly attracted participation from hundreds of institutions. In the final allocation, the top 20 investors collectively received over 85% of the shares and over 88% of the convertible bonds, with approximately 30 long-term investors and approximately 20 technology-focused investment institutions. The high concentration of stock indicates relatively contained secondary market selling pressure and signals that this capital base—centered on fundamental research—is prepared to accompany the company through the next phase.
The stability of this relationship determines whether Zhipu can sustain the cycle of "financing—R&D—commercialization—re-financing." The arrival of $5 billion has given this cycle more generous working room over the next two to three years. The announcement projects that funds will be fully deployed by June 30, 2028, meaning Zhipu has approximately a two-year window to demonstrate that sustained technology investment can translate into sustainable commercial returns.
Globally, the financing pace among large model companies is accelerating. Anthropic completed a $65 billion Series H financing round in May 2026, bringing its post-money valuation to $965 billion. Zhipu's $5 billion is not striking at this scale, but considering it was completed by a Hong Kong-listed company through share placement and convertible bonds and comes just two months after a previous $4 billion placement, this cadence signals two things: the market has intense demand to allocate to scarce Chinese large model core assets, and Zhipu itself has a clear plan for the efficient deployment of capital.