AI is pushing Asia-Pacific data centers from server room expansion toward competition in power, cooling, and network density

Overview

This report is Morgan Stanley's AI data center industry course material for Asia-Pacific investors, focusing on demand, supply, business models, and investment mapping for data centers in China and ASEAN under the AI wave. The China section emphasizes demand growth driven by hyperscaler capex, token usage, and rising self-sufficiency in domestic AI chips, while noting that remote nodes with low electricity prices and abundant green energy are better suited for large-scale AI training and current application workloads. The ASEAN section highlights Singapore's role as a subsea cable and connectivity hub, regional land and power resources, and the potential advantages of telecom operators such as Singtel in platformizing AI data centers.

Core views

The report's main theme is that AI is changing the unit economics and site-selection logic of data centers. On the demand side, China's 2026 data center orders and hyperscaler capex are forecast to grow significantly; on the supply side, the AI era places greater importance on large orders above 50MW, high-density racks above 10-30kW, low electricity prices, green energy, speed of construction, and prefabricated delivery. Remote Chinese nodes such as Ulangab and Zhongwei have advantages in low power prices, wind and solar resources, and scale, but investors still need to watch permitting, networks, customer commitments, and policy constraints. In ASEAN, cluster formation is jointly determined by connectivity, subsea cables, land, power, policy, and data sovereignty, with investment opportunities expanding from pure landowners to telecom networks, GPUaaS, liquid cooling, high-density operations, and integrated platforms.

Analysis framework

The report uses a multi-dimensional framework including demand forecasting, supply site selection, TCO cost breakdowns, cloud-era versus AI-era data center comparisons, REIT financing yields, data center KPI bridges, and platform SOTP valuation. It first assesses demand growth, then evaluates which nodes, business models, and companies can convert capacity into leasing, utilization, EBITDA, and ROIC.

Methodology notes

• Demand forecastingHyperscaler capex and order capacity forecasting

  Use capex, ordered GW, and token usage to measure AI data center demand.

  The report cites 66% YoY growth in China's 2026 data center orders, hyperscaler capex of about Rmb600bn, a base-case 3.4GW, and a bull-case 4.5GW to support its view of expanding AI demand.

• Supply and site selectionRemote data center TCO and resource constraint analysis

  Assess data center node competitiveness based on electricity prices, PUE, land, policy, scale, and latency.

  The report compares nodes such as Ulangab, Zhongwei, and Hebei, emphasizing that low electricity prices, green energy, policy support, and scalable campuses in remote areas are more attractive for AI training.

• Operating modelKPI bridge from capacity to financial metrics

  MW capacity must be converted through leasing, utilization, pricing, power pass-through, customer mix, and interconnection services into revenue, EBITDA, and ROIC.

  The report suggests investors should not look only at built capacity, but also at contracted load, billing models, ramp-up pace, PUE, interconnection services, and customer mix.

• Business modelFrom colocation to an integrated AI infrastructure stack

  An AI data center platform is jointly composed of connectivity, AI DC, GPUaaS/AlaaS, and orchestration layers.

  Using Singtel as a case study, the report explains how telecom networks, subsea cables, edge nodes, Nxera data centers, GPUaaS, and platform orchestration can form a low-latency demand funnel and differentiated monetization.

• Valuation and financingREIT yield and SOTP value identification

  Use REIT issuance yields and SOTP breakups to assess asset recycling and platform value.

  The report notes that data center REITs trade at yields of about 3.5-4%, but estimates a fair issuance range of 5-7%; it also believes Singtel Digital InfraCo may account for more than 10% of SOTP, but is not yet fully priced in.

Asset mapping & comparison

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

• Remote data centers in China

  Potential primary carriers of AI training and some current AI application workloads.

  Strengths

  Low electricity prices, abundant green energy, ability to build large-scale campuses, and policy support from some remote regions.

  Weaknesses

  Farther from core cities, with higher dependence on network connectivity, customer migration willingness, and construction permits.

  Comparison

  Compared with tier-one cities, the report gives a CAGR of 60% for remote nodes versus 19% for tier-one cities.

  Risks

  Policy permitting, power reliability, network latency, demand realization, and customer concentration risks.

• Tier-one city data centers in China

  Still serve low-latency cloud business, enterprise demand, and interconnection in core cities.

  Strengths

  Close to customers, low latency, and mature existing ecosystems.

  Weaknesses

  Constrained by power quotas, land, costs, and new supply.

  Comparison

  In the AI era, large orders favor remote low-cost nodes more, and tier-one city growth is lower than that of remote nodes.

  Risks

  Rising costs, supply constraints, and migration of AI training workloads out of core cities.

• Ulangab and Zhongwei

  Representative destinations for remote AI data centers in China.

  Strengths

  Ulangab is close to Beijing with latency of about 4ms, while Zhongwei has larger remote capacity; both have advantages in cool temperatures and wind/solar resources.

  Weaknesses

  Zhongwei is about 1,000km from Beijing with latency of about 8-10ms; both locations still need to prove sustained demand and project delivery.

  Comparison

  Ulangab currently has capacity of about 490k racks and ~1.2GW; Zhongwei has about 300k racks and ~750MW in 2025.

  Risks

  Power, permits, customer rack deployment pace, and regional competition.

• ASEAN data center clusters

  Serve regional AI infrastructure, cloud regions, and data sovereignty demand.

  Strengths

  Singapore's subsea cable hub, regional connectivity, and combinations of land and power resources can create cluster advantages.

  Weaknesses

  Countries differ significantly in power prices, reliability, regulation, taxes, land, and natural risks.

  Comparison

  Compared with a single-market model, the ASEAN model relies more on cross-border connectivity, joint-venture partners, and local telecom resources.

  Risks

  Regulatory changes, data sovereignty restrictions, permitting cycles, natural disasters, and geopolitics.

• Singtel / Nxera / Digital InfraCo

  The ASEAN AI data center platform case in the report.

  Strengths

  Has telecom networks, subsea cables, edge nodes, AI DC platforms, GPUaaS, and orchestration capabilities, enabling a low-latency demand funnel.

  Weaknesses

  Platform value realization depends on capex, partners, local power and land resources, and customer signings.

  Comparison

  Compared with pure data center owners, the Singtel model covers connectivity, capacity, GPU services, and enterprise deployment.

  Risks

  Execution, capex, competition, insufficient valuation rerating, and joint-venture partner coordination.

• Data center REITs

  A tool for changes in asset recycling and financing models for data center operators.

  Strengths

  Can release capital, support new project development, and provide exit or refinancing channels for mature assets.

  Weaknesses

  There is a gap between trading yields and issuance yields, and the report estimates the fair issuance range is higher than current trading yields.

  Comparison

  DC REITs trade at yields of about 3.5-4%, while the estimated fair issuance range is 5-7%.

  Risks

  Rising interest rates, lower asset valuations, insufficient lease quality, and volatility in distribution capacity.

Key data

• China data center orders in 202666% YoY growthThe main text of the report forecasts significant growth in total data center orders in 2026.
• Hyperscaler capex in 2026About Rmb600bn, up 41% YoYUsed to support the assumption of expanding AI data center demand.
• 2026E order capacityBase case 3.4GW; bull case 4.5GWCapacity forecast from the chart footnotes.
• Remote data center growth rate60% CAGR; 19% for tier-one citiesThe section title shows remote node growth significantly outpacing tier-one cities.
• China AI chip self-sufficiency rate70% by 2030EThe chart footnotes show the self-sufficiency rate is expected to rise.
• Ulangab and Zhongwei capacityUlangab about 490k racks, ~1.2GW; Zhongwei about 300k racks, ~750MW in 2025Both locations have average temperatures of 8°C and abundant wind and solar resources, with electricity prices of about 0.26-0.29 yuan/kWh.
• Rack density in the cloud era vs. AI eraAbout 5kW in the cloud era; currently 10-30kW in the AI era and higher in the futureThe AI era also corresponds to 50MW+ hyperscale orders and faster delivery cycles.
• H200 sample TCOTotal cost of 44,224 for Ulangab; total cost of 46,356 for HebeiIn the sample, Ulangab has power prices of 0.40 per kWh and PUE of 1.20, while Hebei has power prices of 0.60 per kWh and PUE of 1.25.
• Data center REIT yieldsTrading yields 3.5-4%; estimated fair issuance range 5-7%Reflects the tension between securitization and financing costs.
• Singtel Digital InfraCoEBITDA growth of more than 30%; Digital InfraCo may account for more than 10% of SOTPThe report believes the related value has not yet been fully reflected.

Impact & implications

The investment implication is that AI data center opportunities should not be screened solely by traditional server room area or rack count; instead, the focus should be on access to power, low-cost green energy, liquid cooling and high-density operations, subsea cable and network connectivity, customer lock-in capability, construction and delivery speed, REIT financing channels, and platform service capabilities. Both remote nodes in China and connectivity hubs in ASEAN benefit from AI compute demand, but returns depend on whether capacity can be converted into long-term leasing, utilization, and ROIC.

Risks

• AI data center orders, token usage, or hyperscaler capex may come in below expectations.
• Power, land, permitting, or construction progress for remote data centers may fall short of expectations.
• The difficulty of improving high-density racks, liquid cooling, and PUE may be greater than expected, affecting TCO and ROIC.
• Customer signings, rack deployment, and utilization ramp-up may lag capacity build-out.
• REIT financing yields may not be attractive, limiting asset recycling and funding sources for new projects.
• ASEAN markets face risks related to data sovereignty, regulation, taxation, natural disasters, and geopolitics.
• The report discloses issues involving U.S. Executive Order 14032 and export controls, and related securities or entities may face compliance restrictions.
• Morgan Stanley has investment banking, shareholding, or other service relationships with some covered companies, and investors should pay attention to conflict-of-interest disclosures.

What to watch

• Whether China's 2026 data center orders, the 3.4GW base case, and the 4.5GW bull case will materialize.
• Whether growth in hyperscaler capex and token usage will continue.
• The rise in domestic AI chip self-sufficiency and its pull on local AI compute demand.
• Electricity prices, PUE, permitting, deployment speed, and customer mix at remote nodes such as Ulangab and Zhongwei.
• As AI rack density continues rising beyond 10-30kW, the implications for liquid cooling, power, and networks.
• Whether actual issuance yields for data center REITs can fall within the feasible 5-7% range.
• Whether Singapore's subsea cables and ASEAN cross-border connectivity will continue supporting regional AI infrastructure clusters.
• Singtel Digital InfraCo's EBITDA growth, share of SOTP, and whether the market will reprice it.