China's computing power industry is undergoing a deep-seated redistribution of profits. As the national integrated computing network is elevated to the level of national infrastructure, the billing logic has shifted from "selling cabinets" to "selling tokens". The profit center of the industry value chain is accelerating its move from upstream hardware manufacturers to midstream scheduling operators and downstream application service providers, rendering the traditional competitive logic, which was centered on cabinet scale and GPU inventory, obsolete.

Sinolink Securities points out in its latest industry research report that daily token call volumes in China have surged from 100 billion in early 2024 to 140 trillion by March 2026, a more than 1,000-fold increase in two years. The three major telecom operators have taken the lead in launching token-based tariff packages, and GaaS (GPU-as-a-Service) is becoming the high-end form of computing power operation. At the same time, the "15th Five-Year Plan" has incorporated the national integrated computing network into 109 major projects, elevating computing infrastructure alongside water and electricity to the status of national strategic assets. The industry is transitioning from the "East Data, West Computing" physical layout phase to a new stage of "computing-network integration, nationwide scheduling, and computing-electricity coordination".

This transformation is reshaping the valuation logic of the entire industry. Computing power utilization, token output capacity, scheduling technology level, and compliance qualification barriers have become new valuation anchors, replacing the traditional metrics of cabinet count and power capacity. The Sinolink Securities report believes that entities with cross-regional scheduling and industry operation capabilities will capture a larger share of profits, while small and medium-sized service providers that rely solely on hardware leasing will face sustained operational pressure. The report argues that a paradigm shift is occurring in valuation anchors: investors should shift their focus from cabinet scale and GPU inventory to computing power utilization, token output capacity, scheduling technology, and compliance qualifications. Those operators lacking these capabilities will see intensifying operational pressures.

Pricing Revolution: From Selling Cabinets to Selling Tokens

The Sinolink Securities report characterizes the main theme of this round of change as a "pricing system restructuring". Traditional computing power leasing uses cabinets, individual GPU cards, or server hours as billing units, leading to a highly chaotic market pricing system with wide discrepancies between different service providers, chip architectures, and regional data centers, making low-price vicious competition the norm. Token-based metered billing shifts the pricing target from hardware resources to the actual AI output, directly matching costs with business needs. Shanghai Telecom has launched a token-based computing package where 1 yuan corresponds to 250,000 quota points, while Shanghai Mobile, in collaboration with leading internet platforms, has launched an AI-native workbench where 1 yuan can purchase 400,000 tokens.

The report divides computing power operation models into four tiers: basic computing resource leasing, tiered computing subscriptions, full-chain GaaS solutions, and industry-specific customized computing. Profit margins increase sequentially across these tiers. The core logic of the GaaS model is that operators are no longer constrained by physical cabinet counts and can flexibly schedule domestic NPUs alongside overseas GPUs, leveraging cross-regional heterogeneous scheduling to capture scheduling premiums and token revenue sharing. Revenue structures evolve from hardware rental income to task-based service revenue.

However, the report also highlights practical constraints: the industry has yet to establish a unified national standard for token metering, conversion, and settlement. The statistical calibers of "quota points" and "tokens" vary across vendors, and some government and enterprise clients still have traditional budgeting systems that are aligned with cabinet procurement models. Token commercialization remains in its early exploratory phase.

Where Do the Profits Go: Three Paths of Value Migration

Sinolink Securities breaks down the specific mechanisms of profit migration from three dimensions. First, the shift from "chip price differentials" to "scheduling premiums": operators with unified management and intelligent scheduling capabilities for heterogeneous computing resources can gain service premiums through resource optimization. Second, the shift from "cabinet rental" to "token revenue sharing": under the GaaS model, operators not only collect usage fees but also earn a share from the token consumption of model inference. Third, the shift from "one-time construction" to "ongoing services": long-term contracts in industry-specific customized computing generate stable cash flows, with high customer switching costs and strong collaboration stickiness.

Analysts specifically note that this value migration will be further reinforced under supply constraints. The global shortage of key hardware such as HBM has pushed up hardware procurement and holding costs, compressing the profit elasticity of the asset-heavy model of buying cards and renting them out. Meanwhile, operators with cross-regional scheduling and existing resource revitalization capabilities have seen their relative bargaining power increase. In terms of market size, according to CAICT projections, China's core AI industry scale is expected to exceed 1.2 trillion yuan in 2025. Liu Liehong, Director of the National Data Administration, has publicly projected that by the end of the "15th Five-Year Plan" period, China's AI-related industry scale could reach the 10 trillion yuan range, indicating a significant shift in the value distribution pattern across the industry chain.

Competitive Landscape: Four Camps Emerging, SMEs Under Pressure

Sinolink Securities divides the current Chinese AI computing power operation market into four major camps: public cloud vendors, telecom operators, third-party IDC service providers, and local state-owned computing platforms. Public cloud vendors are at the forefront in cross-regional scheduling, heterogeneous computing integration, and GaaS services, but their "multi-tenant" model faces trust challenges in high-compliance scenarios such as finance and government affairs. Telecom operators possess well-established network infrastructure and government-enterprise customer relationships, having taken the lead in launching token-based tariff packages, but they lag in technological innovation efficiency and ecosystem openness. Third-party IDC providers are caught in a dilemma between high liquid cooling retrofit costs and transformation pressure, with the industry undergoing a deep shuffle. Local state-owned platforms enjoy policy support in land, electricity, and energy consumption indicators, but they have shortcomings in technical capabilities and operational efficiency.

According to CAICT data, the scale of China's intelligent computing services market exceeded 130 billion yuan in 2025, with AI cloud enterprise-level services accounting for approximately 80 billion yuan. In the AI cloud segment, according to Omdia data, China's AI cloud market reached a total scale of 56.7 billion yuan in 2025, with Alibaba Cloud leading at 38.1%, followed by Volcano Engine (20.4%), Baidu Cloud (9.4%), Tencent Cloud (6.3%), and China Telecom Cloud (5.9%). The top five together account for over 80% of the market. The report explicitly states that small and medium-sized computing enterprises face triple pressure: price wars initiated by industry leaders leveraging scale advantages, substantial capital investments required for liquid cooling retrofits and scheduling technology upgrades, and a preference among major clients for full-stack capable leading service providers. Operators lacking hub resources, scheduling capabilities, and industry barriers will face continuously intensifying operational pressure, with some entities pivoting to vertical tracks such as industry-specific and edge computing to survive.

Valuation Restructuring: New Framework Centers on Utilization and Token Output

The report proposes that the valuation framework for computing power operators is undergoing systematic restructuring. The traditional asset-heavy valuation system centered on cabinet scale, server count, and power capacity can no longer accurately reflect the true value of enterprises. The new valuation system needs to incorporate four dimensions: computing power utilization, token output capacity, scheduling technology level, and compliance qualification barriers. Among these, computing power utilization is the core metric for measuring operational efficiency. Cross-regional unified scheduling can alleviate resource idle time, allowing the same physical assets to generate more revenue and profit, serving as the key dividing line between high-quality and inefficient computing enterprises. Operators with superior token output efficiency deserve valuation premiums. Enterprises capable of unified heterogeneous computing management and cross-regional low-latency scheduling can capture technology premiums. Compliance qualifications such as Class 3 Information Security Protection and security-reliability certifications serve as the "entry ticket" to high-premium markets like finance and government affairs.

The report offers clear recommendations for investors: avoid traditional generic IDC projects lacking differentiated competitiveness, actively steer clear of computing projects in non-hub regions, and avoid operators heavily reliant on a single overseas chip supply. In the short term (2026-2027), focus on liquid cooling supporting equipment and 800G/1.6T high-speed optical modules. In the medium term (2028-2029), position in domestic AI chips and computing scheduling software companies. In the long term (2030 and beyond), concentrate on computing-electricity coordination supporting infrastructure and edge computing operators.