Executive Summary and Key Findings
In 2025, a series of numbers began appearing with increasing frequency in China's top-level technology policy documents: eight national computing hubs, 174 EFLOPS, 330 EFLOPS, a half-trillion yuan industry. These figures did not emerge from thin air — behind them lie countless servers running day and night in data center halls carved into Guizhou's mountains, rows of server racks in Inner Mongolia consuming green electricity generated by wind turbines, and a single Huawei Ascend 910C accelerator card in a high-density rack in Shenzhen's Nanshan district, repeatedly processing a financial institution's customer risk assessment requests. Computing power has become one of the most critical infrastructure categories in the country, ranked alongside water, electricity, roads, and gas — and growing far faster than any of them.
This report provides a comprehensive, in-depth study of China's data center and AI computing industry, covering full-year FY2025 data and the latest developments through Q1 2026. It focuses on four core market segments (IDC operations, AI computing services, domestic AI chips, and infrastructure manufacturing) and maps the complete supply chain.
Summary of Key Findings
China's data center and AI computing market reached approximately ¥500 billion (RMB) in 2025, comprising roughly ¥150 billion in IDC operations, ¥250 billion in AI computing services, and ¥100 billion in domestic AI chips. The national stock of in-service standard server racks surpassed 12.5 million, with AI computing capacity reaching approximately 215 EFLOPS — growing nearly 40% year-on-year.
On competitive dynamics, IDC operators have polarized sharply: Runze Technology's net profit surged 182% year-on-year to lead the industry, while GDS and ChinaData maintained steady growth and Guanghuan Xin'an recorded a net loss of over ¥760 million due to goodwill impairment. In AI chips, Huawei Ascend held approximately 49% of shipment share to dominate the domestic market; Hygon Information crossed ¥14 billion in revenue for the first time; Cambricon achieved its first-ever annual profit; and domestic AI accelerator cards collectively reached 41% overall market share.
On the technology front, liquid cooling penetration rates rose rapidly from roughly 5% to approximately 20–25%, with the three major telecom operators reporting over 26% liquid cooling penetration in newly-built data centers. PUE policy pressure (mandatory ≤1.3 for all new large data centers) has pulled liquid cooling commercialization from the future into the present. The eight national computing hubs under the East-Data-West-Computing (EDWC) initiative account for 70% of national computing capacity, with Inner Mongolia's Horinger node ranking first nationwide on the green computing index.
On risks, the U.S. GPU export controls (the H20 ban added in April 2025) represent the most important single external variable, forcibly accelerating domestic substitution while also creating short-term computing shortages. The domestic GPU process-node ceiling (best mass-produced domestic chips approximately equivalent to 7nm, roughly two generations behind TSMC) is the most critical mid-term technical constraint. Uncertainty in the pace of large language model commercialization remains the most significant demand-side volatility factor for AIDC.
Looking ahead to 2030, China's AI computing market could exceed ¥1.2 trillion, with liquid cooling penetration surpassing 50%, domestic GPU market share rising to over 60%, PUE 1.2 becoming the universal industry standard, and Chinese IDC operators establishing a meaningful presence in Southeast Asia and the Middle East.
This report unfolds across twelve chapters covering: definitions and supply chain overview, global competitive landscape, PEST analysis, Chinese market sizing, supply chain decomposition, key company deep-dives, geographic industrial clusters, eight specialized sub-sector analyses, technology evolution, risk landscape, forward projections, and research conclusions — providing a systematic reference framework for policymakers, industry investors, and practitioners.
Chapter 1 Definitions and Supply Chain: The Complete Map from Racks to Computing Power
Data centers, a term that in the 2010s referred merely to back-office IT infrastructure, have by 2025 become central arenas of national strategic competition. From AI startups in Beijing's Haidian district to hyperscale halls deep in Guizhou's Wumeng mountains, every large language model inference request, every real-time financial settlement, every frame of cloud-rendered video must be computed on a server somewhere in a rack. Computing power has become one of the most critical means of production of the twenty-first century.
Scale Classification: Large, Hyperscale, and Edge Nodes
Data centers are typically stratified into four tiers by installed capacity. Small/micro data centers (under 500 square meters, 10–100 racks) primarily serve internal enterprise or localized needs. Medium-scale data centers (500–5,000 square meters, hundreds to thousands of racks) represent the traditional telecom-grade co-location mainframe. Large data centers (5,000–20,000 square meters, typically over 5,000 racks) are the standard configuration for internet platforms and cloud providers. Hyperscale data centers (over 20,000 square meters, tens of thousands of racks or more) are the exclusive domain of super-platforms such as Google, Microsoft, Alibaba, Tencent, and ByteDance. Edge data centers are the other extreme — minimal installed capacity, deployed at urban edge nodes close to end users, specifically serving low-latency scenarios.
Use-Case Classification: General IDC, AI Data Centers, and Edge Computing
General Internet Data Centers (IDC) primarily host storage, networking, and general computing rentals. AI Data Centers (AIDC) are specifically equipped with GPU/NPU clusters to support large model training and inference. Edge computing nodes focus on millisecond-latency scenarios such as autonomous driving, industrial real-time control, and video surveillance. These three forms differ significantly in business model: general IDC revenues from rack rentals are stable but unit prices are low; AIDC charges by GPU-hour at unit prices several to ten times higher than general IDC; edge nodes are typically sold as bundled computing service packages.
Power Efficiency Classification: PUE Tiers
PUE (Power Usage Effectiveness) is the core metric for measuring data center green performance. PUE = total data center energy consumption / IT equipment energy consumption, with an ideal value of 1.0. China's National Development and Reform Commission, in its "Special Action Plan for Green Low-Carbon Development of Data Centers" issued in July 2024, explicitly required: by end-2025, national data center overall rack utilization rates to reach at least 60%; newly-built large-scale and above data centers to have PUE no higher than 1.3; new data centers within national computing hub nodes to have PUE no higher than 1.25.
Supply Chain Overview: The Supply Network Behind a Single Data Hall
Building a data center involves seven core layers. The deepest layer is land and energy — land parcels, power capacity approvals, and green electricity supply. This is the fundamental logic behind the EDWC (East-Data-West-Computing) strategic placement: western regions' electricity costs are 40% lower than eastern regions, with abundant wind and solar resources. The second layer is civil construction and facility infrastructure, including lightning protection, seismic-resistant structures, and fire compartmentalization. The third layer is power systems, covering high-voltage distribution equipment, UPS (Uninterruptible Power Supply), precision air conditioning, and diesel generators — typically the single largest category of data center investment, accounting for 30–40% of total capex. The fourth layer is cooling systems, covering traditional computer room air conditioning (CRAC/CRAH), chillers, and liquid cooling equipment (cold-plate and immersion types). As per-rack power density leaps from 10 kW to 30 kW to 100 kW+, liquid cooling is rapidly displacing traditional air cooling. The fifth layer is rack servers and their core compute components: AI accelerator cards (GPU/NPU), CPUs, memory, and high-speed interconnects. The sixth layer is networks and optical fiber, covering internal backbone networks, external internet access circuits, and inter-datacenter dedicated circuits. The seventh layer is operations software and services covering DCIM (Data Center Infrastructure Management), IT operations, security, and energy management systems.
The industrial logic of this seven-layer supply chain determines who captures value in the data center industry: in China's current phase of rapid expansion, civil construction and power systems represent the largest share of investment; but the highest unit margin goes to core computing hardware (GPU/NPU accelerator cards) and specialized operations services. As the industry matures, operating efficiency and per-watt computing power density will become the core differentiators.
Chapter 2 Global Landscape: Hyperscale Platform Dominance, US-Europe Lead, China Catching Up
Equinix: The "Airport" Model of the Internet
Equinix is the world's largest neutral co-location data center operator, with over 260 International Business Exchange (IBX) data centers across 33 countries and revenue approaching $9.2–9.3 billion in FY2025. What makes Equinix irreplaceable is not the quality of its hardware, but the network effects of its ecosystem: within a single Equinix facility, a customer can directly connect to over 10,000 networks, over 3,000 cloud on-ramp services, and thousands of enterprise customers through Cross Connect cables (direct physical fiber connections between cages within the same facility). This network density constitutes Equinix's true moat — switching an Equinix deployment is not simply moving servers, but destroying an entire web of interconnection relationships.
Digital Realty: The "Industrial Park" Batch Model
Digital Realty represents a different path. It focuses on large-block leasing — leasing entire buildings or entire data halls to hyperscale cloud customers (AWS, Google, Meta). Its FY2025 revenue was approximately $5.9 billion, with the majority coming from long-term (10–20 year) wholesale leases to a small number of large customers. This model provides exceptional cash flow stability (near-zero occupancy fluctuation) but creates high customer concentration risk. Digital Realty's core differentiation lies in large-scale engineering delivery capability — from site selection and permitting to civil construction and equipment installation, it can compress the delivery cycle for a 50-megawatt large data center hall from the industry average of 18–24 months to 12–15 months.
AI Era Structural Shifts: The NVIDIA CUDA Moat
NVIDIA's competitive position in the AI accelerator chip market is not primarily about hardware performance — it is about software ecosystem lock-in. The CUDA programming framework accumulated over a decade is the deepest moat: over 4 million registered CUDA developers, over 370,000 CUDA-accelerated applications, and deep coupling between the mainstream deep learning frameworks (PyTorch, TensorFlow, JAX) and CUDA underlying calls. This ecosystem makes "switching from NVIDIA to domestic GPUs" not merely an issue of replacing hardware, but rebuilding all underlying operator adapters and performance optimization code in a non-CUDA world — far higher migration costs than are visible on the surface.
CoreWeave: The Pure-Play AI Cloud Paradigm
CoreWeave has grown from near-zero to a valuation of tens of billions of dollars in just a few years, providing the clearest market validation for the specialized AI cloud model. Its core bet is: general cloud providers (AWS, Azure, GCP) are optimized for broad workloads and cannot fully optimize for GPU-intensive AI workloads; specialized AI clouds focused on GPU clusters can, through customized GPU scheduling, low-latency high-bandwidth interconnect optimization, and AI-specific operational services, achieve 30–50% better GPU utilization and cost efficiency than general clouds. Its customer backlog exceeded $55 billion, with Meta committing $14.2 billion and OpenAI committing $6.5 billion in multi-year computing procurement contracts — providing extraordinary long-term revenue certainty. The CoreWeave model provides an important reference for Chinese specialized AI cloud operators (such as Ziguang Cloud Intelligence and Baishan Cloud).
Chapter 3 PEST Analysis: Policy and Technology as Dual Engines, Risks and Opportunities Coexisting
Policy Dimension: East-Data-West-Computing's Deep Logic
The "East-Data-West-Computing" (EDWC) national strategy, formally announced in February 2022, is not merely a data center zoning policy — it represents a comprehensive strategic design for China's computing power spatial layout, green energy use, and digital economy equalization. The eight national computing hubs each serve distinct strategic functions.
The Beijing-Tianjin-Hebei hub (anchored on Zhangjiakou-Chengde, Langfang-Baoding) is the most special: its geographic proximity to Beijing (Beijing's data centers are strictly restricted) makes it the primary承接 overflow zone for Beijing's real-time business, with high land costs but unparalleled location advantages. The Yangtze River Delta hub (Wuxi-Changzhou in Jiangsu, Jiaxing-Huzhou in Zhejiang, Shanghai's suburbs) is China's IDC most densely invested area in industry, with mature supply chain, abundant talent, but fierce competition. The Guangdong-Hong Kong-Macao Greater Bay Area hub (韶关 as the core) aims to承接 the Guangdong demand overflow from Guangzhou and Shenzhen.
The Western hubs — Guizhou Gui'an, Inner Mongolia Horinger and Wulanchabu, Gansu Qingyang, and Ningxia Zhongwei — are the strategic core of EDWC. These areas feature power costs 40% below the national average, abundant renewable energy (water power in Guizhou, wind power in Inner Mongolia, solar in Gansu and Ningxia), land costs nearly zero compared to eastern tier-1 cities, and suitable climates (Inner Mongolia and northwest regions allow extended hours of natural cooling, significantly reducing PUE). The challenge has been demand: western nodes struggled with low rack utilization rates in their early years, as the latency cost of moving computing 2,000 km to the west was unacceptable for many real-time businesses.
Economic Dimension: The "Scale Effect" Arithmetic of AI Investment
The most profound structural change driving China's data center investment is the large model training "arms race." A single training run of a 100-billion-parameter model requires approximately 1,000–10,000 NVIDIA A100 GPUs running continuously for one to three months, consuming computing resources worth hundreds of millions to billions of yuan. As the frontier of model scale moves from hundreds of billions to trillions of parameters, the computing demand grows geometrically. The training computing requirements for GPT-4 are estimated to be 10,000+ times those of GPT-2 five years earlier — this is the most fundamental driving force behind the global data center investment boom from 2023 to 2025.
U.S. GPU Export Controls: The Forced Accelerator
The timeline of U.S. export controls on advanced chips to China has had profound effects: 2022 restrictions on A100 and H100; the October 2023 expansion restricting H800, A800, and other circumvention products; April 2024 tightening of license requirements even for lower-performance chips; and the most important April 2025 action formally restricting H20 (NVIDIA's product custom-designed for the Chinese market) from export to China. This series of escalating measures has created a Chinese market that is essentially completely cut off from NVIDIA's highest-performance products.
The paradoxical effect of export controls: while creating short-term computing shortages, they have objectively accelerated domestic substitution. Before the H20 ban, many Chinese AI companies still had access to a "just-barely-not-banned" NVIDIA product; after the ban, there was no cushion — companies had no choice but to accelerate adoption of Huawei Ascend, Hygon DCU, and Cambricon products. This forced substitution has rapidly pushed up domestic GPU shipment volumes and given domestic chip companies an unprecedented window of market opportunity.
Chapter 4 China Market Sizing: The ¥500 Billion Structure and Three Growth Curves
Market Segmentation: Four Layers with Distinct Dynamics
China's data center and AI computing market in 2025, using a comprehensive four-layer calibration, totals approximately ¥500 billion: IDC operations (third-party co-location and hosting services) approximately ¥150 billion, growing at 12–15% annually; AI computing services (GPU cloud rental, dedicated AI computing clusters, intelligent computing center operations) approximately ¥250 billion, growing at 50–70% annually; domestic AI chip market (Huawei Ascend, Hygon DCU, Cambricon, Iluvatar CoreX, Moore Threads) approximately ¥100 billion, growing at over 100% annually; data center infrastructure manufacturing (power equipment, cooling equipment, network equipment, server hardware) totaling approximately ¥500 billion as a supply chain.
Rack Count and AI Computing Capacity
China's total in-service standard server racks reached approximately 12.5 million by end-2025, with an annual net increase of approximately 2 million racks. Of these, AI computing-grade racks (configured with GPU/NPU accelerator cards, supporting single-rack power of 30 kW and above) increased from approximately 800,000 to approximately 1.5 million. National AI computing capacity reached approximately 215 EFLOPS (exaFLOPS) in 2025, with nearly 40% year-on-year growth; projections suggest this will exceed 330 EFLOPS by end-2026.
The Three Growth Curves: General IDC, AIDC, and Edge Computing
General IDC follows a "steady" growth curve of 10–15% annually, driven by traditional internet, fintech, e-commerce, and SaaS applications. AIDC follows a "hockey stick" growth curve, with 50–80% annual growth driven by large model training and inference, policy-driven intelligent computing centers, and the digital transformation of enterprises and government. Edge computing follows a "latent" growth curve — currently fragmented and difficult to commercialize, but expected to accelerate significantly as smart manufacturing and autonomous driving applications mature after 2027.
Chapter 5 Supply Chain Decomposition: Eight Value Layers from Land to Service
Data Center Site Selection: A Six-Variable Decision Framework
Data center site selection is not a simple real estate decision — it is a multi-variable optimization problem that determines a project's long-term economic viability. The six core variables are:
Power availability and cost: This is the single most important site selection variable. Power capacity approvals, grid reliability, and electricity unit pricing directly determine a data center's operating costs (electricity costs account for 50–70% of total operating costs). Western regions' electricity prices of ¥0.25–0.35/kWh versus eastern regions' ¥0.50–0.70/kWh represent a 30–50% difference in long-term operating cost structure.
Network connectivity: Data centers need high-quality, redundant internet access circuits. Core hub locations (Beijing, Shanghai, Guangzhou, Shenzhen) have abundant optical fiber resources and carrier neutral exchange points; western nodes have relatively limited network resources, requiring significant investment in inter-regional dedicated circuits.
Climate and natural cooling: Northern China (Inner Mongolia, Xinjiang) and high-altitude areas (Guizhou, Yunnan) have natural advantages in free-cooling hours — annual hours of air temperatures below 10°C, when data centers can use outdoor air directly for cooling without mechanical refrigeration, significantly reducing PUE. Inner Mongolia's Horinger can achieve over 7,000 hours of natural cooling per year.
Seismic and natural disaster risk: Data center facility investment is enormous, and once built, relocating is nearly impossible. Earthquake zones (southwestern China's Sichuan, Yunnan, etc.), flood plains, and typhoon corridors are all risk factors that must be avoided.
Policy and regulatory environment: Local government support, data center admission standards, land pricing policies, and tax incentives all directly affect project economics. Inner Mongolia, Guizhou, and Ningxia have introduced particularly generous incentive policies to attract data center investment.
Talent supply: Data center operations require specialized maintenance personnel (electrical engineers, mechanical engineers, network engineers, security specialists). Western regions have a structural disadvantage in talent supply, which is a realistic challenge for long-term data center operations.
The Third-Party IDC Technical Service Evolution
Traditional IDC operators primarily provided rack space, power, and cooling — "empty box" services. With the rise of AI computing, the technical requirements of data center customers have expanded dramatically. Modern high-value IDC services have evolved to include: power system design and optimization (high-efficiency transformer selection, distributed UPS design, green energy access solutions); precision cooling system design and commissioning (cold-plate liquid cooling loop design, immersion cooling tank engineering); AI cluster network architecture design (Leaf-Spine topology, RoCE v2/InfiniBand selection, network partitioning strategy); AI server installation, commissioning, and burn-in testing; GPU cluster health monitoring and fault early warning (GPU utilization, memory bandwidth, interconnect link quality real-time monitoring); and AI application performance optimization consultation.
Chapter 6 Key Companies: From Financial Results to Competitive Moats
Runze Technology (300442): The High-Profit IDC Pioneer
Runze Technology reported FY2025 revenue of ¥5.674 billion, net profit of ¥5.050 billion (+182% YoY), net profit margin of approximately 89% — a figure that almost defies belief in the manufacturing and services industry, and that reflects Runze's fundamental structural advantage. The 182% profit growth came primarily from three sources: early layout of large-scale liquid cooling AIDC in Langfang (Hebei), with GPU clusters filling up earlier than peers and achieving high utilization rates; land and power assets acquired at earlier (lower) cost, with book value far below replacement cost; and the surge in GPU rental prices following U.S. export controls making existing NVIDIA H800 GPU assets rapidly appreciate (creating both rental income growth and asset appreciation).
Runze's 89% net profit margin is not replicable for latecomers. Replication requires: completing land and power approvals in core areas adjacent to Beijing before the tightening of EDWC policies; completing the first batch of GPU cluster construction and customer signing before GPU exports were banned; and having sufficient capital to cover the 2–3 year construction period before achieving profitability. This "window advantage" belongs entirely to Runze, and will not recur.
GDS Holdings (万国数据): The International Path
GDS Holdings, as a NASDAQ-listed international IDC operator, has built data center assets in mainland China, Singapore, Malaysia, and Japan. Its strategy is differentiated from purely domestic-market operators: its international operations serve Chinese internet companies (ByteDance, Alibaba) in their international market deployments, as well as local financial and government cloud demands in Southeast Asia. GDS's international expansion logic is clear — as China's internet companies expand globally, they need reliable co-location services in local markets from partners that understand their needs, and GDS's long-term partnerships with Chinese customers give it natural competitive advantages in international markets.
ChinaData (世纪互联): Stable but Pressured
ChinaData is China's largest neutral third-party IDC operator by facility count, with operations in over 20 cities. Its challenge is structural: while having broad geographic coverage, its average data center scale is smaller than Runze and GDS, making it harder to build economies of scale in high-density GPU cluster hosting. Its FY2025 revenue showed moderate growth, but with profitability under continued pressure due to infrastructure renovation costs and competitive price pressures.
Cambricon (688256): The Breakthrough Story
Cambricon's FY2025 performance represented the most dramatic turnaround in China's AI chip sector: revenue of ¥6.497 billion (+453% YoY), net profit of ¥2.059 billion — its first-ever full-year profit. The key driver was cloud inference deployments: Cambricon's MLU series chips achieved significant penetration in Alibaba Cloud, Baidu Cloud, and China Telecom cloud inference instances, generating rapidly growing high-volume revenue. Cambricon's strategic choice — prioritizing inference scenario software adaptation before extending to training — proved to be the right commercial sequencing.
Hygon Information (688041): The Compatibility Play
Hygon Information's FY2025 revenue reached ¥14.376 billion (exceeding ¥10 billion for the first time), with net profit of ¥2.545 billion. The DCU (Deep Computing Unit) product line was the primary growth driver, benefiting from the export control-driven domestic substitution wave. Hygon's x86-architecture compatibility strategy — enabling CUDA ecosystem migration via ROCm-compatible software stacks — gave it lower migration barriers compared to other domestic GPU vendors.
Huawei Ascend: The Ecosystem Builder
Huawei Ascend shipped approximately 812,000 units in 2025, accounting for approximately 49% of domestic market share. Ascend 910B/910C performs at approximately 60% of H100 levels on standard benchmarks, which while significant, is narrowing as Huawei accelerates its 910D and next-generation roadmap. More importantly, Huawei's investment in the Ascend CANN ecosystem — supporting PyTorch/MindSpore adaptation, deploying engineer teams to assist 200+ AI companies in migration — is building the software ecosystem depth that will determine long-term market position.
Sugon Data (曙光数创): The Liquid Cooling Infrastructure Leader
Sugon Data is the domestic liquid cooling infrastructure market leader, benefiting most directly from data center power density increases. Its whole-rack immersion liquid cooling systems have secured supply chain positions with ByteDance, China Mobile, and major intelligent computing center projects. Sugon Data's competitive advantage lies in system integration capability — designing, building, and commissioning complete liquid cooling infrastructure end-to-end, reducing the procurement complexity for data center operators.
Inspur Information (浪潮信息): Scale at the Cost of Margin
Inspur Information reported FY2025 revenue of ¥114.767 billion (+74% YoY), with Q1-Q3 2025 revenue reaching ¥120.669 billion (+44.85% YoY). These growth numbers are spectacular, but the profitability tells a different story: Inspur's business is primarily AI server assembly, with gross margins of 6–8%. The key components (GPUs, memory, high-speed interconnects) are procured at market prices, leaving Inspur capturing primarily systems integration value. Its strategic positioning depends on whether it can move up-stack toward AI solutions and managed services to escape pure hardware assembly commoditization.
Equinix and Digital Realty: International Benchmarks
Equinix FY2025 revenue approximately $9.2–9.3 billion, demonstrating the durability of the interconnection ecosystem model. Digital Realty FY2025 revenue approximately $5.9 billion, maintaining wholesale hyperscale lease dominance. Both serve as benchmarks for Chinese operators on the path to international expansion.
Chapter 7 Industrial Clusters: The Geographic Map from Guizhou Data Valley to Southeast Asia
The Five Western Hubs: Differentiated Positioning
Guizhou Gui'an New Area is China's first national big data comprehensive pilot zone, with policy continuity advantages, excellent climate (average annual temperature 14°C, minimal mechanical cooling needed), and water-powered green electricity guarantees. Its disadvantage is distance from major demand centers, suitable only for batch computing and cold data storage. Guizhou's data center operations industry has matured, with Huawei, Tencent, and Apple iCloud China all having established data centers here.
Inner Mongolia Horinger ranks first nationally on the green computing index, with wind power generation exceeding 80% of electricity supply. The region's extremely low electricity cost (approximately ¥0.25/kWh) and strong cooling climate (below-freezing winters enable extended natural cooling) make it the preferred location for batch training computing tasks. Wulanchabu, another Inner Mongolia hub, leverages similar advantages and has attracted ByteDance and Alibaba large-scale investment.
Gansu Qingyang and Ningxia Zhongwei are relatively newer hubs, primarily utilizing northwest China's abundant solar and wind resources. Policy subsidies are generous, but demand-side challenges remain — the combination of location disadvantages and talent scarcity means these hubs remain more policy-driven than market-driven.
East Coast: The Premium Scarcity Assets
Beijing-Langfang is arguably the most valuable data center corridor in China. Beijing's strict capacity cap on new data centers creates artificial scarcity, while Langfang (just 60 km away) combines acceptable network latency with far lower land and power costs — making Langfang the inevitable承接 zone for Beijing's data center overflow. Runze Technology's Langfang complex is the showcase asset of this corridor.
The Shanghai-Kunshan-Jiaxing corridor serves identical overflow demand from Shanghai. Kunshan (within 30 minutes' drive of downtown Shanghai) has become China's most densely invested data center area per square kilometer outside of Beijing-core, with dozens of data centers from leading operators including GDS, Chindata (ChinaData), and CapitaLand's Ascendas.
Southeast Asia: The Hottest International Market
The investment boom in Southeast Asian data centers from 2024 to 2025 has been historic. Malaysia's Johor state is the most important landing zone for Chinese IDC companies going international. Johor's proximity to Singapore (as close as 1 km across the Johor Strait, with round-trip network latency under 2 ms) means it effectively functions as a technical extension zone for Singapore's data center ecosystem. Meanwhile, Johor's industrial electricity prices are approximately one-fifth of Singapore's, and land costs approximately one-tenth — a rare "adjacent to high-priced market + own low cost" location combination. Aofei Data's flagship Southeast Asia project is in Johor, targeting Chinese internet companies (e.g., ByteDance's Southeast Asia-facing applications) that cannot store data in China but need Asia-Pacific nodes, as well as local financial institutions and government cloud demand.
The Middle East (Saudi Arabia, UAE) has rapidly become one of the world's most active emerging data center investment regions from 2024 to 2025. Saudi Arabia's Vision 2030 digital transformation strategy is investing massively in domestic AI infrastructure; the UAE (especially Dubai and Abu Dhabi) continues attracting large IDC investments with its mature business environment, low tax rates, and global internet exchange position. Chinese data center infrastructure products (intelligent room solutions, liquid cooling equipment, high-performance AI servers) are becoming important procurement sources for Middle East data center projects.
Chapter 8 Sub-Sector Deep Dives: Nine Tracks Examined
Track 1: General IDC → AIDC Transformation Path
The most important structural change in China's IDC industry is the "AI transformation" of traditional data centers. General IDC space cannot directly accommodate AI GPU clusters — the core challenge is power density: traditional IDC designs for 5–10 kW/rack, while AI GPU clusters require 20–80 kW/rack (and the next generation exceeds 100 kW/rack). This means traditional IDC cannot simply add GPU servers; it requires systematic infrastructure upgrades across power distribution, cooling, and structural reinforcement.
The transformation path for traditional IDC operators: first, assess which facilities can be upgraded (based on transformer capacity, floor load capacity, and cooling system scalability); second, determine the upgrading scheme (distributed small liquid cooling modules added inline vs. full liquid cooling renovation); third, establish customer relationships (securing a large model company or cloud provider anchor tenant before renovation begins is often prerequisite for project approval). The estimated total investment for transforming a medium-scale data center (3,000 racks) from general IDC to AIDC capability ranges from ¥200–500 million.
Track 2: Liquid Cooling Supply Chain
The liquid cooling supply chain in 2025 has evolved from "laboratory showcase" to "large-scale commercial deployment." The supply chain includes: upstream fluorinated liquid manufacturers (3M Novec series, domestic substitutes in early stages), aluminum/copper heat sink and cold plate manufacturers, pump and pipe fitting suppliers, secondary distribution unit (CDU) integrators, and final end-to-end system integrators (Sugon Data, Inspur, Lenovo Solutions). The supply chain's bottleneck is currently shifting from "can it be done" to "how to do it fast and reliably at scale."
Track 3: Domestic GPU Software Ecosystem
The true bottleneck for China's domestic GPU market is not hardware — it is software ecosystem maturity. Building CUDA-level software ecosystem depth requires years of developer community cultivation. Huawei's Ascend CANN framework has achieved mainstream framework (PyTorch, MindSpore) adaptation; Hygon's ROCm-compatible approach lowers migration barriers; Cambricon has prioritized inference scenario operator coverage. The question is whether China can build a self-sufficient domestic GPU software ecosystem within the Chinese market — not matching CUDA's global depth, but sufficient for Chinese AI applications to operate efficiently without CUDA.
Track 4: Training vs. Inference Computing
Training and inference have fundamentally different hardware requirements. Training demands high-bandwidth memory (HBM), high-precision floating point (FP16/BF16), and massive inter-chip communication bandwidth. Inference can often use lower precision (INT8/FP8), prioritizes throughput and latency, and can often be served on fewer, less expensive chips. In 2025, inference computing growth rate has exceeded training computing growth rate, as large model commercial applications (AI assistants, code completion, AI search) scale dramatically. This is driving specialized "inference-optimized" chip products from domestic vendors.
Track 5: Green Computing and Carbon Footprint
China's data centers consumed approximately 240 TWh of electricity in 2025 — about 3% of total national electricity consumption — and are projected to exceed 350 TWh by 2030. The carbon footprint of this electricity consumption is enormous: at China's average grid carbon intensity of approximately 500 gCO₂/kWh, national data centers emit approximately 120 million tons of CO₂ annually. Policy pressure to reduce this carbon footprint is intensifying: through green electricity certificates (RECs), direct renewable energy procurement agreements (PPA), and carbon market integration, data centers are being pushed to increase renewable energy usage and reduce carbon intensity per unit of computing.
Chapter 9 Technology Evolution: The Power Density Revolution from 30 kW to 100 kW+
The Three Inflection Points of Liquid Cooling Commercialization
The first inflection point occurred in 2023, when major cloud providers (Alibaba Cloud, Tencent Cloud, Huawei Cloud) began specifying liquid cooling as a standard requirement for new AIDC projects. This converted liquid cooling from an optional premium to a baseline requirement for AI-grade data centers.
The second inflection point came in 2024, when the national standard team released the first batch of liquid cooling data center design specifications, followed by the three major telecom operators formally incorporating liquid cooling standards into their data center procurement requirements. With government operators standardizing the technology, the supply chain scaled rapidly and unit costs fell by 30–40%.
The third inflection point arrived in 2025, as next-generation GPU products (NVIDIA B200/GB200, Huawei Ascend 910D) drove single-rack power to 80–120 kW — beyond the thermal limits of any conventional air cooling system. At this power density, liquid cooling is no longer an option but a physical necessity.
Cold-Plate vs. Immersion: TCO Over Five Years
The two dominant liquid cooling technologies — cold-plate and immersion — have distinct economics. Cold-plate liquid cooling (attaching liquid cooling plates directly to CPU/GPU surfaces while leaving other components air-cooled) can typically handle up to 50–60 kW/rack, has a lower implementation cost (approximately 15–25% premium over conventional air-cooled solutions), and is compatible with standard server form factors. Immersion liquid cooling (submerging entire server boards in dielectric liquid) can handle over 100 kW/rack, has a 30–50% implementation cost premium, but achieves PUE as low as 1.05 (versus ~1.3 for cold-plate), delivering significantly better long-term energy cost economics for the highest-density deployments.
Leaf-Spine Network Architecture Evolution
AI cluster networking is fundamentally different from traditional internet data center networking. For AI training clusters, network latency between GPUs (the "all-reduce" collective communication bottleneck) is often the limiting factor for training throughput, not GPU compute performance itself. This requires ultra-low latency, ultra-high bandwidth interconnects — NVIDIA InfiniBand HDR/NDR (400/800 Gbps) or RoCE v2 (Remote Direct Memory Access over Converged Ethernet) at 400–800 Gbps. The network topology evolves from traditional three-tier (access/aggregation/core) to Leaf-Spine fat-tree architectures that minimize any-to-any latency within the GPU cluster.
AIOps: AI Managing AI Computing
The irony of the AI computing era is that managing AI infrastructure increasingly requires AI itself. AIOps (Artificial Intelligence for IT Operations) in the data center context means: AI-driven cooling control (Google's DeepMind application in its own data centers achieved approximately 40% cooling energy reduction), predictive hardware failure detection (identifying GPU, memory, and storage failure patterns days or weeks before they occur), automated GPU cluster workload scheduling (balancing training and inference workloads dynamically across the cluster to maximize GPU utilization), and intelligent network congestion prediction and rerouting.
Chapter 10 Risk Landscape: Six Systemic Threats in Depth
Risk 1: U.S. GPU Export Controls — The Largest Supply Chain Uncertainty
U.S. semiconductor export controls on China represent the largest single external risk to China's data center industry. With H100/H200/B100/B200 fully banned and H20 restricted from April 2025, China's market is completely cut off from NVIDIA's highest-performance GPUs through normal channels. For domestic IDC operators, the primary impact is: existing NVIDIA GPU inventory has become scarce and valuable; new AIDC buildouts are forced to use domestic GPUs, shifting supply chain reliability and software ecosystem compatibility risks to domestic vendors. For large model companies, NVIDIA supply disruptions mean constrained training capacity scaling, potentially falling behind overseas competitors with access to complete Blackwell clusters.
Risk 2: Domestic GPU Process Node Ceiling — Physical Constraints on Competitiveness
Domestic GPUs face a fundamental constraint in China's semiconductor manufacturing process maturity gap with TSMC. Current best-in-class domestic fab (SMIC) mass-produces chips at approximately 7nm equivalent, while TSMC is shipping 3nm volume with N2 about to launch — approximately two generations behind. This means equivalent-compute domestic chips have larger die area (higher cost), higher power consumption (higher thermal density, greater cooling demand). Breakthrough paths: domestic fabs accelerating catch-up (constrained by EUV lithography equipment restrictions); architectural innovation through Chiplet multi-die interconnect and advanced packaging (COWOS-like technology) to improve system compute at the package level.
Risk 3: Large Model Commercialization Uncertainty
Data center investment logic largely rests on assumptions of accelerating large model commercialization. But LLM commercialization faces significant uncertainty: most enterprise LLM deployment projects are still in "pilot" or "small-scale procurement" stages; the largest consumers of computing power (Tencent, Baidu, ByteDance, Alibaba) are primarily self-building compute and don't rely mainly on third-party IDC; whether LLM POC (proof-of-concept) projects across traditional industries convert to long-term compute contracts depends on whether ROI becomes clear.
Risk 4: PUE and Energy Compliance Pressure
PUE policy constraints are not just a technical challenge but a compliance risk. Existing data centers unable to achieve PUE ≤1.3 within specified timelines face administrative risks including restrictions on new rack approvals and local government remediation mandates. Systematic upgrades of power distribution and cooling systems require large capital expenditures and may involve months of downtime — a major operational challenge for availability-critical operators. Some older data centers (especially those 10+ years old) may face forced retirement if renovation costs are too high.
Risk 5: Geopolitical Risk Abroad
Chinese IDC companies expanding internationally will increasingly face geopolitical risk post-2025. Southeast Asian governments (especially Indonesia, Vietnam) are tightening data security reviews of Chinese technology companies. Some countries (India) have actual policies restricting or prohibiting direct Chinese tech company investment in data centers. Even in relatively friendly Malaysia, 2025 saw data center electricity quota controls prioritizing local demand, affecting some Chinese companies' expansion plans.
Risk 6: Customer Concentration in Hyperscale Platforms
Leading IDC operators (GDS, Runze, Aofei) have customer bases highly concentrated in a small number of hyperscale platform customers (ByteDance, Tencent, Alibaba, Baidu), with the top three customers typically contributing 50–70% of total revenue. The downside: weak bargaining power, high customer-departure or self-build risk. Tencent, Alibaba, and ByteDance have all been continuously increasing their self-built IDC ratios, reducing reliance on third parties. If a major customer doesn't renew at contract expiry, it can be devastating for smaller IDC operators.
Chapter 11 2026–2030 Projections: Five Trends' Evolution Paths
Trend 1: AI Computing Market Scale Projections
Base case scenario: China's AI computing market (AI computing services + AI chips) will reach approximately ¥600 billion by 2027 and approximately ¥1.2 trillion by 2030, representing a 5-year CAGR of approximately 35% from 2025. This projection assumes: large model commercial applications (AI assistants, AI search, AI-native enterprise software) continue to expand as planned; domestic GPU shipments can meet demand growth (no severe supply shortages); and no major policy disruptions (no large-scale government rollbacks of intelligent computing center programs).
Bull case scenario (2030 market ¥1.5 trillion+): assumes "algorithmic efficiency plateau" — as models approach Scaling Law limits and inference efficiency improvements decelerate, per-unit-of-intelligence compute demand grows faster than efficiency gains, maintaining or accelerating demand growth. This scenario is possible but requires breakthrough large model application scenarios (such as autonomous driving full penetration, AGI-level enterprise applications going mainstream).
Bear case scenario (2030 market ¥600 billion): assumes significant commercialization setback — most enterprise AI projects fail to demonstrate clear ROI and are cancelled, training compute demand stagnates after foundation model wars stabilize, and inference demand falls far short of expectations. This scenario has low probability (current evidence suggests strong ongoing demand), but is not impossible.
Trend 2: Liquid Cooling Penetration to Reach 50%
Liquid cooling's penetration rate growth will be driven by physics: as GPU power density continues to increase (NVIDIA's next-generation Rubin architecture is expected to push single-chip TDP to 1,000W+), air cooling physically cannot dissipate the heat. By 2027, liquid cooling will become the default configuration for all new AIDC projects; by 2030, liquid cooling penetration in new data center racks will exceed 70%, and the overall installed base (including existing data centers) will reach approximately 50%.
Trend 3: Domestic GPU Market Share to Reach 60%+
Domestic GPU market share growth faces two prerequisites: hardware performance continuing to close the gap with NVIDIA; software ecosystem maturing to reduce migration costs. Assuming Huawei Ascend 910D achieves approximately 80% of H100 performance in late 2026, and domestic GPU soft ecosystem (CANN + ROCm + domestic frameworks) achieves mainstream framework compatibility, domestic GPU market share in 2027 could reach 50–55%; by 2030, with another full product generation, 60–70% is achievable.
Trend 4: Intelligent Computing Centers Becoming the New Power Grid
In the next 5 years, national-level intelligent computing center construction will become a standard configuration for all provincial capitals and most prefecture-level cities, similar to how power grids and highway networks were constructed in the last century. Government-driven intelligent computing centers will form a national computing power network, with standards for inter-hub resource scheduling, task migration, and unified billing being gradually established.
Trend 5: International Expansion Establishing Initial Footprint
By 2030, Chinese IDC operators will have established meaningful international presence, primarily in Southeast Asia (Singapore-Johor corridor, Jakarta, Bangkok), Middle East (Dubai, Riyadh), and selected markets along the Belt and Road. This international presence will primarily serve Chinese internet companies' international operations and local government/enterprise digital transformation needs.
Chapter 12 Research Conclusions and Strategic Judgments
Core Judgment 1: Computing Power Has Become Strategic Infrastructure, Market Logic and Policy Logic Must Be Read Together
The most important conclusion of this report is that China's data center and AI computing industry can no longer be analyzed with a purely market-logic framework. The government's determination to build computing power as strategic infrastructure, the EDWC policy's decade-long commitment, and the intelligent computing center programs across all provinces and cities have collectively made computing power a "strategic public good" — its supply is driven partly by market demand and partly by national strategic needs. This dual-drive structure means the industry's growth rate will be structurally higher than pure market logic would predict, but the risk of inefficient investment (building computing capacity that market demand cannot absorb) is also structurally higher.
Core Judgment 2: Export Controls Are Accelerating the Bifurcation of Two Parallel Ecosystems
U.S. GPU export controls have not simply "blocked" China's access to advanced chips; they have accelerated the formation of two parallel AI computing ecosystems globally. The "Western computing system" centered on NVIDIA's CUDA ecosystem will dominate markets in Europe, America, Japan/Korea, Southeast Asia, and India; the "China computing system" built around Huawei Ascend and Hygon DCU will dominate the Chinese domestic market and extend into some Belt-and-Road countries. This "ecosystem bifurcation" may be more significant than the Great Firewall internet split — it will reshape global AI technology sovereignty alignment and human-capital ecosystems for the next decade.
Core Judgment 3: The True Battleground Is Software Ecosystem, Not Hardware Performance
Hardware performance gaps are shrinking — Huawei Ascend 910C at ~60% of H100, potentially reaching 80%+ with 910D, is approaching "good enough" for many commercial applications. The real bottleneck is software ecosystem: whether Chinese AI companies can develop, train, and deploy models efficiently on domestic GPU platforms; whether domestic GPU manufacturers can build developer communities and operator libraries with depth comparable to CUDA; whether academic institutions train the next generation of AI engineers on domestic GPU platforms. If China's domestic software ecosystem achieves "closed-loop self-sufficiency" within the Chinese market, the CUDA moat loses practical relevance for Chinese market participants — and that transition, if it occurs, represents the true inflection point for domestic GPU market dominance.
Core Judgment 4: 天下工厂研究院 Sees Manufacturing's Critical Support Role
天下工厂产业研究院, in mapping the global AI computing geography, notes that the hardware enabling the AI era — from precision air conditioning for data centers, to liquid cooling systems, to high-speed optical modules, to server chassis and racks — is manufactured in Chinese factories. China's manufacturing base is a critical enabler of the global AI computing buildout, supplying the physical infrastructure that makes the software and services layer possible. Understanding where these manufacturing capabilities are concentrated, and which factories are building the hardware of the AI era, is part of understanding the full competitive landscape.
Strategic Recommendations
For IDC investors: the risk/return profile of existing NVIDIA GPU assets (scarce, appreciating) versus new AIDC builds (domestic GPU execution risk, software ecosystem risk) requires clear differentiation. Premium assets remain Tier-1 core city and core-satellite corridor (Beijing-Langfang, Shanghai-Kunshan) existing capacity; greenfield AIDC investment in western nodes requires more patient capital and careful demand underwriting.
For AI chip companies: the software ecosystem is the decisive investment priority. Hardware performance closing the gap is necessary but not sufficient — the companies that will win long-term are those that build deep software library ecosystems, cultivate developer communities, and make migration from CUDA economically and practically feasible for Chinese AI practitioners.
For data center operators: the next 3 years are a pivotal window for liquid cooling competence building. Operators that establish liquid cooling design, procurement, and commissioning capabilities before the 100 kW/rack era fully arrives will command premium rates and preferred customer relationships; those that wait will face both higher supply chain costs and a narrowing window to build the expertise.
For policymakers: the policy priority should shift from "building more computing capacity" to "improving computing utilization and efficiency." Western node rack utilization rates below 30% represent massive sunk investment being underutilized. Policies that improve demand migration to western nodes (latency-tolerant workload classification, pricing incentives for east-to-west compute migration) will unlock more economic value than simply approving more data center capacity.
Data Sources and Key References
Data and analysis in this report draw from the following sources. Financial data for publicly listed companies comes from official corporate filings (A-share annual reports, HKEx/NASDAQ filings) with fiscal year 2025 coverage. Market sizing estimates integrate analysis from Synergy Research Group, IDC Corporation, Gartner, China's Saidi Consulting, and the China IDC Circle; where estimates diverge, the report uses median values and notes the range. Policy and industry statistics reference the National Development and Reform Commission, Ministry of Industry and Information Technology, China Academy of Information and Communications Technology (CAICT), and the National Energy Administration official publications and white papers. International market data references Equinix, Digital Realty, CoreWeave investor relations disclosures and analyst coverage.
天下工厂产业平台工厂数据库 — covering 4.8 million in-production Chinese factories — provided cross-validation for supply chain density analysis in this report, particularly for manufacturing sub-sectors including data center precision cooling, server power supplies, intelligent computing center construction, and optical module manufacturing. This real-world factory distribution data supplements documentary research with ground-truth production capacity mapping.
This report is intended as a research reference for industry participants, investors, and policymakers. All data has been sourced from public information; no non-public information was used. Readers applying this report's findings to investment decisions should conduct independent verification and evaluation based on their own risk tolerance and judgment.