Zhipu는 AI 물량 축적 흐름 속에서 컴퓨팅 용량과 하드웨어 비축을 확보하기 위해 50억 달러 규모의 기술 장부를 편성한다.

The competitive focus of the AI industry has shifted from “can we build it?” to “can we continuously iterate and consistently push boundaries?” The former may stem from a flash of inspiration, but the latter hinges on the accumulation of organizational maturity, product strategy, capital reserves, and long-term vision. Capital, in particular, serves as the essential “ammunition and granary” supporting all other productivity factors. This reality underpins the recent fundraising rounds by leading manufacturers including Zhipu, ByteDance, Alibaba, and Tencent. On the evening of September 13, Zhipu announced the completion of approximately $5 billion in financing, comprising roughly $2 billion in equity placement and $3 billion in convertible bonds. The proceeds are primarily earmarked for model research and development and compute infrastructure. Following the announcement, pressing questions emerged across capital markets and the AI community: Why does Zhipu require this capital? Where will it be deployed? And what strategic advantage will it secure?

The winning factor in the ongoing AI marathon has fundamentally changed. During the first phase, model capabilities were largely attributed to algorithmic breakthroughs. In the second phase, the true differentiator is the ability to stably reproduce those breakthroughs repeatedly. This demands a sophisticated engineering system: experiment scheduling, data-to-compute ratios, and underlying infrastructure cluster orchestration. These operational details rarely appear on benchmark leaderboards, yet they are decisive for sustaining model iteration. While continuous iteration is now industry consensus, few companies can execute it consistently over the long term. Consequently, pacing itself has become a measure of capability. Earlier this year, Google previewed the imminent launch of its flagship Gemini 3.5 Pro at the I/O conference, only to face repeated delays. According to multiple overseas media reports, the postponements stemmed from shortcomings in coding and Agent capabilities. Falling behind on iteration rhythm temporarily made Google a focal point of industry debate in the latter half of the year. In contrast, several domestic AI firms have maintained steadier cadences. Zhipu, for instance, sequentially deployed GLM-5, GLM-5.1, GLM-5.2, and GLM-5.3 between February and August, averaging a major upgrade every two months. Prior releases, including GLM-4.6 and GLM-4.7, adhered to the same rhythm. According to third-party analytics firm Artificial Analysis, over the past eleven months, Zhipu’s flagship model scored on the Composite Intelligence Index rose from 32 to 60, firmly placing it within the global frontier tier. Multiple AI researchers have noted this trajectory to AI Technology Review, observing that such rapid iteration and capability escalation are uncommon in the sector. They emphasize that while isolated high benchmark scores can result from fortunate hyperparameter tuning, sustained multi-generation improvements require robust R&D capacity, organizational efficiency, and engineered systems. “Currently, the bottleneck in large model iteration is rarely ‘whether we can conceive an idea,’ but rather ‘whether we can transform that idea into reproducible results on schedule,’” one researcher told AI Technology Review. “Zhipu’s consistent on-time deliveries indicate that its production planning and engineering capabilities are firmly established.” At least based on public information, Zhipu’s development chain remains uninterrupted. The company’s next move is to leverage ample capital to sustain this momentum, maintain, and potentially lead the evolutionary trajectory of large language models.

Replenishing capital reserves is a priority for more than just Zhipu. In early September, ByteDance reportedly secured a syndicated loan of approximately $29.6 billion directed toward data centers and AI infrastructure. In August, Alibaba completed a new share placement worth HK$80 billion, primarily allocated to global compute resources and AI data centers. Tencent issued nearly $4.7 billion in bonds in June, while other major tech firms like Baidu, Meituan, and Kuaishou have been reported queuing in the bond market. Overseas players are equally active. SoftBank finalized a $40 billion bridge financing round in March to support OpenAI’s super-ecosystem expansion. Furthermore, multiple research institutions project that the combined capital expenditure of the world’s top nine cloud service providers will exceed $886.7 billion by 2026. However, raising capital alone is insufficient. The market closely monitors how these funds will be allocated, as deployment strategies will determine who secures the initiative in the next wave of technological generational gaps. According to Zhipu’s financing announcement, the net proceeds from the equity placement and convertible bonds total approximately HK$39.3 billion. Roughly 60 percent, or HK$23.5 billion, will be directed toward the next-generation GLM architecture, a fully self-training system, and large-scale training, inference, and compute infrastructure. “The dominance of convertible bonds with a premium indicates a preference for long-term capital,” a securities analyst tracking the AI sector told AI Technology Review. “Zero coupons mean the company avoids immediate interest payments, while conversion value is deferred, making it highly suitable for firms with extended R&D cycles.” This clearly signals Zhipu’s long-cycle orientation. Such confidence is largely rooted in a proven commercialization loop. Recent interim financial reports show H1 revenue reached RMB 954 million, a 399.7 percent year-over-year increase. Open platform and API revenue accounted for RMB 825 million, surging 2,736 percent year-over-year, and its share of total revenue jumped from 26.3 percent at the end of 2025 to 86.5 percent. Cloud-based billing services have become the revenue backbone, reducing reliance on project-based delivery. With the foundation solidified, the focus shifts to advancing frontier R&D. Based on the announcement, the budget for next-generation GLM development targets several critical technical challenges: expanding effective scale on the base model, advancing end-to-end unified modeling for ultra-long contexts and native multimodality; increasing effective compute depth for reasoning without significantly raising inference costs, thereby strengthening long-chain reasoning and self-correction; ramping up pre-training, mid-training, and post-training while expanding the exploration space for long-horizon reinforcement learning; and constructing task sandboxes that closely mirror complex real-world engineering environments. From an R&D perspective, these initiatives consolidate into three core directions. First, training scale. Scaling laws remain the primary defensive line for top AI companies. Scaling iterations involve not only enlarging base parameters but also deepening training compute, reasoning compute, and environmental complexity. Zhipu’s August release of GLM-5.3 demonstrates this: by expanding long-horizon task environments and post-training scale, its end-to-end task completion rate increased by over 50 percent, proving that the current scale dividend has not yet plateaued. Second, experimental continuity. Large model training is highly sensitive to interruptions. Historically, GPU scarcity forced serial experiment scheduling, prolonging feedback cycles. Abundant compute removes this bottleneck, allowing iteration rhythms to evolve from bi-monthly milestones to denser, agile cycles. Speed itself is a barrier; capability gaps often reflect how many high-quality “hypothesis-verification-correction” closed loops a team completes annually. Third, pipeline-oriented R&D systems. The funding covers training data generation and filtering, task environment construction, long-horizon reasoning, and underlying compute optimization. Collectively, these investments target the entire training pipeline, creating value that extends far beyond any single model. Swapping base models is a periodic action; pipelines are daily accumulated and optimized engineering assets. This is the core value of deploying large capital into infrastructure. “Task environment construction is the most undervalued component in the pipeline,” a large model algorithm expert previously noted. “It does not directly reflect in single-point benchmarks, but it dictates the ceiling for long-horizon complex capabilities. For models to master multi-day real-world business logic, the training side must first decompose tasks into executable, verifiable, and reproducible environments, requiring integrated systems for clusters, storage, documentation, and codebases.” Across top industry teams, pipeline construction is evident this year, with differentiation lying primarily in investment intensity and execution speed. Synthesizing these directions, Zhipu aims to significantly broaden its exploration space across training scale, experiment frequency, and engineering systems.

If expanding compute and pipelines addresses immediate needs, the most forward-looking technical signal in this financing round is Zhipu’s explicit pursuit of “Fully Self-Training” for next-generation architectures. Under relatively fixed frameworks and compute constraints, large model training revolves around three pillars: data, environments, and infrastructure tuning. All three are increasingly being automated by AI. On the data front, the direct catalyst for “self-producing and self-training” is the approaching ceiling of available datasets. Industry consensus holds that both manually annotated high-quality data and publicly accessible corpora are nearing saturation. Next-generation frontier models require trillions of high-precision incremental tokens, making model-driven generation and filtering the most discussed pathway. Zhipu’s stated approach involves cross-validation: one model generates content while another reviews and validates it. Only data passing adversarial checks enters the training set, shifting the pipeline from procurement to self-production. However, risks persist. “The challenge of letting models train models lies in verifying the correctness of their outputs,” an expert close to Zhipu’s R&D team explained. “Generation and review must be calibrated simultaneously; otherwise, both sides risk drifting together.” The industry debate over synthetic data continues, with no definitive conclusion on whether over-reliance on model-generated data causes performance degradation. In task environments, AI-driven construction is motivated by prohibitively high setup costs. Models require environments mirroring real expert workflows, where closer alignment yields higher-value training signals. Allowing models to extract, generate, verify, and refine processes within real environments is one method to reduce these costs. Infrastructure optimization faces a different hurdle: experience scarcity. Tasks like system scheduling, inference path design, and operator efficiency once relied on veteran intuition but are now being penetrated by AI. Public data indicates that GLM-5.3-powered infrastructure agents are already assisting engineers in optimizing operators, diagnosing performance bottlenecks, and improving deployment stacks. Leveraging previous-generation models to train next-generation ones represents a key technical route following Zhipu’s “Touch High” strategy unveiled in July. This approach is not isolated. Throughout the year, leading labs have increased the compute share allocated to post-training and reinforcement learning, exploring model-generated data and automated training systems. Overseas discussions often frame this as automating the research workflow itself. Zhipu aligns closely with these trends but distinguishes itself by explicitly embedding “fully self-training” into its public strategic narrative. Yet, self-training is exceptionally capital-intensive. Self-produced data, newly built environments, and iterative infrastructure tuning all demand massive compute, high talent density, and extensive engineering overhead. Capital is merely one variable. Time windows and competitor iteration speeds can alter outcomes. “Whether this path succeeds will likely become clear within two to three years,” the R&D expert cautioned. “It is too early to draw conclusions now.”

Within the industry, this financing round is widely viewed as the launchpad for Zhipu’s subsequent capability leap. The critical observation window remains the next twelve months. Three key developments warrant close monitoring. First, the release cadence and magnitude of capability advancement for the next-generation GLM. While GLM-5.3 demonstrated the effects of post-training scaling, any architectural changes in the base model will carry more informational weight than benchmark scores alone. Second, tangible progress on fully self-training. Whether data production scales autonomously, task environments become systematic, and infrastructure tuning becomes model-led will not be immediately apparent from a single product launch but will be embedded in engineering execution. Third, compute dynamics. During the anonymous testing phase of GLM-5.3-Flash, all traffic was routed through domestic chips, with a cluster of 100,000 domestic accelerators handling daily volumes exceeding one hundred trillion tokens. The pace of subsequent domestic compute procurement and partnerships will directly dictate how aggressively experimental scale can be expanded. Ultimately, large model competition boils down to a company’s ability to sustainably train progressively stronger models. Adequate capital ensures the engine operates at higher capacity. How the next chapter unfolds remains to be seen.