Ask a data centre operator whether they are ready for AI and you usually get an answer about hardware. Ask them when they can energise their next hall and the conversation changes tone completely. That gap is the whole story of 2026. The International Energy Agency, in Key Questions on Energy and AI published on 16 April 2026, puts global data centre electricity consumption at 485 TWh in 2025, a 17% rise in a single year, heading for roughly 950 TWh in 2030. Readiness is no longer something you buy. It is something you queue for.
Key takeaways
- IEA: 485 TWh in 2025 to about 950 TWh in 2030, near 3% of global electricity.
- AI-focused facilities grew 50% in 2025 and are projected to triple by 2030.
- Gas turbine orders surged 70% in 2025, a visible supply chain chokepoint.
- EU sites above 500 kW of IT demand report KPIs annually by 15 May.
What the demand curve actually says
The headline number is large but the shape underneath it matters more. The IEA separates general-purpose data centres from AI-focused ones, and only the second category is behaving unusually. Overall data centre electricity demand grew 17% in 2025. Consumption by AI-focused facilities grew 50% over the same period, and the agency expects it to triple between 2025 and 2030, approaching the level of conventional data centre consumption.
By 2030, on that trajectory, data centres would account for around 3% of global electricity demand. It is worth holding that figure loosely. The IEA itself frames AI energy demand as the product of three moving and uncertain trends: efficiency improvements, uptake, and changing model capabilities. A projection built on three variables that each swing by double digits annually is a planning aid, not a forecast to sign contracts against.
Why the constraint moved off the raised floor
The bottleneck is upstream, and it has hardened rather than eased over the past year. The IEA reports that chokepoints across energy supply chains and advanced chip manufacturing have tightened since its previous assessment, and names three that operators feel directly.
- Generation and grid equipment. A 70% surge in gas turbine orders during 2025 exposed how thin that supply chain is, and the same pressure lands on power electronics and transformers, which are the unglamorous items that decide whether a substation exists on time.
- Connection queues and permitting. Planning and regulatory systems are being stretched by the sheer volume of data centre applications. The IEA’s own policy recommendations call for reforming how connection queues are managed and for streamlining permitting, which is a polite way of saying the process is now a material project risk.
- High-bandwidth memory. A shortage of HBM, integral to AI chip production, has developed and is expected to persist through at least the end of 2027.
None of those are problems a facilities team can engineer around. They change the order of operations for a build: secure the power position first, then design the hall around what the interconnection agreement actually allows.
Inside the hall, density is the design decision
The technical divide between an AI hall and a conventional one comes down to how many kilowatts sit in a single rack and where the heat goes. This is the part operators can control, and it is where retrofits either work or quietly fail.
An NVIDIA GB300 NVL72 rack draws in the region of 132 kW to 142 kW depending on configuration and vendor integration, with roughly 90% of the heat captured by liquid and the remainder by air. Set that against the installed base described in Uptime Institute survey work, where average rack densities have historically sat well under 10 kW and few facilities exceed 30 kW. These are not two points on the same curve. They are different buildings.
| Design parameter | Conventional enterprise hall | AI training hall |
|---|---|---|
| Typical rack power | Single-digit to low tens of kW | Above 100 kW per rack |
| Primary heat path | Air, with containment | Direct liquid to chip, air for the remainder |
| Load profile | Relatively stable | Sharp swings during training runs |
| Binding constraint | Floor space and cooling capacity | Grid connection and water or heat rejection |
The practical consequence is that partial readiness is a real and defensible position. A site can host inference workloads at moderate density on an air-cooled floor and be genuinely useful, without ever being suitable for training clusters. Pretending otherwise leads to the worst outcome, a hall rebuilt for a density the incoming power will never support.
Efficiency is improving, just not where your meter reads it
Two efficiency stories run in parallel and they are easy to confuse.
At the workload level, gains have been dramatic. The IEA notes energy consumption per individual task dropping by at least an order of magnitude annually in recent years. At the facility level, movement has been far more modest: Uptime Institute survey data has shown average PUE broadly flat for years, held back by legacy infrastructure and by climate-specific limits on efficient cooling, with a worldwide average around 1.56 in its 2024 survey.
That divergence explains why aggregate consumption keeps rising while every individual component gets more efficient. Efficiency per task is being converted into more tasks, not into a smaller electricity bill. Any business case that assumes otherwise is assuming demand stays still, which nothing in the current data suggests it will, and which sits awkwardly alongside the tech trends currently disrupting the IT sector.
The reporting obligation European operators keep underestimating
If your facility sits in the European Union, readiness has a compliance dimension that has nothing to do with GPUs. Commission Delegated Regulation (EU) 2024/1364, adopted on 14 March 2024 under the Energy Efficiency Directive, requires operators of data centres with an installed IT power demand of at least 500 kW to report to a European database.
Reports are due annually by 15 May for the previous calendar year, and the indicators are the ones a poorly documented site cannot produce on short notice: PUE for energy, WUE for water, ERF for reused energy and REF for renewable supply, alongside data on temperature set points and waste heat use. Water is the one that catches operators out, because liquid cooling choices made for density reasons show up later as a WUE figure someone has to defend.
A readiness test worth running before the next board paper
Rather than a maturity model, we would ask five questions and treat any vague answer as a no.
- What is your firm energisation date, in writing, from the utility, for the next increment of load?
- What is the lead time on your transformers and switchgear as quoted this quarter, not as remembered from the last build?
- Which halls can physically accept liquid, meaning floor loading, pipework routes and a heat rejection path, without structural work?
- Can you produce PUE, WUE, ERF and REF for last calendar year from instrumentation rather than estimation?
- What happens to your power draw when a training job stops abruptly, and has anyone modelled that swing with the utility?
An operator who can answer all five has a plan. An operator who can answer the first two has a schedule, which is more than most. The rest is engineering work with a known cost, and engineering work has never been the reason an AI buildout slipped.
Wondering how the energy maths plays out beyond data centres?
The same tension between decarbonisation targets and capital discipline is reshaping heavy industry.
Read our take on industry going carbon-neutral without going broke
Sources: International Energy Agency, Key Questions on Energy and AI, published 16 April 2026, for data centre electricity consumption of 485 TWh in 2025 rising to roughly 950 TWh in 2030 and around 3% of global electricity demand, growth of 17% in 2025 overall and 50% for AI-focused data centres with a tripling expected to 2030, the 70% surge in gas turbine orders in 2025, pressure on power electronics and transformers, the high-bandwidth memory shortage anticipated to persist through at least the end of 2027, the stretching of planning and permitting systems and the recommendation to reform connection queue management, energy consumption per task falling by at least an order of magnitude annually, and the framing of AI energy demand as the product of efficiency, uptake and model capability. Uptime Institute Global Data Center Survey for average worldwide PUE of about 1.56 in its 2024 edition, the flat multi-year PUE trend attributed to legacy infrastructure and climate-specific cooling limits, and average rack densities with few facilities exceeding 30 kW. NVIDIA GB300 NVL72 vendor documentation from HPE, Lenovo and Schneider Electric for rack power in the 132 kW to 142 kW range depending on configuration and the approximate 90% liquid, 10% air heat split. Commission Delegated Regulation (EU) 2024/1364 of 14 March 2024, supplementing Directive (EU) 2023/1791 on energy efficiency, for the 500 kW installed IT power demand threshold, the annual reporting deadline of 15 May and the PUE, WUE, ERF and REF indicators. Updated August 2026.
