Power Emerges as Critical Bottleneck for AI Data Centers; Growth Opportunities Across Multiple Asian Markets

Overview

This report synthesizes key insights from Goldman Sachs’ May 2026 Asia TMT Conference, focusing on the evolving supply-demand dynamics in the data center industry under the AI wave. The conference featured senior executives from three operators—Empyrion Digital, Digital Edge, and PaleBlueDot AI—who shared frontline perspectives. The core conclusion is that while GPU shortages are expected to ease by 2028–2029, power availability has become the primary bottleneck constraining data center construction and delivery. Meanwhile, the rapid demand cycle from AI-native customers is reshaping industry operating standards, with strong growth potential emerging in Asia-Pacific markets such as India and Japan.

Core views

Power Bottleneck and Public-Private Collaboration Models: Grid infrastructure development typically requires 5–10 years, making it a longer-term constraint than GPU availability. To address this, governments must take the lead, while private operators and hyperscalers invest directly in power generation and renewable energy to bridge the gap. Malaysia is cited as an effective example of public-private collaboration. Additionally, markets must balance energy efficiency (PUE) against water consumption based on local resource endowments—for instance, Digital Edge achieves a PUE below 1.25 in Mumbai by using treated wastewater instead of potable water. Community acceptance is now recognized as the fifth critical infrastructure pillar, following power, water, connectivity, and permitting. Diverging Customer Needs and Operational Strategy Adjustments: The client landscape has shifted significantly in the AI era. Single-tenant GPU customers typically lock in orders 15–24 months before facility construction begins, whereas traditional multi-tenant clients require only 3–6 months of lead time. Emerging AI-native customers (Neoclouds) compress contract signing to just 6 weeks—much faster than the 4–6 months typical for traditional hyperscalers. To accommodate these differences, operators commonly adopt modular designs and hybrid air-liquid cooling systems as standard configurations, support rack densities ranging from 20kW to over 1MW, and delay final fit-outs as long as possible to retain flexibility. Delivery speed has become a core competitive advantage—PaleBlueDot, for example, delivered 6,000 GPUs in Japan within 89 days. Regional Growth Outlook and Demand Structure: India is highly promising due to its demographic dividend, engineering talent pool, and proximity to the Middle East; Digital Edge is already advancing a 300MW campus project there. Japan is expanding aggressively with strong government backing, while the Philippines attracts attention due to streamlined permitting and scalable power infrastructure. On the demand side, inference now accounts for approximately 70% of AI compute demand, with significant room for further expansion as application-layer use cases mature. The industry broadly views AI-driven incremental demand as non-zero-sum—traditional cloud businesses will continue growing in parallel, and supply-side consolidation is expected to enhance efficiency and alleviate resource constraints.

Analysis framework

The report employs an integrated framework combining 'bottleneck identification across the value chain,' 'customer behavior comparison,' and 'regional differentiation analysis.' First, by contrasting GPU normalization timelines (2028–2029) with grid build-out cycles (5–10 years), it precisely identifies the industry’s core constraint as having shifted from hardware shortages to infrastructure readiness. Second, through quantitative comparisons of contracting timelines and technical requirements between traditional hyperscalers and AI-native clients, it reveals the business logic behind operators’ need to maintain asset 'replaceability' and flexibility. Finally, by evaluating national policies, resource endowments, and geopolitical factors, it pinpoints regions with structural growth opportunities. This approach helps investors understand that in the AI infrastructure cycle, mere capacity expansion is insufficient—resource integration capability and delivery speed now constitute the new moat.

Methodology notes

• Industry/Value Chain Analysis FrameworkSupply-demand framework

  Core Bottleneck Shift Analysis

  By comparing the GPU supply normalization timeline (2028–2029) with the 5–10 year grid infrastructure build-out cycle, the report concludes that the industry’s primary constraint has shifted from equipment shortages to power infrastructure limitations—a classic case of dynamic supply-demand mismatch analysis.

• Competitive & Strategic Framework

  Asset Replaceability and Flexibility Strategy

  Faced with vastly different demand cycles between AI and traditional cloud clients, operators maintain asset versatility through hybrid cooling, wide rack density compatibility, and delayed fit-outs—avoiding sunk costs from over-optimization for a single workload type.

• Industry/Value Chain Analysis FrameworkPenetration S-curve

  Inference Demand Share and Growth Stage Assessment

  By tracking the 70% share of inference demand, the report infers that AI compute consumption is transitioning from early-stage model training to large-scale application deployment, signaling accelerating downstream commercialization.

Asset mapping & comparison

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

• Digital Edge

  Conference participant showcasing localized operational strategies

  Strengths

  Achieves PUE <1.25 in Mumbai using treated wastewater; supports rack densities from 20kW to over 1MW; developing a 300MW campus in Navi Mumbai, India

  Comparison

  Compared to pure AI-focused operators, places greater emphasis on compatibility between AI and traditional workloads and adaptation to local resource constraints

• Empyrion Digital

  Conference participant providing insights on customer segmentation and community relations

  Strengths

  Identified the trend of single-tenant AI clients locking in orders 15–24 months in advance; proposed community acceptance as the fifth infrastructure pillar; advocates that AI demand represents non-zero-sum growth

  Comparison

  Emphasizes delaying facility fit-outs to balance flexibility between AI and traditional customer requirements

• PaleBlueDot AI

  Conference participant validating inference demand surge and ultra-fast delivery capability

  Strengths

  Delivered 6,000 GPUs in Japan in 89 days; inference accounts for 70% of demand; positions itself as a GPU-as-a-Service provider converting energy into tokens

  Comparison

  Delivery speed significantly exceeds industry average, highlighting the agility advantage of AI-native service providers

Key data

• AI Customer Contract Signing Cycle6 weeksSignificantly shorter than the 4–6 months typical for traditional hyperscalers, reflecting urgent AI demand
• Grid Infrastructure Development Cycle5–10 yearsFar exceeds GPU supply recovery timelines, now the hardest constraint on data center delivery
• Potential Supply Absorption Period3–6 monthsIf unconstrained, current demand would deplete available capacity within this timeframe
• Inference Demand ShareApproximately 70%Per PaleBlueDot data, AI compute demand has shifted decisively toward inference
• Mumbai Project PUE<1.25Achieved by using treated wastewater instead of potable water in a water-scarce region
• Rack Density Compatibility Range20kW to >1MWDigital Edge facilities support extremely wide power density ranges to accommodate diverse workloads

Impact & implications

For data center operators, securing stable power resources and rapid delivery capability have superseded simple cabinet scale as core valuation drivers. Companies with public-private collaboration experience, hybrid cooling expertise, and the ability to respond flexibly to short-cycle AI customer demands will command premium valuations. For upstream equipment vendors, products supporting high density, liquid cooling, and modular deployment will become standard. Regionally, India, Japan, and the Philippines—each with unique policy or resource advantages—are positioned to capture the next wave of AI infrastructure spillover demand and emerge as new growth poles in Asia-Pacific.

Risks

• Expansion delays due to constraints in power, water, and permitting approvals
• Rising community opposition potentially blocking new projects or increasing compliance costs
• Tighter geopolitical regulations raising supply chain due diligence burdens and operational complexity
• Over-optimization for AI workloads potentially reducing compatibility with traditional cloud loads

What to watch

• Concrete progress in public-private partnership (PPP) models for grid infrastructure development across countries
• Ongoing divergence in contracting terms and delivery expectations between AI-native and traditional hyperscaler clients
• Policy support intensity and power capacity expansion in key markets like India, Japan, and the Philippines
• M&A and consolidation activity among data center operators and its impact on supply efficiency
• Commercialization pace of inference-layer applications and further evolution of compute consumption patterns