China Data Centers Move Toward Scaled AI Infrastructure

Overview

This report focuses on the expansion of China's data centers and AI infrastructure. The core view is that AI training and inference demand is driving Chinese data centers toward higher-density, wholesale, and hub-based development, and operators need to capture growth through larger-scale capacity, capital recovery, financing capability, and access to high-end chips and power resources. The report also discusses GPUaaS as an emerging business model, including price changes, economics, and risks, and evaluates expansion opportunities in western and northern compute hubs.

Core views

Goldman Sachs believes Chinese hyperscale internet and cloud companies still spend less on AI infrastructure than U.S. peers, but the growth trend is clear; data center capacity and demand are both expanding, with demand growing faster, which should help improve utilization. At the company level, GDS is viewed as a scale and customer-advantaged wholesale platform among Chinese operators; VNET is in an acceleration phase as it transitions from traditional retail IDC to wholesale IDC; Kingsoft Cloud benefits from demand from Xiaomi and other AI customers; and SUNeVision benefits from growth in data and AI demand in Hong Kong.

Analysis framework

The report analyzes industry supply and demand, hyperscale capex, GPUaaS pricing, data center capacity distribution, major compute clusters, and company valuation methods. Company valuations use SOTP, EV/EBITDA multiples, and DCF, and separately list target price risks.

Methodology notes

• Valuation methodsSOTP

  Sum-of-the-parts valuation

  GDS's 12-month target price is based on a SOTP valuation of GDS China and DayOne, and applies a 10% holding company discount.

• Valuation methodsEV/EBITDA

  Enterprise value/EBITDA multiple

  The target prices for VNET, SUNeVision, and Shanghai Athub use target EV/EBITDA multiples based on 12-month forward or 2027E EBITDA.

• Valuation methodsDCF

  Discounted cash flow

  Kingsoft Cloud's target price is based on a DCF model, with assumptions disclosed in the report including a 10.3% WACC and a 3% TGR.

• factor_frameworkGS Factor Profile

  Growth, financial returns, valuation multiples, and a composite factor profile

  Goldman Sachs' factor profile compares individual stocks with the market and peers on growth, financial returns, valuation multiples, and composite indicators.

• corporate_actionM&A Rank

  M&A probability ranking

  Goldman Sachs uses a 1-to-3 M&A Rank to assess the probability of a company becoming an acquisition target, with 1 representing high probability, 2 medium probability, and 3 low probability.

Asset mapping & comparison

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

• GDS Holdings / 9698.HK

  China operator and wholesale data center platform

  Strengths

  One of the leaders in China's carrier-neutral data center market, with expandable capacity in key compute clusters and a stronger balance sheet through financial discipline and capital recycling.

  Weaknesses

  Projected 2025-2028E monthly service revenue (MSR) in RMB/kW at -8% CAGR, affected by renewals of existing contracts and new orders delivered at lower historical prices.

  Comparison

  Among peers, it has the largest operating and pipeline capacity; the chart shows total capacity of about 5.4 GW.

  Risks

  Slower-than-expected migration demand and utilization improvement, slower ramp in overseas revenue/profitability, China and overseas pricing trends below expectations, customer churn, and a slower de-leveraging process.

• VNET Group

  Chinese data center operator transitioning from traditional retail IDC to wholesale IDC

  Strengths

  Wholesale IDC is expected to account for 35% of revenue in 2025, and 2025-2028E revenue/EBITDA CAGR of 37%-38% could drive re-rating.

  Weaknesses

  Faces share selling by the largest shareholder and potential financing pressure.

  Comparison

  Smaller capacity scale than GDS and Range, but wholesale IDC contribution is expected to rise.

  Risks

  Unable to finance growth targets, weaker-than-expected order execution, AI-related geopolitical risks, further decline in legacy business, and AI model training demand changing faster than expected.

• Kingsoft Cloud

  Chinese cloud services and AI infrastructure beneficiary

  Strengths

  AI is expected to contribute 31% of revenue in 2025, Xiaomi-related business drives a 46% revenue CAGR in 2025-2028E, and the company benefits from Xiaomi's Rmb60bn AI investment over the next three years.

  Weaknesses

  Business growth depends on AI spending and capex capacity from key customers.

  Comparison

  AI revenue contribution is high among Chinese cloud providers.

  Risks

  Supply chain disruptions or inability to secure high-end chips, intensifying peer competition, AI spending by key customers including Xiaomi falling short of expectations, inability to secure capex funding, and financing dilution.

• SUNeVision Holdings

  Hong Kong data center provider

  Strengths

  The largest data center provider in Hong Kong by operating capacity, benefiting from Hong Kong data and AI demand and the expansion of later phases of MEGA IDC.

  Weaknesses

  Growth is affected by the local Hong Kong market, policy, and land/power supply conditions.

  Comparison

  Compared with mainland China operators, it is positioned more toward Hong Kong local capacity and demand.

  Risks

  Rising industry competition, especially from Chinese peers; Hong Kong-specific geopolitical and economic risks; data privacy, regulatory changes, or increased land supply; dilutive M&A, cost overruns, rising financing costs, or convertible bond conversion.

• Shanghai Athub

  Data center operator tied to AIDC expansion

  Strengths

  The target price is based on 20x EV/EBITDA on 2027E EBITDA, showing valuation interest in its AIDC projects.

  Weaknesses

  Highly sensitive to the delivery and migration pace of the Langfang AIDC project.

  Comparison

  The chart shows capacity smaller than GDS, VNET, and Range.

  Risks

  Delivery/migration for the Langfang AIDC project higher or lower than expected, high customer concentration, slower-than-expected expansion execution, AIDC expansion underperformance, and contract renewal uncertainty.

Key data

• China data center capacity4 GW in 2017; 43 GW in 2028EThe chart shows capacity growing by roughly 10x.
• China data center demand2 GW in 2017; 27 GW in 2028EDemand grows faster than capacity, implying utilization improvement.
• China hyperscaler capexUS$9bn in 2022; US$94bn in 2028ESample includes ByteDance, Alibaba, Tencent, Baidu, and Kingsoft.
• U.S. hyperscaler capexUS$156bn in 2022; approximately US$909bn in 2028EThe scale is significantly higher than China, with the 2026E capex/cash flow peak at about 89%.
• GPUaaS U.S. priceUS$1.73/hour in Dec. 2025; US$2.35/hour in Mar. 2026Up about 36% versus Dec. 2025.
• China H100/H200 monthly priceH100: Rmb57 in Feb. 2026, Rmb70 in Apr. 2026; H200: Rmb63 in Feb. 2026, Rmb83 in Apr. 2026The chart labels this as up about 20.30% versus Feb. 2026.
• Share of the eight major compute clusters79% of intelligent compute; 70% of incremental compute supply from 2021-2025; 47% of data center racks in 1Q26Indicates intelligent compute and new supply are highly concentrated in key clusters.
• Range Intelligent capacity6 GW total in China, including 750 MW in operation and 5.25 GW under construction and in reserveFrom the table excerpt.

Impact & implications

If China's AI capex continues to expand, the growth focus of the data center industry may shift from traditional retail IDC to wholesale IDC, high-density racks, GPUaaS, and major compute clusters. Operators with customer relationships, financing capability, delivery execution, and power/land resources are more likely to benefit; but if pricing weakens, customer migration accelerates, financing is constrained, or AI training demand changes, earnings and valuation upside may be capped.

Risks

• China and overseas data center pricing trends below expectations.
• Migration demand, utilization improvement, or customer order execution weaker than expected.
• Disruptions to high-end chip supply or inability to obtain key hardware.
• Insufficient financing capacity, inadequate capex funding, or dilutive financing.
• AI training demand changing faster than expected or in unexpected ways due to technological developments.
• Geopolitical tensions, Hong Kong local policy, data privacy regulation, and land supply changes.
• Customer churn, high customer concentration, and contract renewal uncertainty.

What to watch

• Whether China's hyperscale players can sustain growth in AI infrastructure capex.
• Whether the rebound in GPUaaS pricing continues and how supply-demand dynamics for high-end GPUs such as H100/H200 evolve.
• New supply, utilization, and power constraints in major compute clusters.
• Order intake, delivery, migration, and financing progress at GDS, VNET, Kingsoft Cloud, SUNeVision, and Shanghai Athub.
• Whether the gap between AI infrastructure investment in China and the U.S. narrows.