The International Energy Agency published its World Energy Outlook 2026 on November 19. In its Stated Policies Scenario, electricity consumption by data centres doubles to about 1,000 terawatt-hours a year by 2030. The outlook brings together a rapid buildout of AI computing and a much slower process of building the electricity infrastructure that supports it. Its most consequential finding for a technology project is the mismatch between a pipeline of large proposed sites and the connections local power systems can deliver. A global demand total gives a scale for planning. The date on which a site can use power remains a local question. IEA report · Overview and findings

For a company waiting on compute, the practical signal is how far a planned project has travelled. A land announcement, a building under construction and a hall drawing its intended load represent different stages. Each can appear in an account of investment in AI. Only some are producing the electricity demand expected from a completed site. The forecast becomes more useful when it can be connected to those stages and to the equipment, permits and agreements between them. That is also where a change in a few large projects can move a local outlook far more than a gradual improvement in the efficiency of one application.

Read the scenario before the headline#

The IEA’s Stated Policies Scenario, usually shortened to STEPS, reflects policies already in place and particular measures formally put forward. The 2026 outlook also includes a Current Policies Scenario, based on firmer existing rules, and other cases exploring different responses to energy pressures. The scenarios set conditions under which energy demand and supply develop. A figure from one of them needs its label when it is quoted. This matters for data centres because government choices about electricity infrastructure, supply and demand can affect the route from a computing proposal to a fully used building. IEA’s scenario descriptions

The comparison is annual consumption. A terawatt-hour counts energy used over time. A gigawatt is a rate of power, which can describe the size of a planned connection or the load measured at a particular moment. A one-gigawatt facility drawing that much continuously for a 365-day year would use 8.76 TWh. The same nominal facility with a lower average draw uses less. This arithmetic makes a headline about a large site easier to interpret. It also shows which additional information is missing when a project announces only its eventual capacity.

That distinction becomes substantial across a portfolio. Suppose one set of projects has an announced combined capacity of ten gigawatts. Multiplying by the hours in a year produces 87.6 TWh at continuous full load. If those sites come on line at different points in the year and their equipment fills gradually, actual first-year consumption could be far lower. A second portfolio of smaller sites that are already heavily used could consume more during that year. The capacity figures describe the possible rate. A forecast of annual demand must also make assumptions about timing and use.

The units determine how to interpret electricity’s relationship with AI. More computer work can raise demand. Faster hardware and better software can reduce energy needed for a particular task. Changes in how many tasks people ask the systems to do can outweigh either effect. A large centre also runs equipment beyond the chips executing the request. To compare sites, it is useful to keep the boundary of each reported measure clear, including which equipment is covered and over what period. An average for a whole facility tells a reader something different from a device rating or a figure for one inference job.

A useful comparison also keeps averages and peaks separate. Two facilities can consume the same annual electricity while placing different demands on a connection at their busiest hour. The network must be able to serve the actual load pattern with appropriate operating margins. An annual total smooths that pattern away. A planner who can see both the expected maximum and the likely sequence of normal use has more to work with. So does a company deciding whether a proposed supply arrangement suits the jobs it wants to run. This is why the measurement period is a material part of a power claim, even when all of the figures use familiar units.

CNBC’s November 19 coverage focused on the prospect of electricity use doubling and the constraints on building the sites. Its headline was carried in the day’s news roundups. Bloomberg’s coverage drew attention to the concentration of large AI projects. These are accounts of the same IEA release. Neither creates an independent set of power readings. The report and its methodology remain the place to examine the numerical assumptions, while the coverage helps show which parts of the forecast are becoming relevant to technology and energy decisions. CNBC report · Bloomberg report

The history already pointed toward a doubling#

The broad direction predates this release. In April 2025, the IEA’s Energy and AI report estimated that data centres used 415 TWh in 2024, about 1.5% of global electricity consumption. Its base case put 2030 demand at about 945 TWh. Those values were an estimate and a scenario from that report. The latest figure of roughly 1,000 TWh belongs to a later outlook and its STEPS assumptions. Comparing them shows a broadly similar scale for 2030. It does not by itself establish why the estimate changed or which region accounts for the difference. IEA’s 2025 baseline and projection

The IEA’s April 2026 update gives another dated point. It reported that data centre electricity demand grew 17% in 2025, while global electricity demand grew 3%. In that update, demand for 2030 remained close to the earlier trajectory, with a central projection around 950 TWh. The comparison of sector growth with total growth matters because the denominators differ. A small sector can grow much faster and still account for a limited share of a global total. The same increase can be dominant inside a small region where those facilities cluster. IEA’s April 2026 update · Updated energy and AI analysis

A reader following successive reports should preserve both the date and the definition of the series. An updated estimate of an earlier year can shift the apparent growth rate even if the future total barely changes. An expansion of what is counted as a data centre can do the same. A forecast whose line looks stable can also rely on a different mix of future equipment and utilisation. Those are good reasons to avoid treating a modest change in one distant total as evidence that nothing important happened during the intervening year.

This comparison becomes more useful at a nearer horizon. If a project expected to start operating next year is delayed, that can lower demand for the next year without cancelling the longer-term plan. The later years may then rise more steeply. If a company changes a design and permanently reduces its intended electrical capacity, that is another kind of revision. An account of the change should separate timing from scale. Project-level dates and operating milestones can help explain the distinction before it becomes visible in an annual global total.

A building is one part of a working AI site#

For its new outlook, the IEA describes using satellite imagery to follow 93 large sites purpose-built for AI. It inferred potential capacity associated with observed development of about 17 GW as of September 2026, ten times the level two years earlier. The report puts combined capacity of those sites at around 50 GW when completed. That is a record of observed physical development and an estimate of what it can support. It is different from reading meters at 93 fully occupied facilities. The difference is central to judging how quickly a construction boom will turn into actual power demand. IEA’s site-tracking analysis

Space-based observations are valuable when private construction progresses faster than comprehensive public statistics. A site can show new structures and equipment arrangements before a company discloses detailed operating data. Interpreting what those observations mean still requires assumptions about the facility. The useful evidence is stronger when physical changes are checked against permits, company disclosures and available connection information. A growing building footprint can then be discussed alongside the other steps that make the site usable. It helps a reader see where the measurement ends and the inference begins.

The next steps can occur on different schedules. A shell can be ready before electrical equipment. A power connection can be available before every rack is installed. A company can bring a first phase into use while later phases remain proposals. Treating the eventual campus as if it all began drawing power on the day the first building opened makes a demand series jump too soon. Treating the whole campus as unfinished until its last phase is complete hides use that already exists. A good timeline records each operating stage at the scale the available evidence can support.

For a technology company, the operational question is whether capacity will be ready when the workload needs it. A new training programme may be flexible about the month it starts. A service committed to customers in a particular region can have less room to wait. The same site can therefore be attractive to one buyer and difficult for another, even with identical headline compute and power capacity. Availability includes the timing, the location and the services the site can actually provide. A forecast that accounts for delays can still be consistent with a company experiencing a local shortage.

The global share can hide local concentration#

The IEA reports large prospective data centre capacity relative to today’s peak loads in several jurisdictions, including Texas, Northern Virginia, Finland and Denmark. It also says that all of the prospective capacity will not be built. A queue of proposals is evidence that many developers want access. It can contain different stages of commitment and more than one way for a company to meet the same need. The comparison with current peak load is useful for illustrating scale. It cannot be read as a prediction that local demand will suddenly increase by the full amount of every application. IEA’s regional comparison

The part that makes the concentration difficult is often geographic. Electricity can be abundant across a wider market while a specific substation or line lacks room for another large connection. An available plant elsewhere is helpful only if power can reach the place and time it is needed. A local connection therefore depends on more than the annual amount of generation in the country. Existing customers, maintenance and operating margins also belong in the decision. That is why comparing one new site with the output of a remote power source can oversimplify what the local network must do.

A planner has to decide how much proposed demand to build for. Building too late can leave ready projects waiting. Building far ahead of credible demand can leave customers paying for assets that remain lightly used. These choices become harder when individual sites are large compared with the local network. Milestones that show a developer’s commitment can improve the evidence, as can a clearer picture of how a first phase differs from an optional later expansion. The important question for a forecast is how those commitments move through time and how often they turn into operating load.

The IEA’s October report on integrating data centres into power systems gives policy context. It surveyed 34 jurisdictions covering 90% of global data centre capacity and examined the rules and commercial arrangements around integration. The report highlights the value of credible connection requests, flexibility and responsibility for the costs of new assets. These are ways to make a queue more informative and to allocate risk. They remain sensitive to local rules and the kinds of operation a customer can change. A proposed training campus and a service that must respond continuously can offer different flexibility. IEA’s integration report

Efficiency can change the load without removing the problem#

A more efficient system can do a given job with fewer resources. At scale, the total also depends on how much work is requested. Suppose an organisation reduces the energy for one task by a fifth and doubles the number of tasks. With all other conditions fixed, that use would require 1.6 times as much energy. This simple example does not predict how users will respond to a cheaper task. It shows why an improvement in efficiency per operation and growth in aggregate electricity can be true at the same time. Both belong in an account of the sector.

The change can be important within the same site. If new equipment completes work sooner, it can free capacity for other jobs. The facility may remain near the same electrical load while producing more useful output. Alternatively, the operator may reduce its draw or postpone expansion. Looking only at energy consumed would miss the first improvement, while looking only at work per watt would miss the consequences of growth. The useful measure depends on the question, whether the concern is service output, the need for a connection or the annual energy bill.

The timing of use matters for power systems as well. A batch workload that can shift to another hour may give an operator options that a live service lacks. Moving the task can help with a constrained period if the location and deadline allow it. The effect depends on how much of the facility’s load is genuinely flexible and what else has to keep running. It also has to be observable to the power system. Calling a site flexible is easier than showing that it reliably reduced the load by the agreed amount at the agreed time.

There is a similar distinction in power procurement. A company can sign a long-term agreement that helps finance new generation. The location and timing of that generation still determine its relationship with the site’s load. A facility operating around the clock also relies on the wider network and its balancing resources. A useful account can keep the commercial agreement, the physical connection and the pattern of operation visible together. That gives a reader a better way to evaluate how a supply plan will support the proposed demand.

The records that would strengthen the signal#

The first record to follow is the change in project status. A company can report its first energised phase, the amount of load in use and the schedule for later sections. The power system can report progress on the connection and the network work required. A difference between two dates can be informative even without access to all commercial details. It can show whether a revision came from construction, electrical equipment, permissions or the workload moving elsewhere. Each points to a different interpretation of the same future demand total.

The second is a consistent series of actual use. Comparing the same boundary over successive periods helps separate a ramp in operation from a change in reporting. A record that covers the whole facility should say so. A record limited to a server installation should identify that narrower scope. Weather, workload and the opening of another phase can affect year-to-year comparisons. Keeping those facts alongside the number makes a smaller dataset more useful than a larger list of unconnected announcements. It also makes it easier to check the next version of a forecast against what actually happened.

The third is the fate of proposals that leave the queue. Some disappear because a company chose another location. Some wait for years and then return with a different scale. Others are withdrawn after a project or contract no longer makes sense. Recording those outcomes helps improve how prospective demand is counted. It gives a practical counterweight to a series based only on new applications. The resulting picture may look less spectacular, but it says more about what infrastructure the region will need and when it will be used.

It can help to keep the original estimate beside the revised one. A forecast that said the first phase would start this year can later show whether that happened, whether the phase was smaller and which evidence supported the change. Without that record, a moving target can always look close to its latest projection. A dated trail makes the error useful. Teams can learn which milestones were dependable and which tended to move. Repeating that work across several projects can make the next estimate more accurate even if the public announcements remain optimistic.

For residents and ordinary customers, the relevant outcome includes the effect on the service and the costs they face. A large connection can fund infrastructure that has broader uses, or it can expose other customers to costs if the expected demand never arrives. The outcome depends on the contract and local regulation. A transparent public account can identify who pays for which assets and which conditions protect existing customers. Those details will differ by place. They explain why a global AI electricity projection can turn into a local debate about a particular substation and a bill.

What would change the outlook#

Some developments could lower the path to 2030. Equipment efficiency might improve faster than assumed, applications may require less compute, major projects may be cancelled or connections may arrive later. Others could raise it, including more intensive use of installed capacity or a faster adoption of workloads that users are willing to pay for. These are different channels. A change in one should be connected to the part of the model it affects. That makes a revision intelligible instead of treating every new forecast as a verdict on whether AI demand is rising or fading.

The new report also places data centres inside a wider electrical system. Households, industry, transport and other services are changing their demand. Network expansion and generation investments serve several of those uses at once. A data centre projection can help planners identify pressure, while competing projects and existing needs determine the response. The resulting decisions will shape where companies place computing and how quickly they can bring it into service. The relationship goes both ways, with electricity constraints changing the geography and pace of the computing buildout itself.

The 1,000 TWh figure is a useful scale for the IEA’s stated-policy world. The next evidence to watch is closer to the ground: operating phases, credible connection dates, measured use and the proportion of proposed capacity that survives. A doubling at the world level will arrive as a series of individual projects that clear or miss those steps. The signal becomes stronger when those projects can be connected to the annual total with a consistent account of what was built, what is powered and what is still planned.