Who pays when an AI data centre reserves power it never uses?

A developer proposes an AI data centre and asks the electricity network for a connection. The requested capacity is enormous. The grid operator must decide whether to reinforce lines, expand substations or reserve generation capacity. But the developer may not yet have financing, tenants, chips or a construction timetable. If the project never arrives, the grid can be left with expensive infrastructure built around a promise.[1]

The International Energy Agency addresses this problem in Powering Artificial Intelligence, published on 14 October. Its review covers 34 jurisdictions representing about 90% of global data-centre capacity. It identifies 134 relevant policies and regulations, with nearly 40% announced or implemented in the preceding year.[1][2]

The interesting number is not merely how many gigawatts the industry's connection requests total. It is how little a requested gigawatt says about future electricity consumption without a credible project schedule, contractual obligations and an understanding of how capacity will ramp up. Some applications are speculative. Others can be duplicated across prospective sites. Even projects that proceed often occupy and equip buildings in stages.[1]

A queue can count the same ambition more than once

The IEA points to large pending data-centre loads in places including Texas, Virginia, Finland and the United Kingdom. In some jurisdictions, cumulative connection requests exceed existing peak demand. That is an extraordinary planning signal, but it is not evidence that all those new loads will switch on. If planners assume every request is real, they may build infrastructure that remains underused. If they assume the queue is mostly fiction, they could cause genuine projects and other consumers to face delays.[1]

The problem is asymmetric in time. A digital service can change demand or financing assumptions quickly; transmission and transformer projects can take years. Equipment may be ordered long before enough customers sign binding contracts. Reliability obligations mean networks cannot simply wait until the power consumption appears on a meter to begin planning.[1]

Who carries the risk of an empty building?

The financial exposure changes with the project’s stage. A reservation can disappear before construction, become a connected site that ramps up slowly, or leave an expensive network asset behind after an early exit. Each situation changes which commitments would protect other consumers without obstructing a credible development.[1]

The IEA describes instruments tailored to those circumstances: financial commitments at queue entry, milestone tests and staged capacity releases for projects that may never materialise; minimum payment terms or capacity-based network tariffs for customers that underuse reserved capacity; and contract length, collateral or exit fees for customers that depart early. These can protect ordinary consumers, but charging repeatedly for the same underlying risk can deter genuine investment or create a windfall for the network.[1]

Earlier access in exchange for flexibility

A second idea in the report is to connect some large loads sooner on non-firm terms. These agreements allow temporary restrictions during defined periods of grid stress. The IEA estimates that such approaches could unlock 750–900 gigawatts of additional hosting capacity globally under its analysis. That is modelled potential capacity, not a forecast of operating data centres or a claim that an equivalent amount of power is continuously available.[1]

The difficulty is who can actually curtail activity. A warehouse may shift machinery use; an AI inference service serving customers at all hours may be less flexible, especially if the building owner does not control the applications inside. A contractual right to curtail is only useful if the data centre can enforce it without violating its own obligations. Operators need predictable conditions, enough time to react and an economic reason to take the offer.[1]

This is also a question about fairness

Data centres can sometimes lower average grid charges by using existing spare capacity and contributing towards fixed costs. They can also raise bills when they trigger major new investment whose costs are spread across everyone. The economic impact is therefore neither automatically a subsidy nor automatically a benefit. It depends on where the facility is, what infrastructure it requires and how tariffs allocate costs over time.[1]

The report’s practical demand is for better commitments behind the queue: evidence of project maturity, a realistic demand schedule and clarity about who pays if plans change. A large connection request becomes useful for planning when it comes with obligations that distinguish an investable project from an option on a possible future.[1]

Report an error or suggest a correction