AI, energy and infrastructure choices

Align demand, power and the cash commitment

An enterprise can approve an AI application before the grid can supply it. Hardware may arrive before power; power may arrive before useful demand. The resulting idle commitment can disappear in a steady-state cost-per-compute estimate. Leaders need one investment case that aligns application demand, deliverable capacity and the timing of cash—not three plans approved independently.

The IEA's 2026 central outlook projects global data-centre electricity demand rising from 485 TWh in 2025 to 950 TWh in 2030. ¹ This is an updated projection for the entire data-centre sector. It should not be described as AI-only consumption or a guaranteed outcome. The scale is strategically relevant, but the more immediate investment question is local: which connection, substation, equipment supply chain and cooling design will support the proposed workload?

Start with the workload portfolio. Stable interactive demand, schedulable batch work and experimental use have different tolerances for delay, location and interruption. Treating all three as identical can lock uncertain applications into inflexible capacity. mayden proposes matching each class to a capacity commitment that remains useful as demand and technology change.

Keep forecast revisions visible

Forecasts are necessary for long-lived assets, but they need version control. A demand outlook contains assumptions about adoption, hardware efficiency, utilization and the mix of tasks. Improvements in efficiency can reduce energy per task while new uses increase total demand. There is no contradiction: the denominator and the activity level both change. A board should see which assumptions drive its own exposure instead of receiving a single external number presented as certain.

Lawrence Berkeley National Laboratory's 2025 Update, published in June 2026 and hosted by the US Department of Energy, estimates 192 TWh of US consumption in 2024, or 4.7% of national electricity use, and revises earlier historical estimates. ² That revision illustrates a wider measurement issue: even the starting point can change as information on equipment shipments and operations improves. Comparing two forecasts without checking their vintage, geography and scope can manufacture a growth story that is partly a change in estimation.

mayden recommends three internal scenarios anchored in business demand rather than arbitrary percentages: committed services, plausible expansion and an upside requiring additional evidence. For each, specify computing hours, service levels and the capacity that must be available simultaneously. Translate that into an infrastructure range using tested assumptions. Use the scenarios to identify commitments that remain sensible across different futures and decisions that should wait for a clearer demand signal.

Secure the connection, not just the energy

Annual electricity consumption is measured in energy units such as MWh or TWh. A site's connection and equipment are sized around power, measured in MW or GW, with operating constraints and peaks that an annual average can hide. A procurement paper that moves between these quantities without explanation can obscure the real bottleneck. A low annual utilization rate does not necessarily remove the need for a substantial connection if demand is concentrated in particular periods.

The IEA's Electricity 2026 analysis notes that grid infrastructure can take 5–15 years to plan, permit and complete, compared with 1–3 years for data centres. ³ These are broad ranges, not project schedules. The investment implication is to verify the specific connection path before treating announced capacity as deliverable capacity. Land, an energy contract and a hardware order are not substitutes for the electrical infrastructure and permissions needed to operate.

The IEA's September 2026 report on modernising grids broadens the response beyond new construction. It examines monitoring and control, grid-enhancing technologies and AI-enhanced applications as ways to use existing networks more effectively. ⁸ These are complementary routes, with different operational and regulatory requirements. A general technology opportunity does not establish spare capacity at a particular connection point.

For an AI infrastructure investment, ask the grid operator which constraint is binding and which intervention could change it. A thermal limit, a substation bottleneck and an operational security constraint do not have the same remedy. mayden proposes carrying that diagnosis into the capacity contract and project schedule. Until a site-specific assessment supports a change, potential optimization should remain an option in the investment case rather than committed deliverable power.

Gulf site choices need the same local assessment. This report does not assume that all GCC markets share the same electricity economics, network capacity or permitting regime. A site decision requires local evidence about connection timing, reliability, cooling conditions, water and contractual obligations. Ask who bears delay risk and whether the business can use alternative capacity during a shortfall. A staged commitment may preserve more value than an apparently attractive unit price tied to capacity that cannot be energized when required.

Reduce the workload before fixing capacity

The business has more choices than build or buy. It can route different tasks to different models, reuse results where appropriate, schedule flexible work, reduce unnecessary generation and choose a service level that fits the task. These are proposed engineering and procurement options; the report does not claim a universal saving from any one of them. Each needs measurement of quality as well as resource use. A smaller model that creates more retries or review may be less efficient across the complete workflow.

Classify workloads by latency, location constraints, confidentiality, recoverability and variability. Interactive customer services may need predictable response and strict data controls. Internal batch analysis may tolerate a wider processing window. Model training, fine-tuning and inference can have different resource profiles and scheduling options. The right unit of optimization is a successfully completed business task at the required quality, including failed attempts and downstream correction. Tokens generated are a useful engineering measure but an incomplete business outcome.

Use this classification to structure contracts. Reserve capacity for demand that is demonstrably stable, retain flexible purchasing for uncertain demand and specify portability where feasible. Test whether the application can move before relying on exit rights. Software dependencies, data-transfer costs, specialized hardware and operating expertise can make a nominally portable workload difficult to relocate. The CFO and technology leader should examine those frictions together, because they determine whether flexibility is real or merely a clause in a contract.

Separate physical impact from accounting claims

The IEA's 2025 supply analysis distinguishes the electricity physically consumed by data centres from operators' contractual energy mix. ⁴ That distinction should remain visible in executive reporting. Purchasing environmental attributes or signing a renewable contract is different from matching a site's demand with local supply at every hour. Both can be relevant, but they answer different questions. A report should state the accounting method and avoid implying that a contractual claim eliminates all local grid impacts.

Water requires the same care. Google's 2026 environmental reporting describes replenishment projects across watersheds and reports replenishment equivalent to 78% of its 2025 freshwater consumption. ⁵ This is a company-reported stewardship measure, not evidence that every facility is water-neutral or that local consumption has disappeared. mayden's broader recommendation is to examine withdrawal, consumption, source, location and season separately. A global balance can conceal pressure in a particular watershed, just as an annual energy balance can conceal hourly constraints.

The European Commission's data-centre energy-performance work establishes a reporting framework under Delegated Regulation EU 2024/1364. ⁶ Its relevance is the need for comparable information; specific obligations depend on scope and jurisdiction. Procurement teams should ask for operational metrics they can reconcile, with clear site and time boundaries. Avoid combining electricity, emissions and water into one unvalidated sustainability score. Keeping the measures separate makes trade-offs discussable and prevents improvement in one dimension from silently offsetting deterioration in another.

Price the local constraint and the delay

Global averages can understate the effect of concentrated infrastructure. Ireland's Central Statistics Office reports that data centres used 23% of the country's metered electricity in 2025, compared with 5% in 2015. ⁷ This is a national descriptive statistic for all data centres, not an AI-only measure and not a forecast for other countries. Its strategic lesson is that location matters: a modest global share can coexist with a material local demand on infrastructure.

A credible site case should therefore include the community and system boundary. Examine who funds network upgrades, how costs and risks are allocated, what happens during shortages and whether the site's flexibility can support the system. A promise of future jobs or digital investment does not answer these questions by itself. Nor does local opposition automatically prove a project is uneconomic. The decision needs transparent assumptions and a comparison with realistic alternatives, including phased capacity or a different location.

Stress-test the cash profile across connection, commissioning and application milestones. Name the commitments that become sunk before demand is established and the party bearing each delay. Compare the cost of waiting with the cost of committing too early. A low eventual compute price does not compensate automatically for idle equipment or a missed service launch.

Translate useful work into a capacity obligation

A capacity reservation should begin with an account finance and engineering can reconcile. Exhibit 01 uses invented inputs solely to show the units: 1,000 accepted jobs require 1,250 attempts. At an assumed 0.04 kWh per attempt, plus 20 kWh of shared IT activity not already counted, IT energy is 70 kWh. An assumed facility power usage effectiveness of 1.4 gives 98 kWh for the same period and boundary. Over ten hours that is 9.8 kW average facility power. It says nothing by itself about the maximum simultaneous load, memory or network bottleneck. No figure is a vendor price, observed efficiency or recommended design value.

Before using this account in an investment case, replace the assumed attempt energy and shared load with measurements for the intended configuration, including failed work, evaluation and recovery without counting any activity twice. Match facility overhead to the relevant site and observation period; a portfolio average may not describe the selected service. Separately establish the workload schedule, concurrency, required availability and tested peak load. A batch job that can move within a completion window creates different capacity options from an interactive service whose demand arrives together. State which requirement drives the proposed reservation and which demand remains flexible.

Then map that record to a delivery schedule. A facility may have a planned connection date, a separate hardware commissioning date and a later date for the application to use the capacity effectively. Show the cash commitments against each milestone and name the party responsible for delay. Test the consequence of power arriving after equipment, or equipment arriving before proven demand. The point of the exercise is to identify exposures that disappear in a steady-state unit-cost estimate, including idle assets, temporary external supply and staff hired before operations begin.

mayden's proposed decision rule is to reserve irreversible commitments for demand supported by the service case, while pricing flexibility explicitly for the uncertain portion. Compare the premium for a cancellable or transferable reservation with the cost of unused capacity and the revenue at risk from delay. These values are project-specific; this report supplies no universal optimal split. A board may rationally accept an early commitment to secure a scarce site. It should do so with the downside stated and a decision owner, rather than treating a global demand forecast as evidence that its own applications will consume the capacity.

Buy capacity and flexibility for different purposes

Waiting carries a cost where sites and connections are scarce. A staged strategy should therefore price the option to secure capacity, not assume delay is free. Compare committed capacity, conditional expansion and flexible supply against the same workload and service obligations. Each reservation needs a purpose, an exercise trigger and an account of the cost if demand does not arrive.

Bring five questions to the investment decision: which service creates the workload; what quality and availability it needs; when power, cooling and connectivity can be delivered; which environmental and operating boundaries will be measured; and which commitments remain revisable. A missing answer becomes a condition to resolve or a risk explicitly accepted by the accountable executive.

Rebuild the largest proposed capacity commitment from the application portfolio upward. Test lower demand, delayed energisation and a different workload mix. Show the cost of unused capacity beside the cost of being late. Commit where the service and delivery evidence support the exposure; retain options where they do not. Update the decision when either demand or connection evidence changes.