Asian AI data center demand enters a high-growth cycle, with power and computing infrastructure becoming key bottlenecks

Overview

This is a Morgan Stanley investor presentation on Asian AI data centers, covering data center demand, supply, business models, and value-chain mapping in China and ASEAN. The report’s core view is that computing demand from AI training and inference is reshaping data center site selection, power, cooling, networking, and capex structures. Investment opportunities extend beyond data center owners to power, storage, equipment, GPUs, servers, ODM/EMS, networking, optical modules, cooling, telecom operators, and REITs.

Core views

The report believes Chinese data center demand will continue to be driven in 2026 by hyperscaler capex and the ramp-up of domestic GPUs, while remote data centers are becoming a supply priority because of low electricity prices, green energy, and scale advantages. In ASEAN, data centers are upgrading from traditional colocation infrastructure to AI workload platforms, with key differentiation arising from power density, cooling solutions, network connectivity, sovereign AI demand, and low-latency enterprise deployment. Investors should assess value through the chain of “capacity—leasing—utilization—EBITDA—ROIC,” rather than focusing only on nominal MW scale.

Analysis framework

The report combines demand forecasts, supply constraints, TCO comparisons, workload segmentation, KPI bridging, and value-chain mapping. It first estimates hyperscaler capex, GPU TAM, and data center orders, then compares electricity prices, latency, and capacity at remote nodes. It subsequently uses MW, PUE, utilization, contract structures, and ROIC to explain business model differences, and maps AI data center demand to power, servers, semiconductors, networking, cooling, and telecom connectivity assets.

Methodology notes

• Industry demand forecastingHyperscaler capex and order forecast

  Use hyperscaler capex and data center orders to measure the intensity of AI infrastructure demand.

  The report provides approximately Rmb600bn of hyperscaler capex, 3.4GW of orders, and a 4.5GW bull-case scenario for 2026E to assess the elasticity of Chinese data center demand.

• Cost analysisCustomer TCO comparison

  Compare rent, electricity prices, server depreciation, load factors, and PUE across different data center locations.

  Using H200 and data centers in Ulanqab and Hebei as examples, the report shows how lower electricity prices and energy efficiency differences at remote nodes affect customers’ total cost of ownership.

• Operating KPICapacity -> Leasing -> Utilization -> EBITDA -> ROIC

  Translate capacity construction into leasing, utilization, EBITDA, and return on invested capital.

  This framework emphasizes that revenue depends on contracted capacity, billing models, power pass-through, ramp-up pace, interconnection services, and customer mix.

• Workload segmentationTraining vs Inference workload management

  AI training is better concentrated in areas with low-cost power, while inference relies more on low latency, data sovereignty, and proximity to users.

  The report cautions against treating every MW as the same asset because training and inference have materially different requirements for site selection, cooling specifications, capex, and contract structures.

• Value-chain mappingAI data center value chain

  Extend the AI data center opportunity to power generation, grids, storage, GPUs, servers, networking, optical modules, cooling, and telecom.

  The report lists participants across the global AI infrastructure supply chain and emphasizes that equity exposure spans a supply-chain portfolio rather than only data center landlords.

Asset mapping & comparison

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

• Chinese remote data center operators and nodes

  AI training demand, low-cost green energy, and policy support are driving remote data center construction.

  Strengths

  Nodes such as Ulanqab and Zhongwei offer large-scale capacity, relatively low electricity prices, abundant wind and solar resources, and relatively acceptable latency for training.

  Weaknesses

  Distance from core users may constrain inference and low-latency enterprise workloads, while capacity absorption depends on sustained hyperscaler and AI training demand.

  Comparison

  Compared with data centers in core cities, remote nodes have advantages in power costs and scale but are weaker in customer proximity and interconnection ecosystems.

  Risks

  Power consumption, network connectivity, policy execution, utilization ramp-up, and customer concentration risks.

• Data center REITs

  As a capital recycling and issuance/exit mechanism for data center assets, their yield pricing is discussed in the report.

  Strengths

  They can release capital for operators, support reinvestment, and provide investors with exposure to digital infrastructure cash flows.

  Weaknesses

  The report states that DC REIT transaction yields are approximately 3.5-4%, while an estimated 5-7% may be a more reasonable issuance range, highlighting tension between valuation and issuance feasibility.

  Comparison

  Compared with traditional real estate REITs, data center REITs offer stronger growth but also greater exposure to technological change, customer concentration, and electricity costs.

  Risks

  Interest rates, issuance yields, asset injection quality, tenant concentration, and regulatory approval risks.

• Singtel Digital InfraCo / Nxera / STT GDC

  Representative assets for ASEAN AI data center platformization and telecom connectivity capabilities.

  Strengths

  They offer telecom networks, submarine cables, edge connectivity, AI-ready campuses, and multi-market capacity from STT GDC. The report states that Digital InfraCo EBITDA is growing by more than 30%.

  Weaknesses

  Continued capex is required. STT GDC is accounted for under the equity method, limiting near-term dividend impact, and transaction completion remains pending until 2H26.

  Comparison

  Compared with pure-play data center operators, Singtel has telecom networks and access to enterprise customers; compared with traditional telecom operations, digital infrastructure offers higher growth but is more capital intensive.

  Risks

  Transaction closing, capex overruns, pipeline development, competition, customer ramp-up, and unrealized valuation risks.

• Power, utilities, storage, and grid equipment

  AI data centers make power supply, 24x7 reliability, and grid connectivity core constraints.

  Strengths

  Power generation, clean energy, nuclear power, storage, grid equipment, and high-voltage distribution companies may benefit from incremental data center load.

  Weaknesses

  Returns depend on long-term power purchase agreements, regulated electricity tariffs, project approvals, and grid access, while electricity markets vary significantly across countries.

  Comparison

  Compared with IT hardware, power assets have more fundamental and longer-duration demand, but their growth elasticity and pace of valuation re-rating may be lower.

  Risks

  Electricity price volatility, fuel prices, grid connection delays, environmental permits, capacity shortages, and policy changes.

• GPU, server, ODM/EMS, networking, and optical module supply chains

  Higher AI rack power density and cluster scaling directly drive demand for GPUs, servers, HBM, high-speed networking, optical modules, and power components.

  Strengths

  The report lists Nvidia InfiniBand, Arista, Juniper, Ciena, Innolight, Eoptolink, Delta, Quanta, and Wiwyn among the relevant segments, showing that demand is spreading broadly across the hardware supply chain.

  Weaknesses

  The supply chain is significantly affected by technology paths, customer bargaining power, export controls, and inventory cycles.

  Comparison

  Compared with data center landlords, the hardware chain is more sensitive to AI capex; compared with utilities, it faces greater cyclicality and technology substitution risk.

  Risks

  Export controls, GPU supply bottlenecks, technological iteration, margin compression, customer concentration, and capex cuts.

• Cooling and power distribution equipment

  As AI GPU rack density rises from 40-200kW toward several hundred kW or higher, demand is increasing for liquid cooling, direct-to-chip cooling, immersion cooling, and high-current power distribution.

  Strengths

  High-density racks create structural equipment upgrade demand, shifting cooling from room-level planning toward rack-level thermal management.

  Weaknesses

  Technical standards are still evolving, and different customers and chip platforms have different cooling architecture requirements.

  Comparison

  Compared with traditional air-cooling equipment, liquid cooling and high-density power distribution solutions have higher value but also longer validation cycles and greater operational complexity.

  Risks

  Technology path changes, reliability, maintenance costs, customer certification cycles, and project delays.

Key data

• 2026E China hyperscaler capexApproximately Rmb600bn, up 41% YoYUsed to support the forecast for Chinese AI data center demand.
• 2026E China data center orders3.4GW; 4.5GW in the bull-case scenarioThe report title indicates that total data center orders are expected to grow 66% YoY.
• Domestic GPU TAMExpanding to US$70bn by 2027EThe report views domestic GPUs as a more important demand driver.
• Domestic GPU self-sufficiency rateRising to 70% by 2030EReflects the localization trend in China’s AI computing supply chain.
• Total Chinese data center capacityReached 34GW in 2025The report states that data centers consume approximately 2% of total social electricity consumption.
• Ulanqab remote data center capacityApproximately 490k racks, approximately 1.2GWApproximately 300km from Beijing, with latency of approximately 4ms and electricity prices of approximately Rmb0.26-0.28/kWh.
• Zhongwei remote data center capacityApproximately 300k racks, approximately 750MWApproximately 1,000km from Beijing, with latency of approximately 8-10ms and electricity prices of approximately Rmb0.27-0.29/kWh.
• Traditional enterprise rack power density5-15kWLower than AI GPU racks.
• Current AI GPU rack power density40-200kWDriving cooling from room-level planning toward rack-level thermodynamics.
• Next-generation AI platform power densitySeveral hundred kW or higherRequires high-current power supply, liquid cooling, immersion cooling, and high-speed networking.
• STT GDC transaction platform scale50 data centers, approximately 673MW operating capacity, 12 marketsThe consortium acquisition involving Singtel brings scarce operating capacity and a customer base.
• STT GDC design capacity and pipelineApproximately 2.8GW of design capacity; approximately 1.7GW of cross-market pipelineThe report states that this could accelerate the development of a global AI data center platform.
• Singtel investment in STT GDCS$740mn for a 25% stake; additional capex of approximately S$400-500mn over the next three yearsManagement expects the transaction to close in 2H26.
• Singtel valuation read-throughApproximately 4% increase in enterprise valueBased on Morgan Stanley’s estimate.

Impact & implications

The investment implication is that the AI data center value chain is shifting from standalone colocation capacity toward integrated infrastructure combining “power + computing + networking + software orchestration + sovereign AI services.” The Chinese market is more focused on domestic GPUs, remote nodes with low-cost electricity and green energy, and data center REIT exit mechanisms. The ASEAN market is more focused on cross-border connectivity, submarine cable hubs, telecom edge networks, GPUaaS/AIaaS, and multi-market platform expansion. For portfolios, the report recommends identifying beneficiaries from a supply-chain perspective rather than buying only data center operators.

Risks

• AI demand or hyperscaler capex falls below expectations, causing data center orders and utilization ramp-up to undershoot expectations.
• Although remote data centers have low electricity prices, they may face risks related to network connectivity, latency, customer proximity, and actual load absorption.
• High-density AI racks impose higher requirements on power supply, cooling, land, permitting, and grid access, potentially delaying project delivery or causing cost overruns.
• The gap between data center REIT issuance yields and market transaction yields may affect securitization and capital recycling.
• Export controls, U.S. Executive Order 14032, and related compliance restrictions may affect the investment feasibility of certain entities or securities.
• The path toward greater domestic GPU self-sufficiency remains uncertain because of technology, supply, customer adoption, and external restrictions.
• Multi-market expansion in ASEAN involves differences in regulation, land, power, foreign exchange, partners, and sovereign data rules.

What to watch

• Whether 2026E Chinese hyperscaler capex approaches the approximately Rmb600bn forecast.
• Whether Chinese data center orders reach the 3.4GW base-case scenario or 4.5GW bull-case scenario.
• Whether domestic GPU TAM and the self-sufficiency rate follow through on the path toward US$70bn by 2027E and 70% by 2030E.
• Whether the rack deployment rate, customer signings, utilization, and electricity-price advantages of remote nodes such as Ulanqab and Zhongwei persist.
• Changes in the mix of AI training and inference workloads, and whether inference drives demand for regional hubs and edge data centers.
• The pace at which high-density rack cooling migrates from air cooling toward direct-to-chip, immersion, and liquid-cooling systems.
• Whether the Singtel-STT GDC transaction closes in 2H26 as scheduled, and whether Digital InfraCo EBITDA and valuation contributions materialize.
• Changes in data center capacity, submarine cable connectivity, and power approvals in ASEAN countries including Malaysia, Thailand, and Singapore.