Artificial intelligence is often described in terms of chips, models and software. A growing share of the capital behind it is committed to something more physical: buildings, cooling equipment, electricity connections and the power plants and networks needed to keep them running. That changes the investment question. A data centre can be technically ready, commercially attractive and still unable to begin full operations on its intended timetable if the power connection does not arrive.
The International Energy Agency projects that electricity generation serving data centres will rise from about 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030 under its base case. That is a scenario, not a guaranteed outcome. Its value for investors lies in showing the scale of infrastructure that may be needed alongside computing equipment. Demand forecasts, power availability and project completion dates all have to align before anticipated revenue materialises.
The bottleneck is local
Global energy supply is an imperfect guide to the viability of an individual site. A developer needs deliverable capacity at a particular location, on a particular date and at a cost the business can bear. The relevant questions include the grid connection, transmission upgrades, reliability, local tariffs, land use, cooling water and the power provider’s ability to meet peak demand. A region can have abundant electricity in aggregate while a specific connection remains constrained.
The US Department of Energy cited research estimating that data centres accounted for about 4.4% of US electricity use in 2023 and could account for roughly 6.7% to 12% by 2028. The range is wide because demand depends on deployment, technology and efficiency. It also illustrates why investment cases built on a single precise forecast deserve scrutiny. A lender assessing a proposed campus needs a timetable and downside case, not only a national demand projection.
Grid queues complicate that timetable. The Federal Energy Regulatory Commission’s 2025 markets review describes rising load forecasts and large interconnection queues in the United States. A queue position alone does not establish when an individual generation or transmission project will be built. Financing should distinguish a completed power supply arrangement from a proposed one and a confirmed connection date from an estimate.
The financial exposure extends beyond the building
Data centre investments can involve a sequence of commitments made before full demand is visible: land acquisition, construction, equipment orders, energy contracts and network upgrades. Cash flows may arrive later than debt service or construction costs. If power is delayed, a tenant may defer commissioning, negotiate phased occupancy or shift workloads elsewhere. The effect depends on the contracts, but the mismatch between spending and revenue is a common financing concern.
This is an analytical inference from the physical project structure, not a claim that all AI projects face the same risk. Established facilities with secured connections and contracted customers differ from speculative projects that still depend on planning approvals or future grid investments. Investors should also distinguish the credit of a large technology tenant from the credit of a single-purpose project vehicle. A strong prospective customer does not automatically remove construction, delivery or counterparty risk.
Energy procurement adds another layer. A long-term power purchase agreement can improve price visibility, but it is not the same as guaranteed supply at every hour. The IEA’s energy supply analysis finds that renewables may meet nearly half of the additional data centre electricity demand to 2030, while natural gas and coal also provide substantial near-term supply in its base case. The actual mix depends on geography. An investment case should describe both its contractual electricity purchases and the physical system that serves the site.
Efficiency can change the forecast
The amount of computing delivered per unit of electricity is not fixed. Chip design, cooling, server utilisation and the efficiency of AI models may alter the electricity required for a given amount of useful work. Greater efficiency could also make new uses economical, increasing total demand. Neither effect should be treated as inevitable. An IMF working paper on AI energy demand examines the uncertain link between AI adoption and power use, underscoring the need for a range of outcomes.
Financiers can test those outcomes explicitly. One scenario assumes that customers use the reserved capacity as expected. Another delays energisation or occupancy. A third assumes weaker pricing because capacity comes online faster than customer demand. Each scenario changes the period during which interest, construction costs and operating expenses must be funded. The useful result is not a single forecast return; it is an understanding of which assumption would break the capital structure.
Who bears the infrastructure cost
Transmission and generation upgrades may require investment by utilities and other parties, raising questions about cost allocation. A project can appear attractive if it assumes that someone else funds network expansion, but the resulting tariff, connection charge or contractual obligation still affects economics. Developers, regulators and utilities may reach different solutions across markets. The US Department of Energy has pointed to rate structures and grid investment as part of the response to data centre growth.
There are broader market risks too. The Financial Stability Board has noted that AI adoption may amplify third-party dependence and market correlations within finance. For investors in the underlying infrastructure, concentration may arise when many projects depend on a small group of technology buyers, chip suppliers or power-constrained locations. That observation supports careful exposure mapping; it does not prove that today’s spending is excessive.
A contract is only as useful as its conditions
An investor should read the conditions attached to a power commitment. Does the agreement specify capacity at the site or only an intention to supply? Are delivery dates conditional on permits, upstream transmission or a generation project? Does the customer pay for unused contracted power, and can the price change materially if network charges rise? These terms determine whether an apparently secured input supports a reliable forecast or merely shifts uncertainty into a different document.
A phased build can reduce exposure. If a facility opens in several blocks as power becomes available, construction spending and tenant commitments may be aligned more closely with operational capacity. That approach can also increase unit costs or complicate procurement. The relevant comparison is the risk-adjusted cost of phasing against the cost of completing an entire campus before its power and customers are certain. A lender should see the cash-flow effect of both choices.
Different capital providers see different risks
An equity investor may welcome the option to expand a site if future AI demand is strong. A lender is more concerned that contracted cash flows cover interest and repayment through a downside case. A utility financing new equipment needs confidence that the associated load materialises. These different incentives shape negotiations over guarantees, minimum payments, construction milestones and termination rights. A headline project value obscures which party ultimately bears a delay.
Residual value also matters. Computing hardware can become outdated more quickly than power infrastructure or a well-located site. If expected computing density changes, the same building may require different cooling or electrical equipment. That possibility argues for modular designs and careful assumptions about replacement spending. It also means a long tenant lease should be analysed alongside any capital obligation to upgrade the facility during its term.
What would a stress case show
A useful sensitivity analysis should combine setbacks rather than vary one input at a time. A grid connection delayed by a year can coincide with higher financing costs and lower occupancy. A new power plant can be completed on schedule while transmission work lags. Customer demand might arrive earlier than expected, leaving the operator to secure costly interim supply. Stress testing these combinations reveals which commitments are firm and which are contingent.
The analysis should distinguish a national forecast from contracted demand at the site. The IEA scenario helps frame the scale of the sector, but it does not validate an individual developer’s utilisation rate. Evidence includes signed leases, customer concentration, commissioning milestones, connection agreements and a realistic contingency budget. Investors can then compare projected returns with the probability and financial effect of delays, rather than assuming that all demand for AI translates into demand for every building.
The grid relationship will shape returns
Data centres could sometimes provide flexibility by shifting workloads or using storage, but the value depends on the type of computing and the local market. Some tasks can be scheduled; others need constant availability. Claims that a site will support the grid should identify the equipment, contracts and response times that make flexibility real. Revenue from grid services should be treated as conditional until the mechanism and performance obligations are clear.
Power availability is therefore an issue of underwriting, not a prediction that AI adoption will stall. The more useful question is which projects can connect, operate and earn cash under a range of credible scenarios. That requires technology investors and infrastructure financiers to examine the same physical constraints, even when their business models and time horizons differ.
The practical question is whether power, compute and customer demand arrive in the right order. AI may create durable demand for infrastructure, but a powerful demand story does not make every site financeable. Projects with credible connection dates, transparent cost sharing, resilient supply and capital structures that withstand delays offer a clearer basis for evaluation than projects that treat electricity as a footnote.
Sources
Federal Energy Regulatory Commission’s 2025 markets review