Data center projects are hindered, and the global AI revolution faces an infrastructure bottleneck.

Event Background

The engine of the global AI revolution—data centers—is encountering unprecedented construction obstacles. According to The Guardian, citing data from Uptime Institute, of the 250 large-scale data center projects exceeding 100MW announced between 2021 and 2024, about half may not be completed as planned or face significant delays. From the "Prince William Digital Gateway" in Virginia, USA, which was blocked due to its proximity to a Civil War battlefield, to the cancellation of "Project Range" in Arizona and the "Cyberjaya Campus" project in Malaysia, energy supply, community opposition, supply chain bottlenecks, and inexperienced developers are collectively dragging down the pace of infrastructure expansion.

This predicament directly threatens the computing foundation of the AI industry. Google has publicly acknowledged that its cloud business is "compute-constrained" due to insufficient data centers, unable to meet the growing demand for AI model training and inference. Meanwhile, NVIDIA, the dominant player in AI chips, has seen its GPU shipments surge, but if data centers cannot be brought online in time, the entire value chain will face a capacity mismatch.

Technical Analysis: Why Data Centers Have Become the Bottleneck for AI

AI workloads demand exponential growth in computing power. Training a frontier large model (e.g., GPT-5 level) may require tens of thousands of GPUs running continuously for months, with a single training session consuming tens of megawatt-hours of electricity. Although the energy consumption per inference is lower, concurrent requests from millions of users still require massive GPU clusters for support. All of this depends on high-density, high-reliability data centers providing stable power, cooling, and network environments.

Traditional data centers are typically designed with power densities of 5-10 kW per rack, while AI clusters can reach densities of 40 kW or higher. This means more complex liquid cooling solutions, greater power distribution capacity, and longer construction cycles. A large-scale AI data center usually takes 3-5 years from planning to operation, and the current project backlog and resource competition further lengthen the cycle.

Additionally, data center construction faces grid connection bottlenecks. In California, some completed data centers have sat vacant for years because the grid cannot supply power; in the Netherlands, developers have even sued grid companies for refusing connection. Uptime Institute notes that 80% of new electricity demand comes from U.S. projects, and the local grid is already overwhelmed.

Enterprise Impact Analysis: Costs, Deployment, and Strategic Adjustments

Cost Impact

Delays or cancellations of data center projects directly increase AI companies' capital expenditures (CAPEX) and operating expenditures (OPEX). On one hand, scarce computing resources drive up rental prices—according to industry reports, rent for high-end GPU cloud instances rose by 30-50% in 2025. On the other hand, companies are forced to pay for power capacity reservation fees years in advance, or build on-site power generation facilities (such as natural gas units), further burdening CAPEX.

Deployment ImpactFor cloud users who plan to migrate core AI workloads to large data centers, project delays mean having to accept smaller, more decentralized deployment options or shift to edge computing nodes. This increases management complexity and may impair inference performance due to network latency. Google’s candid admission of being "compute-constrained" epitomizes this dilemma.

Operations and Compliance

The limitation on new data center construction forces enterprises to extend the lifespan of aging facilities, but these older facilities suffer from low energy efficiency and insufficient cooling capacity to support high-density AI clusters. Meanwhile, data center site selection faces growing scrutiny from communities and environmental organizations. As illustrated by the "Prince William Digital Gateway" case, compliance risks related to historical preservation, water consumption, and carbon emissions are significantly rising.

Market Competition Analysis: The Cloud Vendor Landscape Is Shifting

The data center bottleneck is reshaping the cloud service market. The three giants—AWS, Microsoft Azure, and Google Cloud—have invested hundreds of billions of dollars in building new facilities, but project progress varies.

• AWS: With pre-secured power contracts and a globally distributed footprint, it is relatively resilient to risks. Its partnership agreements with nuclear power plants (e.g., direct nuclear power supply with Talen Energy in the U.S.) provide a stable power source.
• Microsoft Azure: Despite its commitment to becoming carbon negative by 2030, rapid expansion still requires a significant amount of fossil fuels as a transition. Some of its projects have faced community opposition, such as protests at multiple European sites.
• Google Cloud: Has publicly acknowledged being compute-constrained and may face market share losses. However, Google’s expertise in liquid cooling and energy efficiency technologies could make its existing facilities more efficient.
• Oracle and Emerging Cloud Vendors: Oracle can flexibly choose sites by deploying relatively small clusters and differentiating itself (e.g., specialized AI clouds), but its scale is limited. Chinese companies like Alibaba and Tencent face similar power challenges in their Southeast Asian projects.

Additionally, data center developers (such as Equinix and Digital Realty) and power companies (such as Constellation Energy) are playing increasingly critical roles. Regions with stable power supplies and fast approval processes (e.g., central U.S. states, the Middle East) will attract more investment.

Industry Trends: Moving Toward Megawatt Scale, but the Road Is Bumpy

Uptime Institute warns that we are entering the era of "gigawatt-scale data centers." In 2024, six projects were planned with power capacities exceeding 5 GW (close to Ireland's peak demand of 6 GW), five of which are in the U.S. and one in the UAE. If these mega-facilities rely on on-site fossil fuel power generation, they will trigger broader carbon emissions controversies.Optimists like Andrew Batson, Head of Global Data Center Research at JLL, believe the industry can overcome challenges: improvements in battery storage technology, on-site power generation (such as natural gas + solar hybrid), and more efficient cooling solutions will alleviate pressure on the grid. He estimates that about 1,200 data centers will be built globally between 2024 and 2030, with AI demand being the main driving force.