26MAY

IEA: 17% growth, $700B capex run-rate

3 min read

11:34UTC

The IEA's 16 April release put global data centre electricity demand growth at 17% in 2025, six times the overall electricity growth rate, with five hyperscalers' capex topping $400 billion and projected to reach roughly $700 billion in 2026.

← Back to Grid wins power to switch off data centres Jump to analysis ↓

IndustryDeveloping

Key takeaway

Capex topped $400B in 2025 and is forecast at $700B in 2026; the binding constraint is no longer capital.

The International Energy Agency (IEA) released its updated data centre electricity report on 16 April 2026, restating the growth curve in maximalist terms. ¹ Global data centre electricity demand grew 17% in 2025, six times the 3% overall global electricity growth rate, with AI-focused facilities expanding by roughly 50%. Five major tech companies collectively topped $400 billion in capex in 2025, a figure the IEA expects to rise a further 75% in 2026 to about $700 billion. ²

The IEA is the OECD's energy modelling agency in Paris, the body whose annual World Energy Outlook sets the baseline most utilities and finance ministries use for grid planning. When it puts a number on hyperscaler capex, the figure is sourced directly from company filings and disclosed corporate energy data, not industry analyst projections. The 2025 total absorbed roughly 485 TWh of global electricity at data centres alone.

Conditional agreements for small modular reactor (SMR) power for data centres jumped from 25 GW at end-2024 to 45 GW by April, a near-doubling of paper nuclear in sixteen months. ³ No commercial SMR has yet powered a data centre anywhere in the world; the GW figures are letters of intent, not interconnection agreements. The pipeline reflects what hyperscalers want to procure, not what the supply chain can deliver this decade.

Capex and chips have stopped binding the build-out. Credit lines remain open at unprecedented scale and the silicon supply has caught up enough to absorb the dollars. What stops a project in April 2026 is grid headroom, planning approval, water permits, and increasingly state legislatures, not capital. The five sections that follow trace what happens when a $700 billion spend curve meets jurisdictions that are starting to refuse the load.

Deep Analysis

In plain English

A data centre is a large building full of computers that run cloud storage, streaming, and AI services. Every computation uses electricity and generates heat. As AI models have grown larger and more complex, the computers running them need far more power per square metre of floor space than earlier generations. The IEA is the International Energy Agency, a Paris-based international body that tracks global energy use. Its 16 April figures show that data centres worldwide used 485 TWh of electricity in 2025, about the same as the entire United Kingdom. That is up 17% from the year before, while total global electricity use grew only 3%. SMRs, or small modular reactors, are a type of nuclear power station being developed at smaller scale than traditional plants. Tech companies are signing agreements to buy electricity from SMRs that have not been built yet, because they do not expect the ordinary electricity grid to have enough spare capacity to meet their needs.

Deep Analysis

Root Causes

Data centre electricity demand is growing at a multiple of overall grid growth because AI inference workloads are power-dense in ways general-purpose computing is not. A 2027-era AI server rack draws power equivalent to roughly 65 UK households; a 2020-era rack drew the equivalent of about six. This is not a scheduling or efficiency problem, it is a physics constraint on the rate at which waste heat can be removed and power delivered.

The SMR pipeline jump from 25 to 45 GW in 16 months reflects a secondary cause: grid operators in the UK, US, and EU cannot connect new load at the pace operators require, so operators are contracting alternative generation sources. SMR agreements are conditional and multi-decade; they signal intent, not delivered capacity.

What could happen next?

Consequence
At the IEA's central 2030 projection, data centres would consume approximately 950 TWh globally, equivalent to adding Japan's entire electricity demand to the sector in five years.
Medium term · Medium
Risk
SMR conditional agreements at 45 GW represent procurement intent, not delivered capacity; if nuclear licensing timelines hold to historical rates, most will not commission before 2035.
Long term · High
Opportunity
Countries with surplus low-carbon generation capacity and permissive planning frameworks, Iceland, Norway, parts of Canada, stand to attract disproportionate hyperscaler investment as consent constraints tighten elsewhere.
Medium term · Medium

Source Landscape

This story draws on neutral-leaning sources

Primary parallel: The 1990s US internet infrastructure buildout saw co-location and backbone capacity grow at comparable compounding rates, culminating in the 2000-2001 overbuilding that stranded 90%+ of laid fibre. The structural parallel: capital committing to anticipated demand before the demand materialises, with the resulting ghost infrastructure sold at distressed prices. The IEA's own report notes that only a fraction of contracted grid capacity will build, echoing that pattern.

Counter-parallel: The 1973-1979 US nuclear construction boom is a counter-case: capital and government commitment were both present, but regulatory friction, community opposition, and cost overruns reduced 230 planned units to 53 completed. The SMR pipeline jumping to 45 GW in conditional agreements is not capacity; it is the starting point of a siting, licensing, and financing process that the nuclear parallel suggests will face severe attrition.

The $400 billion 2025 capex figure spans Microsoft, Amazon, Google, Meta, and Oracle. Of this, the IEA estimates the majority is going into AI-specific infrastructure rather than general cloud capacity.

The structural implication is a concentration of demand in a sector that currently accounts for roughly 2% of global electricity but is on track for 3% by 2030. For power grid operators, the planning horizon for a new grid connection (3-5 years in most OECD jurisdictions) is mismatched with the 12-18 month deployment cycles hyperscalers now target. That mismatch is the origin of behind-the-meter gas, SMR conditional agreements, and the UK queue anomaly.

Consensus view: Lawrence Berkeley National Laboratory's Arman Shehabi, co-author of the 2024 US DC electricity report, and the IEA's Tim Gould assess that 2025 consumption of 485 TWh represents a structural step-change rather than a cyclical peak: AI server rack power density rose 11-fold between 2020 and 2025, with another fourfold increase projected by 2027, compressing a decade of demand growth into roughly seven years.

Counter-view: China's State Grid Corporation has noted that Chinese hyperscale deployments have achieved PUE (power usage effectiveness) ratios below 1.1 for new facilities, against a global average above 1.5, and argues that efficiency gains could significantly flatten the demand trajectory. Beijing's 2025 guidelines mandating PUE below 1.3 for new DCs suggest the IEA's high-case projection assumes continued US/European inefficiency as a constant.

Key tension: Whether the 17% 2025 growth rate represents a one-off AI-buildout surge that will moderate, or the start of a compound trajectory toward the IEA's 2030 projection of 950 TWh, roughly doubling in five years.

Sources:International Energy Agency·International Energy Agency

Mentions:Microsoft →AWS →Google →Meta →United Kingdom →OECD →attrition →Prima →China →Lawrence Berkeley National Lab →SMR →hyperscale →International Energy Agency →

First Reported In

Update #1 · Boom hits wall: grid says no, states freeze

International Energy Agency· 26 Apr 2026

Read original →

Causes and effects

Caused by

Stargate US: $500B headline, 1.2 GW operational

The Stargate US headline-to-ground gap — only 1.2 GW operational of a 10 GW nominal programme and 30-50% 2026 delay forecast — is a consequence of the consent and grid constraints documented by the IEA's April capex projection.

Occurred 1 Apr 2026

Read story →

This Event

IEA: 17% growth, $700B capex run-rate

The headline trajectory for AI infrastructure has not bent; what has changed is what now constrains it.

Led to

Hyperscalers raise 2026 capex to $725 billion

IEA's April 2026 projection of $700bn sector spend (ID:2784) was confirmed as a floor by Q1 hyperscaler earnings coming in at $725bn combined.

Occurred 30 Apr 2026

Read story →

Different Perspectives

Data-centre developers and hyperscale operators

Hill County's rescission on 4 June, seven days after RCM Hill filed a $100 million taking suit, shows the Fifth Amendment threat is faster than any appellate route: the damages clock runs from the day the moratorium passed. Rural counties with no large-load permitting framework face a litigation bill that can exceed their entire annual budget.

Kenya and President Ruto

Kenya's suspension of the $1 billion Microsoft-G42 Olkaria project in early May applies raw-capacity logic at national scale: President Ruto stated the full 1 GW build would mean switching off half the country against a 3 GW installed base. A single hyperscale campus can consume a third of a Sub-Saharan grid with no equivalent constraint in Europe.

Denmark and Energinet

Energinet's 27 May extension of its large-load connection pause, with a 60 GW queue against 7 GW peak demand, demolishes the assumption that surplus renewable generation is a relief valve for compute demand. Denmark has more wind than it can use and still cannot connect data centres, because transmission pace is the binding constraint.

France and EDF

EDF's repurposing of the Bouchain former power-station site for SoftBank's Phase 1 campus gives France a replicable siting instrument, a brownfield nuclear connection bypass, that no other G7 grid operator can match. France's Choose France summit on 30 May secured the boom's largest European bet without a connection-queue fight or community moratorium.

SoftBank Group

SoftBank's EUR 75 billion France commitment on 30 May anchors at EDF's Bouchain nuclear baseload, bypassing the UK's four-times-US electricity cost premium (cited by OpenAI as reason to pause Cobalt Park) and Germany's grid-queue delays. EDF's supply relationship is bilateral; SoftBank never enters the French connection queue.

US residential ratepayers and state regulators

Portland General Electric's 4 June tariff is the first evidence that PJM's cost-transfer warning to governors on 19 May can run in reverse: Oregon households get a 1.3% bill reduction as data centres absorb their grid costs. The 12 other states carrying active cost-attribution bills now have a filed tariff with actual numbers to cite.