AI Data Center Construction Pipeline: 2026-2028 Planned Capacity by Region

The global AI data center construction pipeline has reached unprecedented scale. As of Q2 2026, 458 data center facilities dedicated to or substantially supporting AI workloads are in active development (site selection through construction), representing 85.4 GW of planned IT load capacity. This compares to approximately 62 GW of total operational data center capacity globally at year-end 2025, meaning the construction pipeline alone will nearly triple total capacity if fully realized. The United States leads with 38.2 GW (44.7% of global pipeline), followed by Europe (18.6 GW, 21.8%), Asia-Pacific excluding China (14.2 GW, 16.6%), China (8.5 GW, 10.0%), and Rest of World (5.9 GW, 6.9%). The pipeline has grown 72% from the 49.6 GW tracked in Q2 2025, reflecting hyperscaler and neocloud capital expenditure acceleration.

However, only 28% of planned GW is fully de-risked (funded, permitted, and contracted with construction firms and equipment suppliers). A further 34% is in advanced planning with preliminary funding and site control. The remaining 38% is speculative: early-stage development without secured power, permits, or financing. Based on historical data center construction delivery rates, we project that 62-68 GW of the 85.4 GW pipeline (73-80%) will actually reach operation by 2030, with a risk-adjusted delivery timeline skewed toward 2028-2030 rather than the 2026-2027 target dates many developers advertise. The two primary constraints are power availability and GPU delivery lead times, neither of which can be compressed easily.

Power availability is the binding constraint on AI data center construction globally. A single 500 MW AI data center consumes as much electricity as 200,000 US households. The 85.4 GW pipeline would require approximately 2,900 MW of new power generation dedicated to data centers annually through 2030, equivalent to adding three nuclear reactors or 1,500 MW of solar with 4-hour battery storage per year. US utility interconnection queues show 380 GW of data center capacity in active interconnection requests, more than the pipeline itself, indicating significant oversubscription and competition for grid capacity. Average interconnection lead times for data centers have stretched from 2.4 years in 2022 to 4.8 years in 2026, pushing risk-adjusted delivery timelines to 2029-2031 for new grid connections.

Developers are responding with on-site power generation. In the US pipeline, 35% of planned MW includes on-site natural gas generation, 12% includes behind-the-meter solar with battery storage, and 8% includes nuclear power purchase agreements (both existing plant PPAs and advanced nuclear SMR pre-purchases). Microsoft's Three Mile Island restart PPA (835 MW, contracted through 2028) and Amazon's Talen Energy data center campus co-located with the Susquehanna nuclear plant (960 MW direct connection) exemplify the nuclear co-location trend. However, these solutions are U.S.-centric: in Europe, where carbon pricing adds $40-80/MWh to fossil generation, and grids are less flexible, 62% of planned capacity lacks secured power agreements, the highest percentage of any region.

AI data centers have fundamentally different density profiles than traditional colocation. Traditional enterprise data centers average 5-8 kW per rack; AI training clusters average 40-70 kW per rack for H100 air-cooled and 80-120 kW per rack for B200 liquid-cooled deployments. A typical 100 MW AI data center supports 25,000-35,000 H100-equivalent GPUs at 40-70% utilization, depending on cooling efficiency and cluster design. Liquid cooling (direct-to-chip and immersion) is now standard for new builds: 78% of 2026 construction pipeline capacity specifies liquid cooling, up from 32% in 2024. The shift from air to liquid cooling allows 2-3x higher rack density, reducing the facility footprint per GW by 40-50%.

The GPU cluster architecture in new builds has standardized to 4,096-GPU "super-pod" modules connected via NVIDIA DGX Quantum InfiniBand or Spectrum-X Ethernet. Each 4,096-GPU super-pod consumes 2.8-3.5 MW (H100) to 4.5-5.5 MW (B200) and occupies 4,000-6,000 sq ft of data center white space. A 100 MW facility typically hosts 20-30 GPU super-pods, with 20-30% of capacity dedicated to inference (lower density, higher rack count) and 70-80% to training (high density, liquid-cooled). The modular super-pod design enables phased deployment: infrastructure built in 20-30 MW phases matching GPU delivery schedules, reducing the upfront capital commitment and allowing providers to bring capacity online as supply chains deliver GPUs.

AI data center construction costs have risen 35-55% since 2022 due to supply chain inflation, labor shortages, and the complexity of liquid cooling and high-density power distribution. Average construction cost per MW has reached $10-15 million including site acquisition, shell, power infrastructure, cooling, and fiber connectivity, up from $7-9 million per MW in 2022. A fully equipped GPU data center including servers, networking, and GPUs costs $35-55 million per MW, 70-80% of which is GPU hardware. The total capital required to build the risk-adjusted deliverable pipeline of 64,000 MW is $540-780 billion for construction plus GPU equipment costs of $1.4-2.2 trillion at current GPU pricing.

The capital intensity is driving financial innovation. The first GPU data center REIT (Real Estate Investment Trust) launched in Q1 2026 (GPU REIT, ticker: GPUR) with $3.2 billion in assets, offering 4.8% dividend yield backed by GPU lease cash flows. Data center construction loans have moved from traditional 3-5 year terms to 5-7 year terms with construction-to-permanent structures that match the longer GPU depreciation schedule. The capital formation challenge is significant: to fully realize the construction pipeline, the industry needs to raise $250-350 billion in new equity and debt through 2030, equivalent to 30-40% of all global data center investment historically. This capital requirement will constrain marginal neocloud providers and favor large balance sheet operators with cheaper access to capital.