AI infrastructure is often described in accelerator counts. The physical system that supports those accelerators includes servers, storage, networking, power delivery and cooling. The IEA describes data centers as a combination of IT equipment and the auxiliary infrastructure that keeps it operating. The facility and the machines therefore need to be considered together. [1]

Our interpretation is that an available building is not automatically available compute. An announcement can concern a future campus, a construction phase, an electrical connection, a commissioned hall, or installed servers. Those are useful milestones, but they should not be added together or presented as interchangeable supply.

For a buyer, the relevant milestone is the one needed for the workload’s start date. A long-term plan may justify a conversation about future capacity. It does not establish that a specific configuration can be delivered next week.

In its April 2026 update, the IEA reported that data-center electricity demand grew 17% in 2025. Its central projection rises from 485 TWh in 2025 to roughly 950 TWh in 2030. The 2030 value is a projection, not an observed outcome, and the IEA identifies bottlenecks in electricity connections, equipment supply and project development. [2]

Those global figures explain why energy belongs in the conversation. They do not tell us which provider has capacity in a particular city, which connection is commissioned, or what proportion of a site is available to new customers. Using a global forecast as evidence for a local inventory claim would skip the most important verification step.

Power also needs a defined basis. A campus-wide figure, a phase-specific figure and the usable power for IT equipment do not answer the same question. We would ask a facility operator to state the measurement boundary, operating status and relevant date before making a comparison.

The IEA’s 2026 analysis also highlights increasing AI power density and pressure on electrical equipment supply chains. That is a separate issue from annual energy consumption: a site must deliver power where the equipment is installed, not only have a plausible aggregate energy budget. [2]

Cooling readiness should be evidenced in the same way as electrical readiness. Berkeley Lab’s work on liquid-cooled commissioning identifies pressure, flow and load testing as parts of checking a system. A statement that cooling technology is planned is not equivalent to a completed commissioning result. [3]

The US Department of Energy defines power usage effectiveness as total facility energy divided by IT equipment energy over the same period. If a hypothetical site consumes 12 MWh in total while IT equipment consumes 10 MWh, its PUE is 1.2. That ratio describes facility overhead relative to IT energy. [4]

It does not, by itself, reveal how many useful model responses were produced, how well the GPUs were utilized, or whether the electricity had low emissions. Two sites can have the same ratio while doing different amounts of useful work. We would treat PUE as one clearly bounded metric, not a universal sustainability or performance score.

The purchasing implication is simple: keep the facility record, the compute configuration and the availability commitment distinct. Our preferred question is “What can this site support, in which phase, under which conditions, and from what date?” That is more informative than a single headline number.