China's compute network is entering the infrastructure phase, with western hubs and GW-scale data center clusters becoming the key variables

Overview

This report reviews the key messages from the China AI Compute Industry Ecosystem Development Conference. Participants included data center operators, cloud vendors, token routing platforms, and research institutions such as CAICT and KZ Consulting. Core topics covered the nationwide compute network, GW-scale data center clusters, the ramp-up and bottlenecks of domestic AI chips, token operations, and token exports. The report's central thesis is that China's compute infrastructure is moving from standalone data center construction toward national-scale networking, clustering, and platformization, and industry competition will increasingly depend on power and network resources, cross-domain orchestration capabilities, chip supply structure, and token economics.

Core views

Goldman Sachs believes that over the next five years, China's nationwide compute network could reshape the competitive landscape and geographic distribution of the data center industry. Capital and technology are concentrating in western compute hubs, while data centers in tier-1 cities are more likely to take on roles in low-latency hot computing, edge nodes, and AI inference. The emergence of GW-scale compute clusters implies higher technical, resource, and capital barriers, leading to greater differentiation among data center operators. The ramp-up of domestic chips is a clear direction, but they currently still lag imported chips significantly in performance, cost, ecosystem, and token output efficiency; therefore, domestic substitution does not equate to a simultaneous short-term improvement in profitability.

Analysis framework

The report uses conference minutes and industry comparison methods, combining four analytical threads: policy and infrastructure investment, large-scale data center construction, the economics of domestic versus imported chips, and token business models. The quantitative section mainly draws from CAICT, KZ Consulting, company data, and Goldman Sachs compilations, including investment scale, the number of compute nodes and scheduling platforms, capex for 200MW data centers, domestic chip shipment share, imported compute share among major CSPs, and the gap in token output efficiency.

Methodology notes

• Industrial infrastructure analysisNationwide compute network architecture

  1+M+N node hierarchy and a three-layer architecture

  The report breaks the compute network into the compute-network infrastructure layer, the interconnected resource layer, and the application service layer, and emphasizes that long-distance RDMA, wide-area data transmission, and cross-domain orchestration platforms are critical capabilities.

• Data center economics analysisComparison of 200MW IT power data center capex

  Investment per unit of IT power and per unit of compute under different IT infrastructures

  The report compares Nvidia B300, Huawei CM384, and Panjiu 2.0 solutions, noting that domestic chip solutions have lower capex per unit of IT power, but higher capex per unit of compute and weaker compute output per unit of IT power.

• AI chip supply analysisShare and customer mix of domestic versus imported AI accelerator chips

  Rising domestic chip share alongside persistent performance and ecosystem gaps

  KZ Consulting expects domestic AI accelerator chip shipments in China to exceed 50% share in 2026, with Huawei and T-Head leading, while CAICT also notes that domestic chips still lag overseas peers in performance, cost, and ecosystem.

• Business model analysisToken operations and pricing differentiation

  The impossible triangle of intelligence, cost, and quality

  The report believes token price trends are diverging, linked to trade-offs among low latency, high concurrency, and high availability, as well as competition shifting from single-model breakthroughs to ecosystem competition.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• VNET Group

  A covered China data center company that stands to benefit from changes in compute network buildout and AI inference demand

  Strengths

  Goldman Sachs assigns a Buy rating, with industry infrastructureization and expanding AI demand providing potential support.

  Weaknesses

  The report does not disclose company-specific financial forecasts or a target price.

  Comparison

  Within Goldman Sachs' coverage basket, it shares a Buy rating with GDS and Range Intelligent.

  Risks

  If compute demand, energy access, network capability, or domestic chip efficiency falls short of expectations, operating leverage may be constrained.

• GDS Holdings

  A covered China data center company that may participate in large-scale AI compute infrastructure construction

  Strengths

  Goldman Sachs assigns a Buy rating, and the industry's shift toward scaled and high-barrier clusters favors companies with resources and operating experience.

  Weaknesses

  The report does not provide separate information on GDS's project progress, earnings forecasts, or target price.

  Comparison

  Like VNET and Range Intelligent, GDS is among the data center operators favored by Goldman Sachs.

  Risks

  High capex, financing costs, energy constraints, and the pace of customer demand could affect returns.

• Range Intelligent

  A case study of GW-scale compute clusters and 200MW data center construction

  Strengths

  It has already launched a 200MW, 100,000-chip-scale data center building and believes a 400MW data center is theoretically feasible; Goldman Sachs assigns a Buy rating.

  Weaknesses

  Ultra-large-scale clusters depend on network architecture, power resources, chip supply, and scheduling capability, making execution highly complex.

  Comparison

  The report presents it as a breakthrough case of a 200MW-class data center in China, placing more emphasis on GW-scale cluster capability than traditional data centers.

  Risks

  If domestic or imported chip supply, customer utilization, token economics, or supporting power infrastructure falls short of expectations, project returns could come under pressure.

• Shanghai Athub

  A covered China data center company

  Strengths

  It remains within Goldman Sachs' coverage universe and may benefit from AI inference and low-latency demand.

  Weaknesses

  Goldman Sachs assigns a Neutral rating, indicating less attractive relative return potential than Buy-rated companies.

  Comparison

  Compared with the Buy ratings on VNET, GDS, and Range Intelligent, Goldman Sachs is more cautious on Shanghai Athub.

  Risks

  Changes in the value proposition of tier-1 city data centers, intensifying competition, and uncertain capex returns may limit upside.

• Beijing Sinnet Technology Co Ltd.

  A covered China data center company

  Strengths

  The report discloses it as one of the covered companies, but does not provide positive company-specific arguments.

  Weaknesses

  Goldman Sachs assigns a Sell rating, indicating a weaker relative view within its coverage universe.

  Comparison

  Compared with Buy-rated VNET, GDS, and Range Intelligent, and Neutral-rated Shanghai Athub, Beijing Sinnet sits in the most cautious rating bucket.

  Risks

  A reshaped competitive landscape, changes in the positioning of tier-1 city assets, and rising barriers for AI compute clusters may create relative pressure.

• Domestic AI accelerator chips

  A key supply-side variable for China's AI compute infrastructure

  Strengths

  China shipment share could exceed 50% in 2026, with Huawei and T-Head expected to lead, while internet companies and government entities are important customers.

  Weaknesses

  They still lag overseas chips in performance, cost, and ecosystem, with lower token output efficiency and profitability.

  Comparison

  Compared with imported chips such as the Nvidia H800, Huawei 910B/910C token output is only about 1/6-1/3 as high.

  Risks

  If the efficiency gap proves hard to narrow, domestic substitution may result in higher compute cost per unit and weaker profitability.

Key data

• Investment in the six major networksRmb7tn in 2026Under NDRC definitions, China's 'six major networks' include the water network, new power grid, compute network, next-generation communications network, urban underground pipeline network, and logistics network.
• China data center investmentc.Rmb2tn / US$300bn over next 5 yearsA June 9 Bloomberg report said China's data center investment over the next five years will be about Rmb2tn.
• Penetration of cross-domain orchestration capabilityabout 10%CAICT data show that among 1.1k+ compute nodes and 110+ scheduling platforms, only about 10% have cross-domain orchestration capability.
• Range Intelligent project scale200MW、100k chip-scale；400MW theoretically feasibleRange Intelligent has already launched a 200MW, 100,000-chip-scale data center building this year and believes a 400MW data center is theoretically feasible.
• Share of imported compute among major CSPs70%-90%CAICT says imported chips still dominate the existing compute capacity of China's major CSPs.
• Domestic AI accelerator chip share>50% market share in 2026KZ Consulting expects domestic AI accelerator chips to exceed a 50% shipment share in China in 2026, led by Huawei and T-Head.
• Capex per unit of IT power for domestic chips低40%-50%Compared with imported chips, domestic chips have lower capex per unit of IT power, but lag on efficiency metrics.
• Capex per unit of compute for domestic chips2-4x imported chipsThe report states that capex per unit of compute for domestic chips is 2-4x that of imported chips.
• Compute per unit of IT power for domestic chips10%-30% of imported chipsThe compute generated by domestic chips per unit of IT power is only 10%-30% of that of imported chips.
• Huawei 910B/910C token output1/6-1/3 of Nvidia H800 serversCAICT exhibits show that under typical tasks, daily token output from Huawei 910B/910C servers is lower than that of Nvidia H800.
• Comparison of total capex for 200MW data centersNvidia B300: Rmb61.0bn；Huawei CM384: Rmb41.0bn；Panjiu 2.0: Rmb41.5bnCompany data show significant differences in total capex for 200MW IT power data centers under different IT infrastructures.
• Disclosed prices of covered companiesBeijing Sinnet Rmb13.77；GDS Holdings ADR $31.61；Range Intelligent Rmb90.06；Shanghai Athub Rmb25.80；VNET Group $8.81The prices come from the report's disclosure page, which did not provide a unified target price or expected upside.

Impact & implications

For investors, the key in the data center industry is no longer just rack and land supply, but the integrated competition across compute networks, energy, network transmission, chip mix, and scheduling platforms. Western hubs and GW-scale clusters may receive capital and policy support, and operators with strong resource integration capabilities are more likely to differentiate themselves; meanwhile, assets in tier-1 cities need to reposition through low-latency and inference scenarios. The ramp-up of domestic chips supports supply chain security, but in the near term may depress compute efficiency and token margins. Applications such as token exports, cross-border digital employees, AI toys, and AI short videos may create new demand, but their business models are still at an early validation stage.

Risks

• Domestic chips still lag overseas peers in performance, cost, and ecosystem, which may constrain token output efficiency and margins.
• Insufficient cross-domain orchestration capability—currently only about 10% of compute nodes and scheduling platforms have it—may hinder implementation of the nationwide compute network.
• GW-scale data center clusters require high levels of power, networking, land, cooling, chip supply, and capital expenditure, creating substantial execution risk.
• The value proposition of tier-1 city data centers is shifting toward low latency, edge, and inference, which may put traditional assets under repricing pressure.
• Differentiation in token pricing and trade-offs among low latency, high concurrency, and high availability may lead to unstable business models and margins.
• Token exports and cross-border data processing involve policy, compliance, data security, and international connectivity risks.
• Major CSPs still rely heavily on imported chips for compute capacity, so changes in imported chip supply may affect construction pace.

What to watch

• Progress in nationwide compute network construction from 2026 to 2030, and whether investment in the 'six major networks' translates into actual data center orders.
• The launch pace, utilization, and customer mix of western compute hubs, GW-scale clusters, and data center projects above 200MW.
• Whether domestic AI accelerator chip shipment share exceeds 50% in 2026, and changes in the customer mix of Huawei, T-Head, Cambricon, and Hygon.
• Whether the gap narrows between domestic chips and imported chips such as the Nvidia H800 in token output, capex per unit of compute, and margins.
• The speed of commercial deployment for cross-domain orchestration platforms, long-distance RDMA, and wide-area data transmission capabilities.
• Pilot progress in Shantou City in promoting token exports under the 'Inbound Data Processing' policy.
• Whether applications such as AI toys, AI short videos, and cross-border digital employees can generate sustained token demand.