Data center demand far exceeds supply, 12GW gap by 2025, benefiting the entire industry chain

Overview

This report is published by Jefferies, which believes that the current global data center market faces severe structural supply and demand imbalances. Although demand has surged due to the AI wave, supply is constrained by factors like labor, cooling, power, and equipment, resulting in significant delays in deliveries compared to order demands. A supply gap of approximately 12GW was already evident in North America by 2025, and this trend is expected to intensify over the next few years. The report believes this shortage environment will benefit data center REITs, power equipment suppliers, and emerging developers with fast delivery capabilities, analyzing the constraints at each stage of the supply chain and identifying global investment opportunities.

Core views

Demand Side: Capital Expenditure Surge Driven by AI Capital expenditures of hyperscalars are accelerating rapidly. The report predicts that major cloud vendors (Big 3+META/ORCL) will spend $770 billion on capital expenditures in 2026, a 74% increase year-over-year, nearly five times the level of 2023. Meanwhile, backlogged orders for cloud services reached approximately $2 trillion in Q1 2026, growing much faster than capital expenditures themselves. Based on underlying hardware assumptions, GPU/XPU shipments alone in 2026 imply a massive AI electricity demand of 19.2GW in North America, almost twice the expected additional capacity (10.3GW). By 2028, AI chip-driven electricity demand may reach 63.3GW, while同期新增容量仅为13.1GW, widening the supply-demand gap exponentially. Supply Side: Six Physical Constraints Locking Delivery Limits The report uses its own model to measure that data center deliveries are limited by six physical factors, with different 'shortages' occurring at various stages: 1. EPC/Labor Force (tightest constraint in 2026): Expected to limit the labor force upper bound at 10.4GW in 2026. Despite rapid growth in large contractors' backlog, skilled electricians and construction workers cannot be expanded quickly enough to significantly accelerate project start-up and completion speeds. 2. Cooling Systems (tightest constraint from 2028 onwards): Current upper limit is around 11.4GW. As EPC efficiency improves, cooling manufacturing capacity will become the largest bottleneck after 2028, especially for high-density liquid-cooled components, whose production ramp-up lags behind demand. 3. Power Availability: Grid queue times average 5-8 years, but can be partially alleviated through on-site generation (BTM) and regulated utility models. Power supply upper limit in 2026 is slightly higher than labor and cooling constraints at around 11.6GW. 4. Electrical Equipment & Transformers: High-voltage switchgear delivery periods exceed 12 months, and large power transformers take 24-36 months. Despite increased orders from giants like GE Vernova and Siemens Energy, capacity expansion takes time, constraining projects from being energized in the short term. 5. Backup Generators: Relatively宽松, upper limit in 2026 is 15.2GW, still facing competition from industries like healthcare and defense. Beneficiaries and Investment Strategies Tight supply chains create a 'winner-take-all' landscape: - Data Center REITs (DLR, EQIX): Maintains extremely low occupancy rates of 1-3%, with rising rental and booking volumes. DLR's enterprise-managed bookings have doubled from quarterly $50 million to nearly $1 billion, with a significant boost in AI business. - Power and Equipment Suppliers (CAT, ETN, GEV, PRIM): Order growth is two to four times revenue growth, indicating strong pricing power and earnings elasticity. The report particularly views Primoris (PRIM) as a value revaluation opportunity and Caterpillar (CAT) as a leader in the generator sector. - Emerging Developers (CORZ, WULF): Bitcoin mining companies leverage existing power facilities and land reserves to rapidly transform into AI data center developers, avoiding lengthy grid connection queues and equipment waiting periods, enabling quicker capacity delivery. Regional Opportunities: From the US to Global Expansion Given the challenges in the US domestic market due to NIMBYism and political resistance, along with difficulties in power access, global expansion becomes inevitable. Australia benefits from renewable energy and land advantages becoming a new hotspot; Malaysia, with low land costs and proximity to Singapore, becomes the preferred choice in Southeast Asia; Spain, with abundant renewable energy and submarine optical cable connections, becomes a European data center hub.

Analysis framework

The report adopts an analysis framework combining 'bottom-up hardware push calculation' and 'top-down macro supply-demand modeling.' First, the analyst team derives the specific electricity demand (at the GW level) for AI computing based on semiconductor accelerator (GPU/XPU) shipment data and single-machine power consumption assumptions, serving as hard indicators for the demand side. Next, it constructs a six-dimensional physical capability model for data center deliveries, assessing the supply limits of labor, cooling, power, electrical equipment, transformers, and generators in different years, thereby quantifying the supply-demand gap. Finally, it verifies the authenticity of demand using financial disclosure data (such as Microsoft, Oracle, Meta's uncommitted lease commitments) and identifies geographically advantageous markets and sub-sectors with comparative advantages based on resource endowments and policy environments.

Methodology notes

• Industry/sector Analysis FrameworkSupply and Demand Framework

  Quantifying the difference between demand (calculated AI chip power requirements) and supply (upper limits of six physical constraints) to assess industry momentum

  The report does not solely rely on historical growth rates but calculates the actual electricity demand for AI chips generated by supply limits, providing a clear picture of market scarcity. This is crucial for determining whether the industry is in a seller's or buyer's market.

• Industry/sector Analysis FrameworkUpstream-Midstream-Downstream Transmission

  Analyzing how upstream hardware restrictions cascade through midstream construction and operations, ultimately impacting downstream customer deliveries

  The report clearly outlines the complete chain from GPU shipments (upstream) to electricity demand (midstream input), to EPC/cooling/transformation (midstream construction), and finally to data center launch (downstream output), pinpointing the most vulnerable links that determine the industry's pace.

• Competitive Strategy FrameworkMoat / competitive advantage

  Identifying unique resources (such as pre-existing power and land) enterprises have in resource-constrained environments

  The report particularly highlights the advantage of Bitcoin mining companies transitioning to AI developers due to their existing 'live land' and grid connections, which provide a significant competitive barrier and cost advantage in an otherwise challenging environment for grid connection and equipment procurement.

Asset mapping & comparison

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

• DIGITAL REALITY TRUST (DLR)

  Benefiting from increased leasing demand and rent hikes, with AI-related bookings doubling, strong cash flow

  Strengths

  Wholesale and enterprise mixed platform, AI-related bookings doubling, robust cash flow

  Weaknesses

  Higher valuation, need to watch interest rate environment changes

  Comparison

  Compared to pure wholesalers, stronger cyclicality resilience

  Risks

  If economic downturn leads to tech giants cutting capital expenditures

• EQUINIX (EQIX)

  Global connectivity leader, natural hub for edge computing and inference

  Strengths

  Strong ecosystem network effect, Top 5 new cloud vendors deploying on its network

  Weaknesses

  Expansion limited by site retrofit and new station construction cycles

  Comparison

  Indispensable in interconnection and low latency

  Risks

  Geopolitical factors interfering with international business expansion

• CORE SCIENTIFIC (CORZ)

  Pioneer in Bitcoin to AI development, leveraging existing live facilities for quick delivery

  Strengths

  Largest live land reserve, shorter delivery cycle than traditional developers

  Weaknesses

  Transformation execution risk, dependency on a single technology route

  Comparison

  Responds faster than traditional developers

  Risks

  Volatility in cryptocurrency prices affecting original business cash flows

• TERAWULF (WULF)

  Utilizing power and utilities experience, converting bitcoin mines to AI data centers

  Strengths

  Management experienced in power field, easy access to live land

  Weaknesses

  Smaller scale, potentially higher financing costs

  Comparison

  Focuses on power-rich sites, avoids grid connection queues

  Risks

  Project landing under-delivery

• CATERPILLAR (CAT)

  Leading provider of backup generators, benefiting from high demand for redundant power in AI data centers

  Strengths

  Market share leading, rapidly expanding capacity to meet surge in demand

  Weaknesses

  Raw material cost fluctuations, potential long-term capacity excess risk

  Comparison

  Absolute dominance in large power generation units

  Risks

  Macroeconomic slowdown leading to decreased non-data center demand

• MERLIN PROPERTIES (MRL SM)

  European data center transformation pioneer, beneficiary of Spanish market

  Strengths

  Superior land location, stable energy partnerships, FFO growth expectations high

  Weaknesses

  Limited discount in current valuation, need to digest high growth expectations

  Comparison

  Fastest transformer in European REITs

  Risks

  Fluctuations in local policies or energy prices in Spain

Key data

• Data center supply-demand gap in 2025~12 GWSigned leasing volume (15.6GW) minus actual delivery volume (3.4GW)
• Expected hyperscaler capital expenditure for 2026$770BGrowing 74% year-over-year, nearly five times 2023 levels
• Estimated AI electricity demand in North America for 2026 (chip-based)19.2 GWEquivalent to 1.9 times expected new capacity
• EPC/labor supply upper limit for 202610.4 GWCurrent tightest physical constraint
• Estimated AI electricity demand in North America for 2028 (chip-based)63.3 GWEquivalent to 4.8 times expected new capacity
• Cooling system supply upper limit for 202611.4 GWExpected to become the tightest constraint after 2028

Impact & implications

A常态化的供需失衡意味着数据中心资产将保持极高的利用率和高租金溢价，有利于REITs的盈利质量和估值修复。对于设备商而言，订单积压转化为收入的时间拉长，但价格弹性增强，利润率有望提升。对于投资者，传统的“新建即过剩”逻辑失效，能够解决“最后一公里”交付问题的企业将获得超额收益。同时，政策风险（如美国各州的禁令）将迫使资本加速流向监管友好、资源丰富的海外市场，重塑全球数据中心版图。

Risks

• Regulatory Policy Risk: Multiple states in the US have introduced bills banning data centers, potentially delaying project construction.
• Power Access Delays: Long grid queue times, even with funds, it's difficult to obtain power permits within a short period.
• Supply Chain Breakdown: Long lead times for critical equipment (like transformers, cooling units), causing project delays.
• Reduced Capital Expenditures: If the economy worsens, tech giants may reduce investments in AI infrastructure.
• Geopolitical Conflicts: Impacting global supply chain stability and operational safety of overseas projects.

What to watch

• Quarterly guidance on capital expenditures and backlog of cloud business for major hyperscalars.
• Order backlog and revenue conversion rates for key equipment suppliers (like Vertiv, Eaton, GEV).
• Legislative dynamics and approval progress on data center construction in various US states.
• Further extension of power transformer delivery cycles.
• Actual delivery speed and client signing status of Bitcoin mining companies transitioning to AI projects.