Electricity 2026: Analysis and forecast to 2030 (data-centre demand and grid-queue sections)

Correction to the proposal card — read this first

The proposal for this source (DC-1, and DC1 in docs/context/kb-proposals.md) said it adds "data-centre electricity roughly doubling, 485 TWh (2025) → 950 TWh (2030), AI-focused share tripling". That figure is not in Electricity 2026. The full text was searched for 485, 950 and every sentence that mentions data centres; none gives a global data-centre TWh total.

The 485 → 950 TWh line comes from a different IEA report, Key Questions on Energy and AI (a World Energy Outlook special report, April 2026). That report's PDF was fetched to check the attribution, and it does say data-centre consumption goes "roughly doubling from 485 TWh in 2025 to 950 TWh in 2030", with AI-focused consumption "tripling". That report is not approved and is not ingested here, so the 485 → 950 figure is not entered on this rung from this source. It is recommended as a separate proposal. Everything below is what Electricity 2026 itself says.

What this is

The IEA's yearly report on world electricity, this year with a five-year forecast (2026–2030) instead of the usual three. It is a forecast by an intergovernmental agency, not a measurement. Its value to this rung is that it describes data centres the way the grid sees them: as a share of demand growth and as projects waiting in line for a connection.

Boundary — what is filed here and what is not

The line between energy and datacenters is not yet ratified (proposed 2026-09-05 in kb-scope.md). Under the proposed line, this report splits in two:

• Filed here (datacenters): data-centre demand, data-centre connection queues, how many queued projects are real, and how grid operators are rationing connections to data centres.
• Belongs on energy, and is NOT filed there by this ingest: the generation half — the renewables share rising from 17% to 27% by 2030, coal retirements, gas share, battery-storage costs, wholesale prices. No file under kb/energy/ was written. If the boundary is ratified the other way, this whole source moves to energy and nothing about the reading changes.

Key figures (all read from the report text)

Demand

• World electricity demand grew 3% in 2025 and is forecast to grow 3.6% a year over 2026–2030, with data centres named among the drivers alongside industry, electric vehicles and air conditioning.
• United States: demand rose 2.1% in 2025 and is forecast at nearly 2% a year through 2030, adding more than 420 TWh over five years. Data centres make up about half of that growth ("around half of the total increase", "about 50% of demand growth out to 2030", and in the regional chapter "almost half" and "more than half" — the report uses all four phrasings).
• Across the world, space cooling, data centres and heat pumps together make up almost half of the growth in the buildings sector to 2030; buildings contribute 49% of added global demand.
• China: demand in the ICT and digital-services subsector, which includes data centres and 5G, rose 17% year on year in 2025.
• India: data-centre demand grew by close to 50% in 2025; Maharashtra holds more than half of India's operating data-centre capacity.
• Ireland: data centres were about 22% of metered electricity consumption in 2024.

Connection queues — the capacity side

• More than 2,500 GW of projects — renewables, storage and large loads such as data centres — are stalled in grid connection queues worldwide.
• On the demand side, at least 150 GW of queued data-centre projects are in advanced stages, and one fifth of the global data-centre build-out is at risk of delay because of grid congestion.
• Queues are inflated by duplicate and "phantom" projects. From US utility data, only around 20% of data-centre connection requests materialise in the short to medium term. In Australia, an Oxford Economics study commissioned by AWS puts 8 GW of 44 GW of requests as likely to enter service. In Brazil, requests passed 26 GW by November 2025 (26.2 GW, up 32% from 19.8 GW in September), of which only 6 GW were under review or at late stages.
• Time mismatch: new grid infrastructure takes 5 to 15 years to plan, permit and build; a data centre takes 1 to 3 years. Prices for key grid components have nearly doubled in five years.

How grid operators are rationing access (the terms a data centre now faces)

• Dominion Energy Virginia proposed a rate class for large users with a 14-year commitment to pay 60% of generation costs and 85% of network costs even if they use less power than they applied for.
• ComEd (Illinois) charges a USD 1 million refundable deposit for data-centre requests of 50 MW or more.
• PG&E Electric Rule 30 cuts large-load connection time from 18–22 months to about 2–5 months for applicants that pay for transmission work upfront.
• Ireland (CRU, December 2025): new data centres must bring dispatchable generation or storage matching their maximum import, and source at least 80% of annual demand from renewables generated in Ireland. EirGrid has restricted data-centre connections around Dublin since 2021.
• UK AI Growth Zones give "strategically important" AI data centres of 100–500 MW priority grid access. SPP agreed a 90-day study pathway for large loads paired with generation.
• An "emerging trend" of data centres securing behind-the-meter supply through on-site generation, while usually staying grid-connected for reliability.

Why it matters here — the token-to-task path

A token's price has a floor set by what it costs to keep a megawatt of serving capacity running, and that megawatt cannot serve anything until it is connected. This report moves two numbers on that path:

1. How much of announced capacity is real. If only about 20% of US data-centre connection requests turn into load, a pipeline measured in requested gigawatts overstates future supply about five times over. That matters to anyone forecasting that serving capacity will outrun demand and push token prices down.
2. What connection now costs a builder, beyond the queue. The Dominion terms (pay 60% of generation and 85% of network costs for 14 years, used or not) turn a grid connection into a long fixed commitment. That is a cost sitting in the capex-per-megawatt denominator that the rung's ~$59M/MW figure does not include.

It also cross-checks the rung's existing queue source: Carbon Direct's 3–4-year waits and

300 GW PJM + ERCOT queue now sit beside an IEA view that the queue problem is global and that most of it is not real demand.

Limitations

• It is a forecast. The IEA revises it every year; cite it as "IEA, Electricity 2026 (early 2026)", never as a fixed fact.
• No global data-centre TWh figure in this report. The headline number people quote (485 → 950 TWh) is from Key Questions on Energy and AI, not this report (see the correction above).
• The 20% conversion rate is US-only and is the IEA's own reading of utility data; the Australia figure is a study paid for by AWS, a data-centre builder.
• Four phrasings of the US share ("around half", "about 50%", "almost half", "more than half") appear in different chapters. Quote it as "about half", not as a precise share.

Source: Electricity 2026 — Analysis and forecast to 2030, International Energy Agency. Read from the IEA's PDF at iea.blob.core.windows.net (report page blocks automated fetches).