In the strategic window period when large model technology continues to leap forward and artificial intelligence moves towards generalization, large model cloud, as an important form of the integrated evolution of computing power infrastructure and intelligent algorithms, is accelerating the reconstruction of the artificial intelligence computing power system and the transformation of service paradigms. Relying on ultra-large-scale heterogeneous computing clusters, advanced distributed training architectures, and intelligent resource scheduling systems, large model cloud has built a technical ecosystem covering the entire lifecycle of training, inference, deployment, etc., providing key support for high-dimensional data processing, complex model iteration, and multi-scenario intelligent inference. To adapt to ChinaThe strategic layout of 'Artificial Intelligence+', the direction of digital China construction, and the research work of LeadLeo in collaboration with Frost & Sullivan (Frost &Frost & Sullivan, abbreviated as 'Frost & Sullivan', and the China Academy of Information and Communications Technology have carried out research on the '2025 Large Model Cloud Value Impact Matrix'. The matrix systematically analyzes the value position of large model clouds in promoting the deep integration of artificial intelligence and cloud computing, focusing on their technical characteristics, service capabilities, application scenarios, and industrial structure. The aim is to establish a scientific and authoritative evaluation system, fill gaps in industry research, and help enhance the influence of service providers and industrial collaboration efficiency.
The 'Large Model Cloud Value Impact Matrix' conducts research and evaluation of vendors through three-dimensional diagrams, based on unified assessment criteria. According to the vendors' market, technology, and strategic performance, it vertically divides them into leader, challenger, and follower tiers, and horizontally into roles such as full-stack leader, market winner, strategic visionary, and technology pioneer.An analysis of N sub-segment models clarifies the leading manufacturers in both comprehensive and sub-segment tracks. It helps enterprises make product selection references based on matrix diagrams, assists large model cloud providers in identifying their strengths and weaknesses and capability distribution, empowers enterprises with more comprehensive and balanced product capabilities and solutions for intelligent transformation, and releases the application value of large models.
The research methodology is based on three research dimensions and four levels of progressive evaluation indicators. The market performance dimension assesses the market influence of enterprise large model cloud products from dimensions such as product matrix, market performance, customer coverage scale, revenue growth rate, benchmark customer cases, pricing strategy, etc. The strategic ecosystem dimension considers comprehensive evaluation of the vendor's strategic ecosystem capabilities in terms of open-source project influence, training and certification system, partner network, industry alliances and participation in standard setting, sustainable development, etc. The technical dimension is based on the long-term data accumulation of LeadLeo, Frost & Sullivan, and the China Academy of Information and Communications Technology in standards and evaluation, combined with public information sorting and various research forms such as enterprise interviews. It comprehensively evaluates the technical capabilities of large model cloud products from aspects such as cloud infrastructure, large model platform capabilities, security capabilities, operation and maintenance capabilities, and service stability.
3D Visualization of the Large Model Cloud Value-Driven Influence Matrix
Source: Analysis by Frost & Sullivan
Among them, all-quantum leaders include Alibaba Cloud, Baidu Smart Cloud, and Volcano Engine (NoteThe above sorting is based on the initial letters of the pinyin names (without regard to order). At the same time, the matrix reveals some representative manufacturers from various sub-sectors: market leaders - Mobile Cloud, strategic pioneers - Unicom Cloud, technology leaders - Kingsoft Cloud.
According to LeadLeo, Frost & Sullivan and the China Academy of Information and Communications Technology, current development of large model cloud services exhibits the following characteristics:
1. Market aspect : In the short term(2025-2026): Hybrid cloud architecture and industry model factories will become major trends. This means that enterprises will combine the advantages of public and private clouds while customizing development models according to the characteristics of specific industries. The financial, manufacturing, and government sectors, due to their high level of digitization and clear application scenarios, will be the first to achieve large-scale applications. In the long term (after 2027): The market will gradually form ecological closed loops such as model trading markets and global computing power networks, which will promote efficient resource allocation and widespread technology dissemination. In addition, the emergence of self-evolving systems (such as AutoGPT) will further enhance AI's autonomous learning capabilities, reduce operating costs, and push AI into the 'economies of scale' stage, potentially forming a trillion-level market space.
2. Technology The core capabilities of the Large Model Cloud cover a high-quality model catalog with multimodal and multi-task capabilities, supportingLoRA and other efficient fine-tuning methods provide customized capabilities, elastic reasoning services with low latency and high throughput, as well as a full-process construction system that spans models and applications, comprehensively supporting the rapid implementation and efficient operation of large models.
3. Strategic Aspects In the global wave of open-source large models, China has risen to become a core driver, not only leading the world in terms of the scale of the open-source ecosystem but also achieving breakthroughs in key technologies and cost-effectiveness; with AlibabaDomestic models represented by Qwen have surpassed Meta. Llama, the largest open-source community, has enabled DeepSeek and other models to deliver high performance at costs far lower than mainstream ones, reshaping the cost-performance curve of large models.
Key findings from this research on the large model cloud matrix include the following:
Big Model Cloud Industry Trend 1: CustomizationThe Rise of ASIC and AI Accelerators .The high lifetime costs of GPUs, the risk of single-supply chain lock-in, and low utilization rates during the inference phase are continuously driving up the cost of computing units. This has led to a transformation in the industry towards ASIC customization, which is algorithm-determined, energy-efficient, and highly scalable, and has become an important direction for the upgrade of computing infrastructure.
Source: Analysis by Frost & Sullivan
Industry Trend 2 of Large Model Clouds: Post-training-driven continuous deepening of model cognition During the post-training phase, supervised fine-tuning is performedThe three technical means of SFT), reward model training, and PPO reinforcement learning, combined with human feedback and reward-punishment signals, continuously optimize the model's understanding of prompts and the quality of responses. This enables large-scale pre-trained models to move from 'perception' to 'thinking', achieving a comprehensive improvement in generative quality, reasoning coherence, efficiency stability, and security alignment.
Source: Analysis by Frost & Sullivan
Big Model Cloud Industry Trend 3: Deepening Vertical Segments Large model cloud applications have shifted from the general capability output stage to the industry deepening stage, presentingThe two approaches of 'industry data fine-tuning + core business coupling' are proceeding in parallel: Firstly, based on exclusive high-quality data from fields such as finance, healthcare, and manufacturing, fine-tuning training is carried out to quantitatively improve the model's professional understanding and decision-making accuracy; secondly, the model capabilities are embedded into key areas such as risk control, quality inspection, and operations and maintenance, initially verifying the effectiveness in reducing risks, improving efficiency, and controlling costs.
Source: Analysis by Frost & Sullivan
