The AI Power Debate Starts at the Grid Connection
A data centre can secure land, announce a supply agreement and order computing equipment while still lacking the grid connection needed to operate at the intended scale. That gap makes local electricity infrastructure central to the AI power debate. National generation totals matter, but a project needs deliverable power at a particular place, with an acceptable connection date and a workable operating arrangement.
The International Energy Agency’s Electricity 2026 report places grid constraints and connection queues among the major issues facing expanding electricity systems. It describes queues involving generation, storage and large loads, with speculative or duplicated applications complicating the picture. A queue total is therefore evidence of pressure on the connection process. It should not be read as a simple inventory of projects certain to operate.
For a proposed data centre, begin with the connection status. A submitted request, an engineering study, an agreed connection and an energized supply are different milestones. Ask what capacity is available initially, what depends on network reinforcement and what obligations the customer accepts. An announced ultimate campus size may be much larger than the first operating phase. Reporting should identify the phase rather than letting the final ambition stand in for current capability.
Distinguish power from energy. Megawatts describe a rate of electricity use or delivery; megawatt-hours describe an amount over time. A 100-megawatt load running continuously for one hour uses 100 megawatt-hours. Its annual use depends on the operating profile, utilization and the facilities included in the measurement. Installed electrical capacity, maximum demand and average consumption should not be interchanged merely because they share a project name.
An electricity purchase agreement answers another question. Depending on its structure, it can support generation development, manage financial exposure or account for a quantity of electricity. It does not by itself establish that the local network can deliver the required power to the data centre at every hour. The physical path, timing and system conditions remain relevant. Treat the contract and the connection as related evidence with different functions.
Storage can help with some operating needs, though its contribution must be described in both power and duration. A battery capable of delivering a particular number of megawatts can do so only for a period determined by its usable energy and operating limits. That can support certain flexibility or resilience objectives. It does not automatically provide an indefinite substitute for the grid or solve every local network constraint.
On-site generation introduces a further project with its own requirements. Fuel supply, emissions, permits, operating limits, equipment delivery and maintenance all matter. A backup generator and a plant intended for routine operation have different roles. An announcement that a campus will use on-site power should identify the actual technology, status and expected duty rather than treating “behind the meter” as a complete solution.
Flexibility can change the connection discussion. Some workloads may be shifted in time or location, and some operators may accept constrained or phased access under specified conditions. The value depends on what the facility can actually interrupt, for how long and with what consequences for customers. A claim of flexible demand needs a credible operating plan. It should not rest on the general observation that software can be scheduled.
The cost question is equally local. Network upgrades need financing, and regulators and utilities decide how costs and risks are allocated within their jurisdictions. A large customer’s project can create benefits while also exposing others to costs if demand is delayed or withdrawn. Evaluate the actual tariff, security and upgrade arrangements before asserting that a data centre necessarily lowers or raises everyone else’s bills.
For energy reporting, follow the chain from the announced load to studies, connection terms, network work and actual energization. For project analysis, follow the same milestones alongside financing and computing equipment. The AI power story becomes clearer when the connection is treated as an engineering and commercial dependency. A distant promise of abundant generation matters much less to an operator than power it can obtain where the servers will stand.