An AI service can feel instantaneous while depending on decisions made years earlier. The response on a screen sits at the end of a chain that includes chips, cooling, electrical equipment and a connection to a particular power system. Improving one link does not automatically accelerate the rest.
That mismatch is a useful way to read the current infrastructure debate. A chip order follows a technology and manufacturing schedule. A grid connection follows an engineering, planning and system-capacity schedule. Once the facility operates, its demand follows a sequence of hours and seconds. Each clock creates a different problem, and each needs a different piece of evidence.
The International Energy Agency’s April 2026 update on energy and AI keeps this physical context in view. Its central outlook projects global data-center electricity consumption increasing from 485 TWh in 2025 to about 950 TWh in 2030. That is a projection for data centers as a whole, not measured electricity use by AI alone in 2030. The distinction matters before any company tries to translate the global number into a local project.
Clock one / The investment arrives before the operating record
An announced facility begins as a set of plans and commitments. The promised computing capacity, electrical connection and eventual service may arrive at different times. Some projects change scope or do not proceed. Counting every announcement as an operating load can overstate near-term demand; ignoring credible commitments can leave a power system unprepared.
The relevant record should distinguish stages. A proposed project, a contracted supply of equipment, a facility under construction and an energized operating site are different observations. They can all inform an outlook if labeled properly. Combining them into one installed-capacity number would conceal uncertainty about what exists and when it can be used.
Accelerator efficiency is similarly specific. A chip can perform more computation per unit of energy under a tested workload while the facility’s total electricity consumption rises. More equipment, more use or a shift toward more demanding tasks can outweigh the efficiency improvement. These outcomes are compatible; one does not automatically disprove the other.
The service also needs a definition of useful output. Tokens are a convenient computing measure, but a short answer and a long sequence of unsuccessful tool attempts do not create identical value per token. A system that completes the requested work with fewer attempts can improve its effective energy performance without changing the hardware at all. That improvement should be measured against comparable outcomes rather than assumed from response length.
Procurement commitments can outlast an early estimate of demand. A facility designed around one workload may later serve another, with different utilization and cooling requirements. Planning needs room for that uncertainty without treating every future application as a reason to build the maximum conceivable project immediately. The business case and the engineering case should state which assumptions they share.
Clock two / The connection belongs to a place
A global electricity total cannot establish whether a particular substation can accommodate a particular new load. The location, connection arrangement and timing of demand all matter. A region may have ample annual generation while a local network still needs reinforcement. Conversely, a facility connected where capacity is available can face a different constraint from an otherwise similar project elsewhere.
Power and energy are easy to blur in this discussion. Power describes a rate, commonly expressed in megawatts. Energy describes the accumulated use over time, commonly expressed in megawatt-hours or terawatt-hours. A requested connection capacity is not a statement that the site will draw that power continuously throughout the year.
Take an explicitly illustrative flat load of 100 MW. Operating for twenty-four hours would consume 2,400 MWh, or 2.4 GWh. Operating at that same rate for every hour of a 365-day year would consume 876 GWh. Real utilization, outages and load variation would change the result. The arithmetic is a unit conversion, not a forecast for an actual AI facility.
It also needs a boundary. Is the 100 MW the IT equipment’s draw, the whole facility’s draw or the maximum connection rating? Those quantities need not be equal. A presentation can be numerically accurate and still mislead if it switches among them without saying so. The boundary should appear next to the number, not only in a distant footnote.
The US Department of Energy’s resources on meeting data-center demand identify a portfolio of supply, flexibility and grid-modernization approaches. That is a reason to examine several responses to growth rather than assume there is one universal power source or construction sequence. The appropriate combination depends on the system and the service being supplied.
Reliability arrangements require the same specificity. A source that produces enough energy over a year may not be available at every required hour. Storage, dispatchable supply, network access and demand management can contribute in different ways. Naming one component does not establish that the entire arrangement meets the facility’s operating commitments.
Clock three / The load is a schedule
Once the building is connected, the hourly pattern becomes consequential. A task with a deadline several hours away offers a different scheduling opportunity from a request awaiting an immediate response. Some work may be shifted in time or location; other work may be constrained by latency, data access, software dependencies or contractual requirements. Flexibility is a property of a particular workload and arrangement, not a label that applies equally to all AI computation.
The IEA’s 2026 executive summary describes improving efficiency per task alongside growth in use and more demanding applications. This is why one energy estimate for a simple text query should not be applied to video generation or a long agent workflow. The task, system configuration and measurement boundary must remain attached to the estimate.
A useful workload schedule asks what can be delayed, for how long and at what cost to the requested outcome. A batch task might tolerate an overnight window but depend on a large dataset available only at one site. Moving it could introduce transfer time and other constraints. An interactive service might have little room to delay a response while retaining room to route some requests differently.
Testing flexibility should therefore include the consequence of exercising it. Did the job still finish before its deadline? Was the promised response time maintained? Did shifting one workload create another bottleneck? A claim about reducing peak demand is more informative when it identifies the service maintained during that reduction, instead of treating every curtailed watt as an equivalent success.
The commercial arrangement also matters. A facility can have technically shiftable work without an incentive or an agreed process to shift it when the grid needs help. Conversely, a flexibility commitment that exceeds practical capability can create an unreliable promise. Operating policy, measurement and the relevant agreement have to refer to the same behavior.
The efficiency ratio is not the whole footprint
Cooling and other supporting systems add to IT electricity use. The Lawrence Berkeley National Laboratory’s 2024 US data-center energy report examines facility energy and water use, including measures of supporting overhead. Its national estimates and modeled outlook should not be presented as measurements of a specific new building. They provide context and definitions, not a substitute for site evidence.
For another purely illustrative example, suppose IT equipment uses 100 MWh and the facility uses 120 MWh in total over the same interval. The resulting power usage effectiveness ratio is 1.2. Reducing the total to 110 MWh while preserving the same IT use improves that ratio to 1.1. It does not tell us whether the IT work itself was useful or efficient.
A facility could improve the overhead ratio while its total use grows because it adds more IT work. That is not an arithmetic contradiction. Both the ratio and the absolute consumption should be visible if the question concerns environmental impact or network demand. Selecting only the improving metric can conceal the scale of the change the community experiences.
Water introduces another boundary. Direct site consumption, water associated with electricity production and the local significance of that use are different considerations. A single global average cannot fully describe conditions at a water-stressed location. Cooling choices need an account of the actual system and context, rather than a universal conversion from every AI response into a fixed quantity of water.
The time boundary matters for environmental claims too. Matching annual electricity use with an annual quantity of contracted generation does not describe the physical supply in every hour. Hourly evidence and the details of the procurement arrangement answer a different question. A claim should say which accounting method it uses without allowing a broad adjective to imply results the method does not establish.
Put the three clocks on one project sheet
A practical project review can identify the intended computing service, the timing and certainty of equipment delivery, the connection stage and the expected operating pattern. It should also name the assumptions that could change those dates or loads. This is more useful than treating installed chips, announced megawatts and projected annual energy as three interchangeable measures of progress.
The same sheet should distinguish the counterfactual. Would the work otherwise run at another facility, be completed with a different model or not happen at all? Those alternatives have different implications for demand. A service moved between sites can create a local increase without an equivalent global increase; entirely new work changes the comparison again.
Efficiency claims deserve a comparable baseline as well. If a new model answers a harder question, comparing its energy use with a previous model’s answer to an easier question mixes capability and efficiency. If a smaller model needs several failed attempts before a larger one completes the job, measuring only the smaller model’s first attempt understates the service’s use. A meaningful comparison holds the required outcome and acceptable quality in view, then counts the work needed to obtain them. This can be more difficult than quoting energy per request, because requests vary and success requires evaluation. The difficulty is a reason to describe the uncertainty, rather than to assign every interaction a universal environmental price.
The contract can put a price on the gap between reserved capacity and actual use. AEP Ohio’s explanation of its data-center tariff, approved in July 2025, describes an initial term consisting of a load-ramp period of up to four years plus eight years. Its minimum monthly billing demand uses the greater of a recent-demand test and a contract-capacity schedule. For contracted capacity above 75 MW, the latter starts at 57.5 MW and adds the capacity above 75 MW, subject to an 85% cap.
For an illustrative 100 MW contract, that branch gives 57.5 + 25 = 82.5 MW. This is a billing-demand calculation under the published schedule, before considering the other test and applicable contract provisions. It does not mean the customer physically consumes 82.5 MW continuously, or buys that number of megawatt-hours. The distinction between capacity and energy is now an operating expense rather than a vocabulary exercise.
The arrangement assigns some risk of an underused connection to the customer reserving it. That can support investment planning; it also makes a developer pay for capacity before every server is busy. Whether the balance is appropriate is a policy and commercial judgment. The important analytical point is that a project has a financial load schedule as well as a physical one. A chip-delivery delay can affect both.
Put a chip replacement schedule, a connection timetable and an hourly load profile beside one another. The gaps between them reveal which part of an AI project is waiting and who has to act. New computing equipment may arrive while a substation upgrade is still being planned; a connected facility may still need to shift work around a constrained evening. Those are distinct operating choices. Calling all of them the power problem makes it harder to fund the one that is actually holding a service back.
Sources & context
This article draws on the sources linked in the text. Analysis and illustrative scenarios are identified separately from reported findings.
- April 2026 update on energy and AI
- resources on meeting data-center demand
- 2026 executive summary
- Lawrence Berkeley National Laboratory's 2024 US data-center energy report
- AEP Ohio's explanation of its data-center tariff
