Quick answer: AI data center power planning must address both the electricity available to a site and the infrastructure that delivers it to compute racks. Grid connection timing can constrain expansion, while denser AI systems place new demands on distribution and cooling. Phased deployment and emerging 800V DC architectures address different parts of this challenge and should be evaluated together.

An AI data center can have space for more servers and still lack the power needed to operate them. The limiting factor may be the date when additional grid capacity becomes available, the rating of an electrical distribution path or the cooling capacity serving a particular row.

The IEA’s 2026 report on energy and AI examines growing electricity demand alongside the ability of grids and equipment supply chains to respond. For project developers, this makes power availability an early site-selection and scheduling consideration. IEA energy and AI analysis

A successful power plan needs to connect the utility commitment to capacity that can actually be commissioned and operated at each stage.

Start with available capacity and connection timing

A proposed campus capacity is different from an energized connection. Before setting a launch date, establish the capacity the utility can supply, the conditions attached to that supply and the works required to make it available.

The programme should identify dependencies such as substation completion, transformer delivery and connection testing. Treat unconfirmed expansion capacity separately from the supply available for the first operational phase.

Consider a project intended to expand over several years. If the initial connection supports only part of the final development, the first phase needs its own workable power and cooling arrangement. Installing every rack at the outset would leave some equipment waiting for the supporting infrastructure.

Separate IT demand from total facility demand

Server demand is only part of the electrical requirement. Cooling equipment, power-conversion losses and other facility services also consume energy.

For an illustrative energy budget, 10 MW of average IT demand at an average power usage effectiveness, or PUE, of 1.20 corresponds to 12 MW of average facility demand. This is an energy-planning example, not a rating for a utility connection or switchboard. Equipment sizing also requires peak-load calculations, operating margins and the intended maintenance and failure conditions.

The same distinction applies within the data hall. An acceptable average rack density can conceal a row whose local distribution capacity is insufficient. Map rack demand to the actual feeders, busways and cooling systems that serve it.

Understand where 800V DC fits

Higher-voltage direct-current distribution is attracting attention as AI racks become more power-intensive. At a given DC power level, increasing voltage reduces current. This can reduce conductor requirements, subject to the complete electrical design and protection requirements.

NVIDIA’s August 2026 road-map describes an 800V DC power rack arriving in the second half of 2026, with a row power center expected in 2027. The architecture is intended to simplify power conversion and support denser compute. These are vendor road map statements, so procurement plans should confirm actual availability and compatibility. NVIDIA 800V DC roadmap

For an operator, 800V DC is a distribution decision. It can improve how electricity reaches the racks, but it does not increase the capacity granted by the utility. Engineers must also consider DC protection, isolation, compatible IT equipment and maintenance procedures. NVIDIA architecture overview

Phase power and cooling together

Phased data center deployment can align investment with available power and expected IT demand. Each phase should include enough electrical and thermal infrastructure to operate independently within its design limits.

Future expansion also needs physical provision. Reserve space and routes for additional equipment, define connection points and establish how new systems will be tested beside operating capacity. These decisions influence whether the next phase can be added without disruptive rework.

EPG’s prefabricated IDC approach uses configurable modules and supports expansion as requirements change. This provides a relevant framework for planning capacity in stages. EPG IDC Solution

Evaluate cooling as part of the power budget

Electrical and mechanical teams need a shared load schedule. A rack power increase changes the heat that must be removed, while a cooling-system change affects facility electricity demand.

For compatible equipment, warmer liquid-cooling temperatures can reduce refrigeration demand under suitable site conditions. EPG’s cooling architecture includes secondary inlet temperatures around 45°C, heat recovery and supplementary cooling options. The annual benefit depends on hardware requirements, climate and the selected heat-rejection system. EPG Cooling Solution

Establish an operable first phase

Before committing to a capacity milestone, bring the key assumptions into one design review:

Confirm the electrical supply available on the proposed commissioning date.

Map rack loads to distribution paths and associated cooling capacity.

Check usable capacity during planned maintenance and the specified failure scenarios.

Identify equipment lead times and expansion interfaces that affect the programme.

Define tests for power transitions, controls and cooling response before IT handover.

The result should be a clear statement of what the facility can support on day one and what must change before the next expansion.

EPG’s AIDC Solution combines power, cooling and IT infrastructure through configurable reference architectures. Explore the available approaches to plan an AI facility around its power allocation, rack requirements and deployment schedule.