大模型下半场怎么打？智谱悄悄花几亿“买”下了国家队

This article analyzes the competitive shifts in the large language model (LLM) industry as it enters its second phase. Its key takeaways are as follows:

1. Leading Chinese LLM developers are now widely facing computing power bottlenecks. Moonshot AI, developer of the Kimi chatbot, has urgently suspended new user subscriptions for its C-end service after user demand exceeded its cluster capacity limit. Six months ago, Zhipu AI also encountered insufficient computing power, forcing it to restrict new user access and suffer from inconsistent inference outputs and degraded user experience during periods of high traffic.

2. Two distinct competitive strategies have emerged in the industry: Moonshot AI focuses on advancing model capabilities, prioritizing larger parameter sizes and higher benchmark rankings. In contrast, Zhipu AI has quietly built out its underlying infrastructure, spending hundreds of millions of yuan to acquire中科加禾 (Zhongke Jiahe), a compiler developer incubated by the Chinese Academy of Sciences. It has also invested across the entire AI infrastructure value chain via its affiliated funds, and even purchased an entire office building to support long-term development and prepare for in-house chip development.

3. It is now an industry consensus that LLM competition has shifted from a race for algorithm performance to a competition for underlying engineering capabilities. Stable, low-cost inference output is the core of long-term competitiveness.

This article outlines the latest competitive trends and shifting user demands in the LLM industry, offering valuable insights for LLM brand operators:

1. Shifting consumer trends and user behavior: LLMs have moved from an early-stage capability demonstration phase to large-scale commercial service. Users no longer chase large parameter sizes and benchmark rankings alone; they now prioritize stable output, predictable inference latency and cost-effectiveness. Insufficient computing power that degrades user experience directly triggers customer dissatisfaction and damages brand reputation.

2. Reference for pricing and competition: Zhipu AI’s GLM-5.2 matches and even outperforms top-tier closed-source LLMs from the U.S. while cutting operating costs to 1/6 to 1/7 of the U.S. model’s level. This "high performance at low cost" strategy drove a 27-fold surge in average daily API calls in its first week post-launch, making it a valuable competitive blueprint for LLM brands.

3. Risk alert: U.S. export controls have created supply chain uncertainty for domestic Chinese LLMs. Preparing domestic computing power adaptation capabilities in advance is essential to guarantee long-term stable business operations.

This article clearly outlines the current opportunities and risks in the LLM industry, offering practical insights for AI industry practitioners:

1. Risk alert: The biggest bottleneck facing the LLM industry today is insufficient supply of underlying computing infrastructure. Even leading companies have to restrict new user access and damage existing user experience when hit by explosive traffic growth due to computing power shortages. Small and medium-sized practitioners must plan computing capacity in advance to avoid this risk.

2. New growth opportunities: As LLMs enter the large-scale deployment phase, underlying AI infrastructure has become an industry-wide rigid demand. Clear market gaps exist across the entire space, from computing cluster construction, inference performance optimization and cross-chip adaptation to compiler development, making this a promising new growth track for entry.

3. Referential development model: Zhipu AI’s approach of investing across the upstream and downstream value chain via its industry fund to lock in industrial resources is worth learning from. Against the broader trend of domestic computing substitution, businesses focused on domestic chip adaptation stand to benefit from clear policy and market dividends.

This article’s analysis of LLM industry development paths offers key insights for manufacturing facilities pursuing digital transformation and new business opportunities:

1. Insights for digital transformation: AI deployment has now moved from proof-of-concept to large-scale application, and underlying infrastructure capability is the core guarantee of successful implementation. When advancing AI-driven digital transformation, factories should not focus solely on model parameter sizes and benchmark scores. They must also prioritize engineering capabilities such as domestic computing power adaptation and inference stability, to avoid the common pitfall of models that look impressive on paper but perform unreliably and cost too much in actual use.

2. New business opportunities: China’s domestic AI chip ecosystem is currently fragmented, with incompatible architectures across different chip vendors, leading adaptation costs to grow exponentially as the number of chip types increases. This has become an industry-wide pain point. Manufacturers with relevant technical foundations can enter the supporting field of AI underlying infrastructure to capitalize on the domestic substitution trend.

3. Reference for long-term development strategy: Zhipu AI’s approach of proactively building core end-to-end capabilities offers a useful blueprint for long-term factory growth. Preemptively building core capabilities allows businesses to capture future growth opportunities amid market uncertainty.

This article clarifies the current development trend of the LLM industry and identifies key customer pain points and entry opportunities for service providers, with key takeaways as follows:

1. Industry development trend: Competition in the LLM industry’s second phase has shifted from algorithm performance to underlying engineering capabilities. The market’s core demand is for stable, low-cost, high-throughput token processing systems, and AI infrastructure-related services are entering a period of rapid growth.

2. Core customer pain points: The top challenge for LLM developers today is unstable computing power supply. The fragmented domestic chip ecosystem, with incompatible instruction sets and architectures across vendors, pushes model adaptation costs to grow exponentially with the number of chip types. Additionally, poor inference stability under high concurrency and persistently high operating costs remain major problems.

3. Solution directions: Cross-chip unified adaptation layers, inference performance optimization, and computing cluster management are all urgently needed solutions. Zhongke Jiahe’s unified virtual instruction set solution has already proven effective, delivering up to a 74x reduction in inference latency and a 1.46x improvement in energy efficiency. This technical direction is a promising area for service providers to explore.

This article identifies the core demands of the LLM industry for AI platforms and outlines development directions for platform operators, with key takeaways as follows:

1. Core merchant demands for platforms: The top demand from LLM developers for platforms is stable, low-cost, scalable computing infrastructure services. Most domestic LLMs need to adapt to multiple types of domestic chips, and the existing fragmented ecosystem cannot meet this need. Platforms need to build a unified adaptation layer to cut developers’ adaptation costs.

2. New platform growth opportunities: As LLMs enter large-scale deployment, demand for inference optimization and computing scheduling services is growing rapidly. Platforms can build new value-added service modules centered on underlying infrastructure capabilities to attract more LLM customers and expand revenue streams.

3. Risk mitigation directions: Platform operators should complete computing capacity planning in advance to avoid insufficient computing power after sudden traffic surges that degrade user experience across the entire platform. They can also follow Zhipu AI’s example to invest across the upstream and downstream value chain, integrate industrial resources, and improve both risk resilience and service capabilities.

This article presents the latest competitive dynamics in the second phase of the LLM industry, raises new questions about industry development, and provides fresh research samples for industrial research. Key findings are summarized as follows:

1. Latest industry dynamics: After entering the second phase, the LLM industry has split into two clear competitive strategies. One is the "brute force" approach focused on model capability, competing on benchmark performance via larger parameter sizes. The other is preemptively building out underlying AI infrastructure to create a full closed-loop engineering chain from chips to applications. The former has already hit clear computing bottlenecks, while the latter is better aligned with the development environment of China’s domestic supply chain.

2. New industry challenges: U.S. export controls prevent Chinese LLM firms from freely purchasing high-end NVIDIA GPUs, while the fragmented domestic computing ecosystem and high cross-chip adaptation costs have become the core bottleneck restricting the large-scale development of Chinese domestic LLMs. This problem is worthy of in-depth research to identify solutions.

3. Business model research: Zhipu AI’s model of laying out the entire industrial chain through a combination of core business and an industry fund to integrate underlying infrastructure capabilities represents a new business model exploration in the LLM industry. It provides a representative case study for research on the independent development path of Chinese domestic LLMs.