AI's Next Bottleneck Is Electricity, Not Just Chips

The AI race is usually described in terms of processors, models and investment. But a building full of advanced chips is not useful until somebody can supply it with reliable electricity. That requirement is making substations, grid connections and power equipment increasingly important to the industry's expansion.

The issue is not that the world is about to run out of electricity. It is that a fast-moving technology industry is asking a slower-moving physical system to accommodate large new loads in particular places. A promising national supply forecast does not automatically translate into an available connection at a particular site.

What the headline numbers actually measure

In its April 16 report, Key Questions on Energy and AI, the International Energy Agency projected that global data-centre electricity consumption would roughly double from 485 terawatt-hours in 2025 to 950 TWh in 2030. That would represent around 3% of global electricity demand in 2030. These are the report's central projections, not guaranteed outcomes.^[1]

Crucially, those totals cover data centres generally, not AI alone. The same infrastructure category also supports conventional computing workloads. Within that broader total, the IEA projects electricity consumption from AI-focused data centres to triple over the period. Saying that all data-centre demand will triple, or that AI alone will consume 950 TWh, would misrepresent the analysis.

The wider electricity system is growing too. The IEA's July 23 Electricity Mid-Year Update 2026 forecasts global electricity demand growth of 3.6% in 2026 and 3.8% in 2027. It identifies industry, appliances, cooling, electric vehicles and other electrification alongside data centres as drivers.^[2] AI is an important addition to that demand, not the explanation for every new power plant or higher electricity bill.

Energy is not the same as a grid connection

Terawatt-hours measure energy used over time. Megawatts measure power: the rate at which electricity must be delivered. A facility's annual consumption tells planners something important, but it does not tell them everything about its maximum demand, its fluctuations or the equipment needed to connect it.

A grid may have enough generation across a year while lacking transmission capacity at a specific location. Conversely, a connection sized for an eventual full buildout may initially serve only a fraction of the anticipated load. Both situations complicate investment decisions.

The IEA identifies this mismatch explicitly. Data centres can develop quickly, fill progressively with servers and initially request oversized connections. Power infrastructure moves on a different timetable. Building too little can delay projects; building too much for demand that never materialises can leave expensive assets underused.^[1]

This is why a project's announced capacity is not the same as operating electricity demand. A credible expansion story needs more than a site and a chip order. It needs a plausible path through connection approvals, equipment supply and actual occupancy.

Efficiency improves, but demand can still rise

There is a genuine counterweight to the growth story: AI systems are becoming more efficient. The IEA reports substantial reductions in energy consumption per task as software and hardware improve. That matters, and it makes blanket claims about the electricity cost of an AI query increasingly unhelpful.

But a query is not a standard unit of work. A short text response, a generated video and an agent carrying out multiple steps can require very different amounts of computation. The April report finds that some newer applications can consume hundreds or thousands of times more energy per query than simple text generation.^[1] That comparison concerns different tasks, not a universal multiplier for every AI interaction.

Training and inference also need separating. Training develops a model; inference runs it to produce outputs. Repeated use means the electricity story does not end when training finishes. Nor does the distinction mean every training job is easily postponed or every inference request needs an immediate response. Scheduling flexibility depends on the actual workload and service requirements.

The result is a moving target. More efficient systems can reduce the electricity needed for a given task while growing adoption and more demanding applications increase total consumption. Neither efficiency alone nor today's usage patterns settle the forecast.

A modest global share can mean concentrated local pressure

The projected 3% global share can sound reassuring. For a community hosting several large facilities, however, the relevant question is what those facilities add to the local system. Electricity infrastructure is geographically specific, and global averages cannot reveal a constrained substation or the cost of upgrading a transmission corridor.

That does not mean household bills must rise. The IEA argues that additional demand can improve utilisation and lower costs where supply is underused. Where the system is tight, it can trigger new investment. Who ultimately pays depends partly on tariff design and how those costs are allocated.^[1]

The practical questions are therefore contractual and regulatory as well as technical. What happens if the customer builds less than promised? Who funds the connection and wider reinforcement? What commitments protect other users if a project is cancelled? These are more useful questions than assuming data centres are either automatically beneficial or inevitably a burden.

More supply helps, but timing still matters

The July outlook forecasts renewables overtaking coal in global electricity generation during 2026.^[2] That is significant context, but it does not establish that every additional AI workload runs on low-carbon electricity. The emissions associated with extra demand depend on where and when it occurs and which generation responds at the margin. A national annual average cannot fully answer that question.

Building generation next to a facility is not an automatic shortcut either. The April report notes interest in onsite natural-gas power, particularly in the United States, while warning that turbine supply constraints and other unresolved questions can limit its speed advantage.^[1]

Flexibility offers another route. The IEA recommends exploring demand response and non-firm connections, which allow supply to be limited under agreed conditions, in exchange for potentially faster access. Batteries and suitable workload scheduling can help, but their value depends on operating requirements and incentives. Flexibility must be demonstrated, not merely promised.

The next stage of AI expansion will therefore be measured partly in energised connections, delivered equipment and workable electricity contracts. Chips still matter. So do model capabilities. But the industry's physical limits will increasingly be negotiated with grid operators, regulators and the communities sharing the same wires.