The Next AI Bottleneck Is Electricity

AI prices depend on more than model quality and chip costs. Power, grid access and cooling now sit behind the cost of running AI. Companies putting AI into everyday work will need to decide which tasks justify expensive models and where cheaper systems are enough.

A chatbot gives little indication of what happens after a user presses enter. The answer appears on a screen, while processors elsewhere run the calculation, networking equipment moves data between machines and cooling systems remove the heat.

The International Energy Agency expects electricity demand from data centres to rise to around 945 TWh by 2030, more than double its 2024 level. AI is driving much of that growth, and power demand from AI-optimised data centres is projected to rise more than fourfold by the end of the decade.

Data centres still account for a small share of global electricity use. The pressure appears where large facilities cluster. A new site needs hundreds of megawatts in one place, often within a few years. Transmission lines, substations and generating capacity take longer to add. The IEA estimates that grid constraints put around 20 percent of planned data-centre capacity through 2030 at risk of connection delays.

AI companies spent the first years of the current boom trying to secure advanced chips. Owning the chips is no longer enough. They also need enough electricity to run them.

Every AI request runs on hardware

Large AI models rely on GPUs and other processors built to perform many calculations at once. Training keeps large clusters working for long periods. Once training ends, the power draw continues whenever users generate text, images, code or video.

A service used by millions of people therefore creates a permanent load. Requests also differ sharply in cost. A short text answer requires less computing than video generation. Reasoning models run longer calculations. AI agents can turn one instruction into several searches, model calls and software actions before returning a result.

Hardware manufacturers are reducing the energy needed for each calculation, and developers are doing more work with smaller models. Total demand is still rising because companies keep adding new AI workloads. The IEA reported a 17 percent rise in data-centre electricity demand in 2025, while AI-focused facilities grew faster.Lower power use per calculation does not mean lower power use across the industry.

Chips do not help if the grid cannot connect them

A company can order processors. Connecting a 500 MW data centre to an electricity grid requires a different timetable. Utilities need enough generating capacity in the region, enough transmission capacity to carry the load and transformers and substations able to handle it. A new data centre can be built within a few years. The IEA says new transmission lines in advanced economies often take four to eight years, while waiting times for transformers and cables have doubled over the past three years.

Developers then have three choices: wait for grid upgrades, help fund them or move. The United States shows how large the load could become. Lawrence Berkeley National Laboratory estimates that data centres will use around 649 TWh of electricity in 2030 in its reference case, equal to 11.8 percent of total US electricity consumption. Its lower and upper scenarios range from 521 TWh to 843 TWh.

Those figures describe national demand, but the commercial problem remains local. Half of the US data centres now under development sit in existing large clusters, according to the IEA. Putting more facilities into the same areas raises the risk that developers all arrive at the same grid bottleneck.For an AI company, processors waiting for a connection are machines that cannot earn revenue.

Electricity is changing where AI gets built

Land, fibre connections and taxes still influence data-centre location. Power availability now removes some sites from consideration before the rest of the calculation begins.

Developers want to know how many megawatts the local grid can supply, how long a connection will take and what electricity will cost once the building opens. Established data-centre hubs do not always produce the best answer. In some European markets, grid connection queues already run for years.

Operators are responding by looking beyond the usual locations and by securing power directly.

The IEA expects renewables to provide nearly half of the extra electricity needed by data centres through 2030. Natural gas also supplies part of the increase, while nuclear generation plays a larger role later in the decade.

Grid delays are also pushing some US developers towards onsite natural-gas generation. The IEA estimates that 15–27 GW of onsite gas capacity could supply data centres by 2030, although turbine shortages and the need for backup capacity limit how quickly those projects move.

AI investment is therefore affecting decisions far beyond software and semiconductors. Utilities, power producers and grid operators now influence how quickly computing capacity reaches the market.

The electricity turns into heat

Processors do not only consume electricity. They release almost all of it as heat. Older data centres relied heavily on chilled air. AI servers pack more computing into each rack, which raises the amount of heat operators need to remove from a smaller space. Many new facilities use liquid cooling closer to the processors.

Cooling adds electricity demand of its own and, depending on the design, water use. Berkeley Lab found more than a 10’000-fold difference in water consumption per computing workload across the systems it studied. Server efficiency produced the largest variation, followed by the amount of water used to generate local electricity, server utilisation, cooling technology, climate and other site conditions.

A single figure for the water footprint of “AI” therefore says little about an individual facility. The same workload run in two locations can use very different amounts of water because the hardware, cooling system and electricity supply differ.

Operators also face trade-offs. A cooling system that uses less water can consume more electricity. A cooler climate reduces part of the cooling load. Newer processors complete more work per unit of power but often arrive in denser racks. Data-centre economics now include all of those variables.

Better chips can still lead to higher electricity use

Computers have become more efficient for decades without reducing the amount of computing people buy.

A lower cost per request makes more applications affordable. Companies start analysing every customer conversation rather than a sample. Search products add generated answers. Office software records and summarises meetings. Developers put AI inside products that previously ran on conventional software.

Agents add another source of demand because one user request can trigger several model calls. Video adds far more computation than text. Longer reasoning increases processing time again.

The IEA tested a scenario with faster improvements in software, hardware and data-centre efficiency. Even then, global data-centre electricity consumption reaches about 970 TWh in 2035 because demand for digital services keeps growing.

Efficiency changes how much electricity each task consumes. It does not decide how many tasks companies choose to run.

Electricity eventually reaches the customer

Most AI customers never see an electricity charge. They pay a subscription, an API fee or a cloud invoice. The provider pays for power along with processors, network equipment, cooling, buildings and financing. Higher electricity prices raise the cost of running the models. Grid delays restrict how quickly providers can add capacity.

For a small company using a chatbot occasionally, the effect is easy to ignore. It becomes harder to ignore once AI sits inside thousands or millions of transactions.

A business using AI for customer support, document processing, search or marketing automation has choices about how much computing each task deserves. Routine classification does not need the same model as a difficult research or coding task. A smaller model often costs less to run and answers faster. Sending every task to the largest model adds cost without improving every result.

Cloud computing already works on the same principle. Companies choose different types of servers for different workloads rather than buying the most powerful option for everything.

AI procurement is moving in that direction. The decision is no longer only which provider has the strongest model. Companies also need to know what each task costs at scale.

AI also cuts energy use elsewhere

Energy companies use AI to forecast demand, monitor equipment and manage electricity networks. Industrial companies use it to adjust production and detect waste. Those applications can reduce energy use or operating costs.

The gains do not cancel the electricity consumed by AI data centres as a group. A model that helps a factory reduce fuel consumption has an outcome that can be measured against the computing required to produce it. Generating large volumes of marketing images has another cost and another return. Calling both “AI” says little about whether the energy use paid for itself.

Data remain thin. AI companies publish little consistent information about the electricity consumed by individual models or types of request, and hardware changes quickly enough to make many estimates obsolete. The IEA has called for better reporting as data-centre demand grows.

Businesses buying AI services therefore have far better information about token prices than about the infrastructure costs behind them.

The bottleneck sits on the local grid

Worldwide electricity production is not about to run out because of AI. Data centres account for around one-tenth of global electricity-demand growth to 2030, less than industrial motors, air conditioning or electric vehicles.

A developer trying to open a data centre outside Dublin, Frankfurt or Northern Virginia has a narrower problem. The site needs a large grid connection. The substation has a limit. The transmission network has a limit. Transformers have delivery dates. New generating capacity has to be financed and built.

Spare electricity in another country does not solve any of those problems.

For AI providers, access to power is becoming part of the cost and availability of computing. For companies buying AI, the consequence arrives through model prices and through the choice of how much computing to spend on each job.

The most expensive model is not the default answer to every business task. As AI use spreads through normal operations, matching the model to the work becomes part of cost control. The next constraint on AI is therefore how much computing operators can afford to power.