Data centres and the power constraint

Data centres are becoming a materially larger electricity customer, but their growth does not translate into a uniform shortage across electricity systems. The planning problem is concentrated demand arriving at particular substations, with uncertain commissioning dates and equipment requirements. A global demand forecast cannot establish whether a specific site can obtain reliable power on schedule.

Three evidence sets frame the issue. The IEA estimates worldwide data-centre electricity consumption at 485 TWh in 2025. Ireland's administrative records put the sector at 23% of mains-network metered consumption that year. Berkeley Lab's older US study presents a wide scenario range for 2028. These answer different questions: global scale, local concentration and sensitivity to technology assumptions. ^{1 12 15 16}

Our assessment is that procurement and system planning should prioritise deliverable connections, phased load commitments and responsibility for unused infrastructure. Headline investment announcements and nominal generating capacity provide weaker evidence of when computing capacity will become operational.

The IEA's April 2026 assessment puts annual data-centre electricity use at 485 TWh in 2025 and about 950 TWh in its central projection for 2030. The latter represents around 3% of global electricity demand. The outlook is an energy-consumption projection, not a pipeline of approved projects. The chart deliberately shows only the two directly supported endpoints; it does not invent annual observations between them. ^{1 2 3}

All data-centre electricity use grew by an estimated 17% in 2025, while the AI-focused subset grew by 50%. These overlapping populations cannot be added together. Nor does the faster subset growth establish that every additional unit of sector electricity was consumed by AI. ^{4 5}

The investment implication is conditional. Greater computing demand can coexist with improving efficiency per task. A procurement case therefore needs explicit assumptions about workload volumes, model complexity, utilisation and facility overhead. Extrapolating either a single query's electricity use or a company's total capital expenditure into sector demand would conceal those assumptions.

Ireland provides a useful local-system case because its statistical office publishes annual meter-based consumption. In the July 2026 release, data-centre electricity rose from 3,030 GWh in 2020 to 7,663 GWh in 2025, with increases in every intervening year. The latest figure represented 23% of total metered consumption. This is a concentration measure for Ireland, not a global benchmark or a measure of AI's share. ^{6 11 12}

The chart uses the complete recent annual sequence from one release vintage. That choice matters: the 2023 and 2024 values in this release are 6,339 and 6,973 GWh. Mixing older releases with this edition could introduce revisions that look like economic changes. ^{9 10}

Metered consumption still has boundaries. CSO identifies relevant meters through customer activity, operator names and other checks; the administrative source does not carry a ready-made data-centre classification. Sites may enter the identified population as consumption grows. These records also do not establish the split between AI and conventional computing, or the quantity of electricity generated behind a site's meter. CSO background notes

Berkeley Lab's December 2024 study estimated US data-centre electricity use at about 76 TWh in 2018 and 176 TWh in 2023. Its 2028 scenarios span roughly 325–580 TWh. This edition predates the IEA's 2026 update and is retained as a clearly dated national sensitivity analysis, not presented as a newly issued September 2026 forecast. ^{13 14 15 16}

The study models equipment stocks and operating practices, including cooling, rather than summing a comprehensive census of electricity meters. Its forward range depends on assumptions about equipment shipments, how intensively accelerated servers operate and facility efficiency. The endpoints are scenarios, not statistical confidence limits. Berkeley Lab, executive summary

For planning, the value of this range is the question it forces: which commitments remain viable across different levels of realised demand? A utility's proposed reinforcement should be assessed against customer milestones and downside utilisation as well as an aggregate growth narrative. Subtracting this older US projection from the newer global projection would not produce a defensible rest-of-world forecast.

Annual consumption in TWh describes energy over time. Capacity in GW describes power at an instant. Converting a planned connection into annual electricity demand requires assumptions about utilisation and the timing of the ramp. The same distinction applies to batteries: their power rating does not reveal how many hours they can sustain a load.

The IEA projects approximately 20–25 GW of battery storage at data centres globally by 2030, and around 15–27 GW of onsite gas generation powering data centres, mostly in the United States. These ranges describe different technologies and functions. They should not be added and labelled firm data-centre supply. ^{17 18 19 20}

Its engineering analysis also finds that reliably supplying critical, variable loads with onsite gas can require generation infrastructure 30–70% above demand. This is a design requirement under the analysis, not an observed fleet-wide overbuild rate. ^{21 22}

A credible project assessment should therefore separate continuous energy supply, short-duration load smoothing, contingency reserves and grid access. A battery can address a rapid fluctuation without solving a prolonged energy deficit; redundant generating capacity can improve resilience while increasing capital and maintenance requirements.

The IEA's 17 September 2026 grid report adds a useful qualification to the debate: infrastructure expansion and better operation of existing networks can proceed together. It estimates that selected grid technologies could enable up to 330 GW of additional generation, storage and demand connections without reinforcement. This is system-wide technical potential, not spare capacity reserved for data centres. ²³

The technologies assessed include dynamic equipment ratings, changes to network configuration and control of power flows. Their relevance to a particular site depends on weather, network conditions and reliability constraints. Their potential therefore cannot be allocated to proposed projects simply in proportion to requested capacity. IEA grid assessment

Our operational inference is to evaluate an alternative connection package alongside the conventional expansion plan: staged energisation, an agreed curtailment envelope, network-control improvements and verified onsite flexibility. Each option should state the service it actually supplies and the circumstances in which it becomes unavailable. That creates a basis for comparing earlier access against cost and operational restrictions.

The evidence supports a milestone-based approach to demand rather than treating every announced project as inevitable load. Developers should demonstrate how power becomes available as computing equipment is installed. Utilities should connect reinforcement commitments to credible customer delivery schedules. Policymakers should make the allocation of upgrade and stranded-asset costs explicit.

For an individual project, request four linked disclosures:

• The connection agreement, energisation stages and conditions for firm or interruptible service.
• A load ramp tied to equipment installation and contracted workloads, with downside and delay cases.
• The complete supply design, distinguishing energy purchases, onsite generation, battery power and battery duration.
• The tariff and financial commitments that determine who pays if contracted demand fails to materialise.

These are analytical recommendations, not a claim that the cited agencies prescribe one standard contract. The central test is whether demand, infrastructure and financial commitments become binding on compatible timelines. A larger headline pipeline alone cannot answer that question.

Research closes on 22 September 2026. The global demand baseline and projections come from the IEA's April 2026 assessment; grid context uses its September 2026 publication. Irish figures are the July 2026 CSO release. US figures retain the December 2024 Berkeley Lab vintage. Publication date and data period are recorded separately for each source and statistic.

Historical IEA and Berkeley Lab model outputs are classified as estimates. Irish administrative observations are classified as actuals while retaining CSO's revision caveat. Future scenario endpoints are classified as forecasts; engineering requirements and technical potential are assessments, not observed deployment. No chart combines national meter data with global modeled data, and none converts GW to TWh without an operating assumption.

The CSO page contains an inconsistent year in one release header. We use 7 July 2026, corroborated by its release landing page and statistician's dated statement, while preserving the fetched page unchanged. Release landing page

The report does not estimate AI-only electricity levels, water use, emissions or project-level electricity prices. Those require different boundaries and additional evidence. IEA and CSO figures are adapted and attributed under CC BY 4.0; charts and interpretation are StatOrigin's, without agency endorsement. Author: StatOrigin editorial / AI-assisted research.