02.1 / Compute
The image begins with compute.
Accelerator availability and refresh cadence change rack density, heat loads, equipment choices and the economics of the facility around them.
Flagship Infrastructure Insight
AI has a physical footprint.
The next AI advantage will be measured in energised compute, not announced gigawatts.
Behind every model response is a physical chain of chips, racks, cooling, substations, power, water, fibre, land, equipment and delivery.
01 / Why now
The global percentage hides the local problem.
Data centres can remain a manageable share of global electricity while becoming a system-defining load in specific grids and markets.
415 TWhGlobal data-centre electricity use in 2024IEA source
~945 TWhIEA Base Case for 2030IEA source
23%Ireland metered electricity use by data centres in 2025CSO Ireland source
The 2030 value is a forecast, not an observed outcome. Ireland shows why local concentration can matter even when the global share remains limited.
02.2 / Power
Installed generation is not deliverable power.
Connection timing, redundancy, reinforcement and long-lead electrical equipment can control the operating date.
02.3 / Cooling
Electricity becomes heat.
The cooling architecture must follow the hardware roadmap without becoming obsolete first.
02.4 / Water
Lower direct use can move the demand elsewhere.
Resource claims change when electricity generation, treatment, desalination and heat rejection enter the boundary.
02.5 / Land
A site is valuable only if it preserves optionality.
Power corridors, cooling, fibre, security, logistics and future phases compete for space.
02.6 / Fibre
Digital resilience still depends on physical routes.
Diversity, latency and international links determine what happens when one path is lost.
02.7 / Materials
One specialist package can control the schedule.
Transformers, switchgear, pumps and other long-lead equipment shape the critical path.
02.8 / Delivery
The image is complete only when the system operates.
Infrastructure, equipment, approvals, contractors and commissioning must finish together. Useful compute exists only when every layer converges.
03 / Capacity Ladder
Announced is not commissioned.
Announced
Public ambition can leave site, timing, power, procurement and phasing unresolved.
What proves this stage?
Official announcement or developer statement.
What can still fail?
Site, timing, power, procurement and phasing can remain unresolved.
Power allocated
A credible supply commitment is not yet a physical connection.
What proves this stage?
Executed utility agreement, credible allocation or equivalent supply commitment.
What can still fail?
Physical connection, reinforcement and energisation timing can still change.
Connected
Connection can be energised while data halls and IT remain incomplete.
What proves this stage?
Utility or commissioning evidence that the required connection is physically available and energised.
What can still fail?
Data halls, cooling and IT systems may still be incomplete.
Commissioned
Electrical and cooling systems support installed IT load.
What proves this stage?
Commissioning or operating evidence that electrical and cooling systems support installed IT load.
What can still fail?
Utilisation can remain low and workload deployment may lag.
Utilised compute
Productive workloads are actually running. The conversion rate between stages is more informative than headline pipeline size.
What proves this stage?
Operational evidence that commissioned infrastructure is running productive workloads.
What can still fail?
Actual utilisation can be commercially confidential. Where public telemetry is unavailable, the stage should remain explicitly unverified rather than estimated.
04 / Density
AI changes the unit of design.
~7.5 to 9 kW
Modal-density context from the Uptime Institute 2025 survey.
Open Uptime Institute source50 to 120 kW
Specialised leading-edge AI rack range documented by ASHRAE. This is a different engineering problem, not a new fleet-wide average.
Open ASHRAE sourceDensity propagates.
Power delivery, cooling, heat rejection, equipment space and upgrade risk change together.
05 / System boundaries
The footprint changes when the boundary changes.
Site
Facility electricity, direct cooling water and local heat rejection.
Utility system
Generation mix, transmission, desalination and indirect water.
Lifecycle
Construction, hardware refresh, materials and embodied impact. A “zero-water” or “green” claim is incomplete unless the accounting boundary is explicit.
Open LBNL source06 / Gulf conversion
The opportunity is not the announcement.
5 GW
Wider publicly stated campus ambition.
Open official project statement1 GW
More clearly defined compute cluster.
Open official project statement200 MW
First phase publicly expected in 2026.
Open official project statementExpected is not verified.
Connected, commissioned and utilised stages require evidence as they are reached.
Conversion is the advantage.
Cost, capital, policy, market access and infrastructure quality matter when they reliably become commissioned and productively used compute.
Illustrative sensitivity retained from the paper: a hypothetical 1 GW IT load at 90% utilisation and PUE 1.10 implies about 8.67 TWh annual facility electricity. It is not an official project forecast.
07 / Counter-signals
The footprint can improve. It does not disappear.
Efficiency can improve compute per task. Some workloads can shift in time, place or device. Cleaner generation can lower emissions intensity. Their effects remain scenario-dependent and can reduce or redistribute resource intensity rather than remove the physical system.
08 / Executive decision screen
Before counting gigawatts, ask what can actually be energised.
01What Capacity Ladder stage is being reported?
02What is the earliest evidence-based energisation date?
03Which physical dependency controls the schedule?
04What is the summer cooling and water strategy?
05What remains useful through several hardware cycles?
06What local capability remains after commissioning?
The imagery is illustrative, location-neutral and AI-generated. It explains system components and does not represent a named project.
Sources & definitions
International Energy Agency, Energy and AI
Supports the 415 TWh observed estimate for 2024 and the approximately 945 TWh Base Case forecast for 2030.
Open sourceCentral Statistics Office Ireland
Supports the 23% share of total metered electricity consumption used by data centres in Ireland in 2025.
Open sourceUptime Institute Global Data Center Survey 2025
Supports the single-digit modal-density context. The range is context, not a universal facility average.
Open sourceASHRAE AI Data Center Energy Performance Framework
Supports the specialised leading-edge density context. The 50 to 120 kW range is not presented as a fleet-wide average.
Open sourceLawrence Berkeley National Laboratory
Supports separating direct site water from indirect system water and making the accounting boundary explicit.
Open sourceG42, Stargate UAE
Supports the publicly stated 5 GW campus ambition, 1 GW compute cluster and 200 MW first phase expected in 2026. These statements do not independently verify commissioned or utilised capacity.
Open sourceChaar AR calculation
1 GW × 90% utilisation × 8,760 hours × PUE 1.10 ≈ 8.67 TWh per year. This is a transparent engineering sensitivity, not an official project forecast.