Featured answer: AI compute strategy must account for power availability, connection timing, location, hourly carbon intensity, cooling, water, resilience and workload flexibility. Global electricity share alone can hide local constraints because data centres concentrate demand. Scenario-based planning is more credible than one forecast of model growth or hardware efficiency.
This evidence-led analysis by Jonas Adam Mohamed Osman Abdelghafour examines what changed, where uncertainty remains and how leaders can turn the issue into practical decisions without overstating what current technology can do.
Key takeaways
- The IEA projects data-centre electricity use could reach about 945 TWh in 2030 in its base case.
- Local grid concentration matters more to many projects than the global electricity share.
- Hardware, software and utilisation efficiency should be measured at useful-work level.
- Flexible workloads can create strategic value when aligned with grid conditions.
- Power contracts do not eliminate connection, congestion, resilience or additionality questions.
Compute is becoming an infrastructure decision
Artificial-intelligence strategy is often framed around models, data, talent and chips. At scale it is also a decision about electricity systems. Training and inference take place mainly in data centres whose servers, networking, storage, cooling and backup systems depend on reliable power. The location and timing of that demand can matter as much as its annual energy total.
The International Energy Agency estimates data centres used around 415 TWh of electricity in 2024 and projects about 945 TWh by 2030 in its base case, just under 3% of global consumption. The IEA also stresses uncertainty and uses alternative cases for adoption, efficiency and bottlenecks. These are scenarios, not guaranteed outcomes.
A modest global share can still create intense local pressure because facilities cluster near fibre, users, skilled labour and suitable sites. Grid connections and generation infrastructure can take longer to deliver than a data centre. Compute roadmaps that ignore connection queues and system constraints risk turning a software ambition into a stranded construction plan.
Measure useful work, not only megawatt-hours
Energy efficiency should link electricity to a service outcome: completed training run, validated experiment, inference request, useful token, scientific calculation or business process. Facility metrics such as power usage effectiveness remain relevant but cannot reveal inefficient models, idle accelerators, unnecessary precision or low-value workloads.
Create a layered metric set: facility electricity, IT electricity, accelerator utilisation, workload completion, model-quality outcome and avoided or created business activity. Track embodied infrastructure separately from operating electricity. Compare architectures only on equivalent service and reliability requirements.
Rebound effects need attention. Lower energy per inference can reduce cost and stimulate far more inference. The absolute demand path therefore depends on efficiency, adoption, model size, utilisation and product design. A strategy should test these drivers rather than assume efficiency automatically lowers total consumption.
Scenario and location framework
For location l and time t, expected power demand can be represented as Pₗₜ = Wₜ × Eₜ × Uₗₜ + Oₗₜ, where W is workload volume, E is IT energy per workload, U captures utilisation and redundancy, and O is cooling and other overhead. The expression is deliberately simple; it forces separate assumptions rather than hiding them in one growth rate.
Build at least a base case, high-adoption case, high-efficiency case and constrained-grid case. Stress connection delays, curtailment, cooling limits, fuel or power price volatility, equipment delivery, network loss and outages. Evaluate hourly profiles because annual matching of renewable certificates does not demonstrate power availability or carbon intensity at the moment of use.
Location scoring should include grid connection date, firm capacity, generation mix, congestion, market design, climate, water, cooling options, fibre, latency, land, permitting, workforce, security and disaster exposure. Weighting is a strategic choice, so publish sensitivities rather than presenting one precise ranking.
Hypothetical example: choosing between two regions
A hypothetical technology company compares Region A, with fast permitting but a constrained grid, and Region B, with a later connection but more firm low-carbon generation. Under the base workload forecast, A appears cheaper. Under the high-adoption case, connection restrictions delay expansion and require costly temporary capacity. B produces lower regret across scenarios despite a higher initial site cost.
The company stages its decision: reserve land and grid capacity in B, place latency-sensitive inference in a smaller A facility, and keep deferrable training flexible. It establishes trigger points for accelerator demand, queue milestones and power-contract delivery. The numbers and decision are illustrative rather than a recommendation about a real location.
Governance, procurement and public policy
A joint compute-and-energy committee should connect product demand, infrastructure, sustainability, finance, resilience and public affairs. Each large capacity commitment needs a workload thesis, energy scenario, connection evidence, resilience design and exit or repurposing option. Procurement should disclose how projected efficiency was measured and what happens if utilisation is lower than planned.
Claims about clean power require careful boundaries. Distinguish contractual matching, physical delivery, additional generation, hourly alignment and system effects. Water metrics should reflect local scarcity and cooling design rather than one global intensity number. Community impact, network cost allocation and emergency generation also belong in the decision.
Public policy should seek transparent forecasts, faster but credible permitting, efficient grid investment and incentives for flexible demand. It should avoid assuming all AI load is identical. Some inference is latency-critical; many training and batch workloads can shift in time or place if systems and contracts reward flexibility.
Practical implementation and monitoring
Capital approval should connect technical demand to energy evidence. A proposal should show signed or credible connection capacity, expected energisation milestones, load shape, ramp profile, redundancy, backup duration and the assumptions translating user demand into accelerator capacity. Separate contracted power from deliverable power. Record who bears curtailment, imbalance, network-upgrade and delay costs, and test whether the facility retains value if the preferred hardware or model architecture changes.
Operational teams can build a carbon- and congestion-aware workload scheduler where data, security and latency allow. The scheduler should optimise within service-level constraints rather than move work solely to the lowest reported carbon signal. Track completed work, deadline breaches, energy, marginal emissions estimate and curtailed load. Independent review should test whether the signal is timely, locationally relevant and protected against data failure or perverse incentives.
Public reporting should state boundaries and uncertainty. Annual electricity, renewable matching and facility efficiency describe different questions. Water withdrawal differs from consumption; market-based emissions differ from physical grid intensity. Presenting them together with location and time avoids a misleading single score. The strategy is credible when engineering, energy procurement, sustainability and finance use consistent workloads and scenarios.
How to read the current evidence
Evidence about emerging technology has several layers. Binding law and official implementation dates describe obligations; standards and guidance describe expected practices; peer-reviewed or metrology research tests particular mechanisms; institutional scenarios explore conditional futures; and vendor claims describe products under stated or unstated conditions. These layers should not be blended. A result established in one experiment is not proof of sector-wide readiness, while a scenario is not a prediction. Decision papers should label the evidence type, date and relevant jurisdiction beside each material claim.
Uncertainty should be carried into the decision rather than removed through confident prose. Teams should show which conclusion depends on adoption, performance, cost, infrastructure, regulation or behaviour; identify the observation that would change the conclusion; and set a review date. Where evidence is contested, compare sources and methods instead of taking an average of incompatible claims. This discipline is especially important for technology and sustainability, where technical capability, implementation capacity and social acceptance can move at different speeds.
Limitations and interpretation
This article separates enacted requirements and cited institutional evidence from the author's analysis. Hypothetical examples are explicitly illustrative. Technology performance, legal duties and implementation choices depend on context and may change after 16 August 2026; organisations should confirm current requirements and test claims in their own environment.
What decision-makers should do now
- Create joint compute, energy and grid scenarios.
- Measure electricity per useful service outcome.
- Test local connection and congestion constraints.
- Design flexibility for deferrable workloads.
- Assess hourly power, resilience, carbon and water together.
- Use staged commitments and trigger-based capacity decisions.
Conclusion
The central AI infrastructure question is not whether global electricity systems can supply one forecast total. It is whether specific projects can obtain reliable, timely and socially credible power while technology and demand remain uncertain. Organisations that integrate compute architecture with grid reality can make better location, contracting and workload decisions—and avoid confusing purchased capacity with usable strategic capability.
Frequently asked questions
How much electricity could data centres use in 2030?
The IEA base case projects around 945 TWh globally in 2030, while emphasising substantial uncertainty and alternative scenarios.
Why can local impact be large if the global share is small?
Data centres cluster geographically, so connection capacity, congestion and generation can bind in a particular region even when global share remains limited.
What is the best AI energy-efficiency metric?
No single metric is sufficient. Link facility and IT electricity to a defined useful workload and verified service or model-quality outcome.
Can AI workloads support the grid?
Some training and batch workloads may shift in time or place, but flexibility depends on system design, data constraints, latency and contracts.
Does buying renewable power solve the issue?
It can support decarbonisation, but does not by itself resolve connection timing, congestion, hourly matching, resilience, additionality or local impacts.