Nationwide computing power networks and GW-scale clusters are reshaping China's data center landscape

Overview

This report summarizes the main views from the China AI Computing Industry Ecosystem Development Conference, covering topics such as the nationwide computing power network, GW-scale data center clusters, the ramp-up of domestic AI accelerator chips, token operations, and cross-border token exports. The report's core judgment is that competitive factors in China's data center industry will shift from traditional rack capacity and location toward computing power network scheduling, long-distance RDMA, cross-domain orchestration, power resources, and chip ecosystems.

Core views

Goldman Sachs believes that over the next five years, the nationwide computing power network could reshape the competitive landscape and geographic distribution of China's data center industry: capital and technological resources are concentrating in western China hubs, while the value proposition of tier-1 city data centers is shifting toward ultra-low-latency hot computing, edge nodes, and AI inference. GW-scale clusters imply higher technology and resource barriers and will widen differentiation among data center operators. Domestic chip shipments are rising rapidly, but efficiency and profitability still significantly lag imported chips.

Analysis framework

The report adopts a conference-minutes-style industry research approach, synthesizing views from industry participants and research institutions including data center operators, cloud vendors, token routing platforms, CAICT, and KZ Consulting, and comparing computing power network architecture, GW-scale campus loads, chip capital expenditures, computing efficiency, token output, and API profit margins.

Methodology notes

• Industrial infrastructure frameworkThree-layer architecture of the computing power network

  Compute-network infrastructure layer, interconnected resource layer, application service layer

  The report breaks the computing power network into three layers: heterogeneous computing and high-speed transmission infrastructure, platformized interconnected resources, and application services, emphasizing wide-area transmission, long-distance RDMA, and cross-domain orchestration capabilities.

• Node organization framework1+M+N node hierarchy

  1 national public computing network service platform, M regional platforms, N industry platforms

  This framework is used to describe how resources are organized within China's computing power network, with a focus on cross-regional and cross-industry computing scheduling and resource coordination.

• Chip economics frameworkComparison of computing capex and token efficiency

  Measure chip economics by IT power capex, capex per unit of computing power, computing power per unit of IT power, token output, and profit margin

  The report points out that domestic chips have lower capex per GW of IT power, but higher capex per unit of computing power and lower computing power per unit of power, resulting in weaker token output and API profit margins than imported chips.

• Token operations frameworkThe impossible triangle of intelligence, cost, and quality

  There are trade-offs among low latency, high concurrency, and high availability

  The report uses this framework to explain token pricing differentiation, arguing that competition is shifting from single-model breakthroughs toward ecosystem competition and business model differentiation.

Asset mapping & comparison

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

• VNET Group

  Buy-rated name in Goldman Sachs' China data center coverage

  Strengths

  May benefit from expanding demand for computing power networks, AI inference, and edge nodes.

  Weaknesses

  The report does not provide company-level project, financial forecast, or target price details.

  Comparison

  Rating is more positive relative to Shanghai Athub and Beijing Sinnet.

  Risks

  Industry investment pace, intensifying competition, and constraints on power and network resources may affect delivery.

• GDS Holdings

  Buy-rated name in Goldman Sachs' China data center coverage

  Strengths

  As a data center operator, it may benefit from AI computing demand and expansion by large customers.

  Weaknesses

  The report does not elaborate on company-level operating metrics or project progress.

  Comparison

  Along with VNET and Range Intelligent, it belongs to the Buy-rated group.

  Risks

  Capex pressure, customer concentration, financing costs, and supply-demand mismatch risks.

• Range Intelligent

  Buy-rated name in Goldman Sachs' China data center coverage and cited in the report as a GW-scale cluster case

  Strengths

  A 200MW, 100k-chip-class data center has already begun operation this year, providing a foundation for GW-scale cluster expansion.

  Weaknesses

  Large-scale clusters require high standards for network architecture, power resources, and capital investment.

  Comparison

  The report discloses more specific project progress for it than for other covered names.

  Risks

  The 400MW-scale expansion remains only theoretically feasible, with uncertainties around execution timing, customer demand, and resource approvals.

• Shanghai Athub

  Neutral-rated name in Goldman Sachs' China data center coverage

  Strengths

  Tier-1 city data centers may shift toward ultra-low-latency hot computing, edge nodes, and AI inference scenarios.

  Weaknesses

  The report does not disclose its specific advantages relative to GW-scale western hubs.

  Comparison

  Its rating is below VNET, GDS, and Range Intelligent, but above Beijing Sinnet.

  Risks

  If tier-1 city data centers fail to differentiate through low latency and inference, they may face growth pressure.

• Beijing Sinnet Technology Co Ltd.

  Sell-rated name in Goldman Sachs' China data center coverage

  Strengths

  The report does not provide clear company-level strengths.

  Weaknesses

  It has the lowest rating in the coverage universe, indicating Goldman Sachs is more cautious on its relative return potential.

  Comparison

  Its investment view is the weakest relative to Buy and Neutral names.

  Risks

  Industry resources concentrating in western hubs and GW-scale clusters may weaken the appeal of traditional data center assets.

• Domestic AI accelerator chips (Huawei, T-Head, Cambricon, Hygon)

  Core supply-side variable in China's AI computing infrastructure

  Strengths

  China shipment share may exceed 50% in 2026, with strong demand from internet companies and government entities.

  Weaknesses

  Performance, cost, and ecosystem still lag overseas chips, constraining token output and API profit margins.

  Comparison

  Relative to Nvidia H800/B300, domestic chips have lower capex on an IT-power basis, but weaker computing efficiency.

  Risks

  If ecosystem and efficiency improve more slowly than expected, domestic substitution may be difficult to convert into equivalent profitability.

• Imported chips and Nvidia H800/B300

  Dominant current supply and performance benchmark for major CSP computing power in China

  Strengths

  Better performance in capex per unit of computing power, computing power per unit of IT power, token output, and profit margins.

  Weaknesses

  Higher total capex and possible exposure to supply, policy, and geopolitical factors.

  Comparison

  70%-90% of existing computing power at major CSPs is still based on imported chips.

  Risks

  Import restrictions or supply uncertainty may push customers to accelerate domestic substitution, but with efficiency trade-offs.

Key data

• Core infrastructure investmentRmb7tnUnder NDRC metrics, China's six major networks are expected to attract about Rmb7tn of investment in 2026.
• Data center investment scaleAbout Rmb2tn / US$300bnBloomberg reported on June 9 that China's data center investment over the next five years will be about Rmb2tn.
• Computing nodes and scheduling platforms1.1k+ computing nodes, 110+ scheduling platformsCAICT data shows that only about 10% have cross-domain orchestration capabilities.
• Range Intelligent projectA 200MW, 100k-chip-class data center has already begun operation this yearThe project is part of its GW-scale computing cluster, located on 500 mu of land, and the company believes a 400MW data center is theoretically feasible.
• Domestic chip capex advantage40%-50% lower capex per unit of IT powerRelative to imported chips, domestic chips have lower capex when measured by IT power.
• Domestic chip efficiency disadvantage2-4x capex per unit of computing power; computing power per unit of IT power is 10%-30% of imported chipsThe report believes that gaps in performance, cost, and ecosystem constrain token output efficiency and profitability.
• Imported computing power share among major CSPs70%-90%CAICT says existing computing power at major CSPs in China is still dominated by imported chips, with per-operator computing power of 2.2-4.5 PFLOPS.
• Domestic AI accelerator chip shipment shareMay exceed 50% in 2026KZ Consulting expects China's domestic AI accelerator chip shipment share to exceed half, led by Huawei and T-Head.
• Internet company demand shareHuawei/T-Head/Cambricon/Hygon at 63%/23%/60%/30% respectivelyKZ Consulting says internet companies are among the largest customers for domestic chips.
• Government entity demand shareHuawei/Cambricon/Hygon at 15%/15%/35% respectivelyGovernment entities are also important customers for domestic AI chips.
• Token output gapHuawei 910B/910C are 1/6 to 1/3 of Nvidia H800 serversRelevant CAICT charts show domestic servers materially lag in token output.
• 200MW data center capex comparisonNvidia B300 Rmb61.0bn; Huawei CM384 Rmb41.0bn; Panjiu 2.0 Rmb41.5bnThe table shows that imported solutions have higher total capex, but better capex per unit of computing power and better computing power per unit of power.

Impact & implications

At the industry level, computing power networks and GW-scale clusters will raise barriers to entry and push data centers from simple capacity expansion toward comprehensive competition in computing power, electricity, networking, and scheduling capabilities. At the company level, operators with resource reserves, hyperscale project experience, low-latency nodes, and AI inference scenarios are more likely to benefit; operators reliant on traditional location advantages or lacking resource barriers may face valuation and growth pressure. At the chip level, the trend toward domestic substitution is clear, but in the short term it still faces constraints in efficiency, ecosystem, and profitability.

Risks

• Domestic AI chips continue to lag in performance, cost, and ecosystem, causing token output and profit margin improvement to be slower than expected.
• GW-scale data center clusters require high levels of networking, power, land, and capital, and project implementation may be constrained by resources and approvals.
• Cross-domain orchestration capability in the nationwide computing power network remains insufficient, with only about 10% of current nodes and platforms possessing such capability.
• Token price differentiation and intensifying ecosystem competition may compress API and inference service profit margins.
• Cross-border token exports involve uncertainties in data, compliance, and business models.
• The data center industry has a long capex cycle; if demand or financing conditions change, supply-demand mismatches and lower returns may result.

What to watch

• The pace of implementation for national computing power network investment in 2026-2030 and progress in regional platform construction.
• Changes in the division of roles between western computing hubs and tier-1 city low-latency, edge, and inference nodes.
• The actual number of operational 100k+ chip-scale projects in GW-scale clusters, customer utilization rates, and power supply assurance.
• Domestic AI accelerator chip shipment share in 2026, changes in Huawei and T-Head share, and ecosystem maturity.
• Whether the gap narrows between domestic servers such as Huawei 910B/910C and Nvidia H800/B300 in token output and API profit margins.
• The progress of token export practices in Shantou City, submarine cable connectivity, offshore wind power resources, and cross-border digital employee applications.