China AI compute and infrastructure buildout Report Interpretation

Overview

Morgan Stanley presents a China AI infrastructure outlook centered on the scale, financing and returns of compute investment through 2030. Its base case calls for RMB8.5 trillion of capex, with AI Cloud as the principal growth driver and positive implications for hyperscalers, cloud providers, data-center operators and AI labs.

Core views

Morgan Stanley forecasts RMB8.5 trillion of total AI-related capital expenditure during 2026-30E, supporting IT capacity of 81GW by 2030E and approximately 55GW of compute. The projected five-year outlay includes RMB6.3 trillion of hyperscaler capex, RMB1.5 trillion for neoclouds and RMB0.654 trillion of telecom compute capex. Within hyperscaler spending, RMB4.2 trillion, or about 67%, is domestic, while RMB2.1 trillion, or about 33%, is international infrastructure and investment. Server spending represents RMB5.3 trillion, or roughly 88%, of compute capex; AI servers account for RMB4.9 trillion, or 92% of server capex. AI Cloud is the report's central demand driver. Morgan Stanley forecasts RMB1.2 trillion of hyperscaler/internet capex in 2027E across Alibaba, Tencent, Baidu, ByteDance, Kuaishou, Meituan and Xiaomi. It projects China power demand of 5.6-13GW during 2026-30E and third-party IDC orders of 6.1-11GW. Hyperscalers are forecast to add 34GW of domestic capacity over the period, or 47GW globally. The report also argues that domestic compute deployment is cheaper than in the US, where it cites a cost range of US$15-23 million per MW. Financing is a central constraint, particularly for the RMB2.0 trillion of offshore capex identified as a key consideration. The report notes that hyperscalers occupy multiple positions in the supply chain and that neocloud financing carries heavier debt and lease obligations. It also highlights more limited circularity in China, making the durability of compute demand and the funding structure important for whether the planned buildout can earn acceptable returns. Morgan Stanley tests ROIC and payback across infrastructure business models. For self-owned GPU IaaS, its illustrative outcome is a 44% operating margin, 13% ROIC and a three-year cash payback; rented infrastructure avoids capex but produces roughly a 20% operating margin. For self-built MaaS, the 1P model produces a 53% operating margin, 19% ROIC and 2.5-year payback in the cited case, while a third-party API model produces a 57% operating margin, 29% ROIC and two-year payback. Domestic inference may have lower ROIC, though Morgan Stanley sees room for improvement as performance and utilization develop. The institution states an Overweight view on AI infrastructure—Alibaba, Kingsoft Cloud and VNET—alongside AI labs including MiniMax and Z.AI, and GPU localization. Its company valuation discussion uses differentiated frameworks: DCF assumptions of 15% WACC and 3% terminal growth for the AI labs; a 10% WACC and 3% terminal growth base case for Alibaba; 5.5x 2027E EV/EBITDA for Kingsoft Cloud, below US neocloud peers' 13x average and 10x median because of lower asset yields and supply-chain uncertainty; and a 10-year DCF for VNET using 8.7% WACC, 4.0% cost of debt, 14.0% cost of equity, 50% debt weighting and 3% terminal growth.

Analysis framework

Morgan Stanley starts with a top-down estimate of China AI capex, decomposes the spending by investor type, geography and equipment, then translates it into capacity, power demand and data-center orders. It assesses financing needs and models operating margins, ROIC and payback across self-owned, rented, 1P and third-party API compute models before applying company-specific valuation frameworks.

Methodology notes

• Industry AnalysisSupply-demand framework

  Capex-to-compute, power-demand and data-center-order forecasting

  The report converts forecast AI investment into server spending, capacity additions, power demand and third-party IDC orders to assess the buildout and its beneficiaries.

• Corporate Fundamentals and FinanceROIC–WACC spread

  ROIC, operating-margin and cash-payback analysis for compute business models

  The report compares the returns and payback periods of owned versus rented GPU infrastructure and 1P versus third-party API MaaS models.

• Valuation methodsDCF (Discounted Cash Flow)

  Discounted cash flow valuation

  The report cites DCF assumptions including WACC and terminal-growth rates for several covered companies.

• Valuation methodsEV/EBITDA valuation

  2027E EV/EBITDA multiple valuation

  For Kingsoft Cloud, Morgan Stanley uses a 5.5x 2027E EV/EBITDA multiple and compares it with US neocloud peer multiples.

Asset mapping & comparison

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

• Alibaba Group Holding (BABA.N)

  Overweight AI-infrastructure exposure; AI Cloud demand is a key growth driver.

  Strengths

  Potential upside from stronger core e-commerce monetization, faster enterprise digitalization and AI-driven cloud revenue.

  Weaknesses

  Higher reinvestment costs may weigh on returns.

  Comparison

  Valued using a DCF base case with 10% WACC and 3% terminal growth.

  Risks

  Competition, weaker consumption, slower enterprise digitalization and additional platform scrutiny.

• Kingsoft Cloud Holdings (KC.O)

  Overweight AI-infrastructure exposure and neocloud beneficiary.

  Strengths

  Potential benefit from stronger AI GPU demand, lower rates and improved Xiaomi AI-model development.

  Weaknesses

  Lower asset yields and supply-chain uncertainty support a valuation discount to US neocloud peers.

  Comparison

  5.5x 2027E EV/EBITDA versus US neocloud peers' 13x average and 10x median.

  Risks

  Supply-side procurement restrictions, higher funding costs and slower AI-model development in China.

• VNET Group Inc (VNET.O)

  Overweight AI-infrastructure and third-party IDC exposure.

  Strengths

  Potential upside from wholesale contracts, faster customer move-ins, rate cuts and REIT asset monetization.

  Weaknesses

  Capacity delivery and sales execution are important operational dependencies.

  Comparison

  Valued with a 10-year DCF using 8.7% WACC and 3% terminal growth.

  Risks

  Hyperscaler AI-capex reductions, delivery delays and weak sales execution.

• MiniMax (0100.HK)

  Overweight AI-lab exposure.

  Strengths

  Potential upside from launching a global state-of-the-art model and easing market competition.

  Weaknesses

  Model performance must remain competitive.

  Comparison

  DCF uses 15% WACC and 3% terminal growth; the price target implies 25x 2027 P/S.

  Risks

  Geopolitical risk, intensified competition and price wars, and model performance lagging peers.

• Z.AI CO., LTD. (2513.HK)

  Overweight AI-lab exposure.

  Strengths

  Potential upside from global expansion with overseas cloud-service providers and easing competition.

  Weaknesses

  Computing constraints may limit execution.

  Comparison

  DCF uses 15% WACC and 3% terminal growth; the price target implies 35x 2027E P/S.

  Risks

  Computing constraints and geopolitical risk.

Key data

• Total AI capex forecastRMB8.5 trillion2026-30E total capex forecast
• IT capacity81GWProjected by 2030E
• Compute capacity~55GWSupported by the five-year capex forecast
• Hyperscaler capexRMB6.3 trillionFive-year aggregate capex
• Neocloud capexRMB1.5 trillion2026-30E forecast
• 2027E hyperscaler/internet capexRMB1.2 trillionForecast for Alibaba, Tencent, Baidu, ByteDance, Kuaishou, Meituan and Xiaomi
• Domestic hyperscaler net additions34GW2026-30E; 47GW on a global scale
• Third-party IDC orders6.1-11GWForecast for 2026-30E

Impact & implications

The report argues that the projected compute buildout should favor AI infrastructure, AI labs and GPU localization. However, the investment case depends on AI Cloud demand, financing capacity—especially offshore—and the ability of operators to sustain utilization and attractive ROIC.

Risks

• Geopolitical risk and export-control-related restrictions could affect AI and compute-related entities.
• Intensified competition and price wars could pressure AI-lab economics.
• Supply-side restrictions may create GPU procurement shortfalls.
• Higher interest rates could raise the cost of funding for infrastructure operators.
• Hyperscalers could reduce cloud capex, particularly AI-related investment.
• Capacity-delivery delays and weak sales execution could impair data-center returns.

What to watch

• AI Cloud demand and the pace of hyperscaler capex.
• Progress in funding RMB2.0 trillion of offshore capex.
• Domestic power demand and third-party IDC orders through 2030.
• GPU supply availability, utilization and compute-model ROIC.
• New wholesale contracts, customer move-ins and potential REIT asset monetization for data-center operators.