Datacentres are now designated as critical infrastructure in the United States, the UK, the EU, China, and other countries. National AI training and inference capacity is viewed as vital to competitiveness and security, and has prompted governments to encourage AI infrastructure development.
The industry has responded with a surge of project announcements that is straining energy and water systems and fueling public opposition to large, often confidential, campus plans. With limited evidence of project viability, this wave of proposals has disrupted energy and water infrastructure planning.
A central challenge is uncertainty over future demand for datacentre capacity. Large and midsize technology companies are building campuses to capture projected AI growth through unique, leading-edge training models and inference-based applications for business and consumer uses.
AI’s market opportunity has likely been underestimated and will expand as new applications emerge. However, individual companies may be overestimating the share they can capture. Each planned campus competes with other planned capacity, increasing the risk of overbuilding relative to actual future AI demand growth.
Uptime Institute analysis found that of 479 large datacentre projects (100MW or more of power demand) announced globally from 2021 to 2025, 10% have at least one building in operation (as of July 2026), 41% have begun construction of the first phase of the project, 40% are stalled or delayed, and 9% have been cancelled. Actual construction and completion of operating facilities is moving slowly.
Overall, at least half are expected either not to be built or not to reach their projected full energy demand. Uncertainty remains high about the feasibility of many of the announced projects.
IT Infrastructure
Current discussion of datacentre capacity expansion often centres on high-density AI training infrastructure, including racks of 150kW or more. In practice, growth will span standard compute for traditional public cloud services, AI inference, and AI training infrastructure (Figure 2).
AI training and GPU-based inference environments may require liquid cooling and 480V or 800V electrical infrastructure. Standard compute and ASIC-based AI inference workloads are more likely to use air cooling or supplemental direct liquid cooling with low-voltage electrical infrastructure. Uptime Intelligence analysis of announced projects indicates that 60% to 80% of planned capacity will support lower-density racks of 20kW to 80kW for standard compute and AI inference.
Project viability
For government and infrastructure level project planning to be effective, the planning process has to be able to identify and act on those projects that are truly viable. Several factors can serve as indicators of a project’s viability:
Experienced datacentre developer: Cleanview and others estimate that approximately 50% of all project proposals are submitted by developers with no datacentre experience. An inexperienced developer will not fully grasp the intricacies and complexities of datacentre development. They will struggle to attract tenants and navigate the design, permitting, and contracting processes.
IT operations tenants to occupy the facility: Many projects are proposed without having a contracted IT operator to occupy the facility. Without designated tenant(s) contractually committed to occupying the facility for a specified period of time, a developer is less likely to be able to finalise their financing or engage in the planning process.
Contracted power capacity: Projects with 500MW or more of power demand will need to bring new generation assets onto the grid or the datacentre site to satisfy the facility’s power needs. The datacentre developer must work with the transmission system operator, a utility or energy retailer, and developers of generation assets – on-site and grid-based – to implement an identified solution that meets the proposed facility’s power demand. A viable project will have a clear path to procure power when it is needed.
Project implementation timelines: Projects of 1,000MW or more are likely to be implemented in phases over five to fifteen years. The buildout of a given project will be done in phases, with each phase needing to address key criteria: committed tenant(s), contracted power and water supplies, and access to critical equipment. Government entities and energy and water system planners need to recognise that future phases of a project may not materialise if the planned capacity cannot capture sufficient revenue to justify the project.
Electrical grid interconnection: Grid interconnection times for datacentre projects range from two to seven years or more. Developers can contract for specific new or existing grid capacity (where available) to meet immediate power needs and/or add new grid assets over time to enable or accelerate interconnection. They may start their facility using permitted, on-site generation before the grid interconnection is completed, or agree to participate in demand response programs (removing some or all of the datacentre demand from the grid during grid events) to accelerate the interconnection timeline.
Access to critical equipment: The rapid expansion of datacentre infrastructure has stressed the supply chain for critical cooling (heat rejection) systems, IT space heat removal systems, critical electrical infrastructure such as transformers and uninterruptible power systems (UPS), and standby generation systems. Lead times for this equipment can be two years or more. Developers need committed delivery dates on these infrastructure components to have a viable project.
Governments and electricity grid and water infrastructure operators are struggling to quantify and understand these variables as they work to establish effective planning and permitting processes. They need to rethink their approach, and emphasise the need for developers to demonstrate project viability as projects are proposed and move through the approval process.
An effective project planning process
The explosion of datacentre requests has caught planning, grid, and water authorities unawares. The magnitude of facility size and resource consumption is unexpected, and the implications of large, closely-located projects can be difficult to assess. Public authorities can mitigate these challenges by establishing clear, consistent assessment processes that will enable planning and permitting authorities to examine critical indicators and interactions between announced projects and existing datacentres.
In turn, datacentre operators need to be more open about the details of project proposals. The use of non-disclosure agreements (NDAs) and tight control of project data during initial permitting and approval stages has proved counterproductive. Datacentre operators need to clearly detail the key attributes of their project and engage the responsible authorities and the public with a factual explanation of the project and its impact on the local community and its energy and water infrastructure.
The electrical grid, water supply system, and local municipal and state/province authorities need to be able to assess the following project attributes.
Developer qualifications: The developer should detail the experience and qualifications related to datacentre projects for its team and project team members, including the engineering consultant, construction manager and contractor, and energy system developer. The developer should also disclose the facility operator, either the IT operator (see figure 1 for examples) or the colocation operator that will run the facility and manage the IT operators (tenants) that will occupy the facility.
IT infrastructure type: Developers must explain the type of IT infrastructure the datacentre will host. Standard compute, AI inference, and AI training infrastructure will have distinct energy and water use profiles.
External heat rejection system technology: The type of external heat rejection (cooling) system chosen for the facility sets the facility’s energy and water use profile. Developers should identify whether the system uses dry cooling or evaporative heat rejection and explain how the chosen external heat rejection system optimises and minimises energy and water consumption at the chosen facility location.
Projected energy and water demand and consumption: Projected energy and water demand and consumption for the proposed facilities are critical to understanding the project’s impact on local infrastructure. Monthly, rather than annual values, are critical. Facility system water and energy use will vary by season, with values low in the winter and high in the summer. Infrastructure capacity limitations manifest during periods of extreme cold or extreme heat; infrastructure planning needs to address the 100 to 400 hours of maximum energy and water consumption.
Contracted portfolio of energy assets: Large datacentre projects should identify the dedicated grid generation and transmission assets needed to meet their power demand. Operators should work with developers, utilities, and energy retailers to specify the existing or new grid assets required to serve the site.
Generation assets may be phased to match the buildout of buildings and IT infrastructure over three to 20 years. Where needs extend beyond the typical three- to five-year planning horizon, the near-term power plan should cover committed buildings, with dates for confirming later phases and assigning specific generation assets.
On-site generation assets: Most datacentres use standby generators to maintain operations when grid power is unavailable. At facilities of 500MW or more, these systems are comparable to on-grid power plants that require significant investment and create material emissions. Operators should procure systems with suitable pollution controls and permit them to operate continuously so they can support grid stability through demand response programs.
Air quality assessments should evaluate standby generator emissions, account for other sources in the airshed, and model the longer, recurring operation needed for demand response. For example, an area with several large campuses could host 3GW or more of standby generation that will operate simultaneously during a grid disturbance. Developers should provide a complete assessment of local generation assets’ air quality impacts.
Controlled ride-through of grid disturbances: Datacentres protect IT infrastructure from grid-related voltage or frequency disturbances by conditioning power through an uninterruptible power system (UPS) and rapidly disconnecting from the grid when disturbances exceed set tolerances. For campuses of 500MW or more, or dense datacentre hubs such as are found in Virginia, Dallas, Atlanta, Ireland, and elsewhere, simultaneous disconnection can create supply-demand imbalances that further destabilise the grid and could trigger blackouts.
Grid authorities and standards bodies are defining requirements for disturbance ride-through and disconnect/reconnect processes that protect IT infrastructure while maintaining grid stability. Developers should explain how electrical system hardware and planned operating processes will conform to these emerging requirements and support grid resilience.
Noise study: Large datacentres will have several sources of noise, including construction noise, cooling tower fans, standby generators or other on-site power systems, and periodic vehicle traffic. The developer should perform a noise study to demonstrate that boundary noise conditions do not exceed local noise ordinances or accepted best practices. Developers should publicly release the noise study and explain the engineering controls and mitigation measures incorporated into the facility design to maintain acceptable boundary noise levels.
Government, power grid, and water system planning and permitting authorities need to access and understand the full scope of datacentre projects as described above. Datacentres are complex facilities, and developers and operators are clustering them around network connections and large metro areas to provide the best service to their customers.
The rapid growth of AI and cloud infrastructure requires a more disciplined approach to datacentre planning. Project announcements alone are no longer sufficient for grid, water, and planning authorities to assess infrastructure impacts or allocate scarce resources. Developers should be expected to demonstrate project viability through committed tenants, credible buildout schedules, secured power and water strategies, defined cooling and IT infrastructure plans, and transparent assessments of emissions, noise, and grid stability impacts.
A clearer, evidence-based planning process will not eliminate uncertainty, but it can separate viable projects from speculative proposals and reduce unnecessary strain on public infrastructure planning. By requiring consistent disclosure and phased commitments, governments and utilities can better protect communities while enabling the datacentre capacity needed to support AI, public cloud growth, and broader digital services.
Jay Dietrich is research director of sustainability at the Uptime Institute.
