Electricity planners work in decades. Data centre developers work in quarters. That mismatch has become one of the most consequential problems in the technology industry, because the facilities that host artificial intelligence workloads are now among the largest single sources of new electricity demand in several regions.
The International Energy Agency has documented the trend in its electricity market reporting, noting that data centre consumption has moved from a rounding error to a material factor in national forecasts. The growth is driven less by the number of buildings than by the density of the equipment inside them.
Why AI workloads consume so much
Training a large model runs thousands of accelerators at high utilisation for weeks. Inference, the process of answering requests, is smaller per operation but runs continuously and at enormous volume. Both patterns differ from traditional web hosting, where servers idle much of the time and can be shifted or paused without consequence.
Density compounds the problem. A rack that once drew a few kilowatts may now draw tens of kilowatts, and cooling that heat requires either moving large volumes of air or installing liquid systems. Retrofitting an existing hall is often more expensive than building a new one, which is why operators prefer greenfield sites.
The queue for grid connections
In many markets the binding constraint is not land, capital or chips. It is the queue to connect to the transmission network. Interconnection studies can take years, and a project that misses its slot may wait considerably longer. Operators have responded by signing power purchase agreements early, sometimes before a site is fully designed.
This has changed the geography of the industry. Regions with abundant renewable generation and spare transmission capacity have become attractive even when they are far from the customers being served. Latency limits how far some workloads can move, but batch training is far more tolerant of distance than interactive chat.
Nuclear, gas and the question of firm power
Renewable generation is cheap but variable. A data centre that must run continuously needs either storage, which is expensive at this scale, or firm generation. Several operators have signed agreements with nuclear providers, including small modular reactor developers whose designs are not yet operating commercially.
These deals are long-dated and carry technology risk. Gas remains the pragmatic short-term option in many markets, which puts technology companies in the uncomfortable position of increasing emissions while advertising efficiency gains. The tension is visible in corporate sustainability reports that now distinguish between market-based and location-based accounting.
Water, land and local opposition
Cooling consumes water in many designs, and communities in water-stressed areas have begun to object. Land use is also contested, particularly where facilities occupy farmland or sit close to residential areas. Local planning processes that once approved projects quickly are now slower and more adversarial.
Some operators have responded with closed-loop cooling, which reduces water use at the cost of higher electricity consumption. Others publish water usage figures and fund local infrastructure. Neither approach removes the underlying demand, and the political pressure is likely to increase as more facilities are proposed.
What it means for the cost of AI
Electricity is becoming a larger share of the cost of running a model. Where power is scarce or expensive, that shows up either in higher prices or in a decision to move workloads elsewhere. The era in which compute cost was the dominant line item is ending, and power contracts are becoming a strategic asset.
For buyers, this argues for treating energy as part of the architecture rather than a facilities detail. Teams that can schedule flexible work to match cheap hours, or that can shift between regions, will have an advantage over teams that cannot.
Efficiency gains and the rebound effect
Every improvement in computational efficiency reduces the cost of a unit of work, and cheaper work tends to be consumed in larger quantities. This rebound effect explains why efficiency gains in data centre design have not reduced total electricity demand, and why forecasts that extrapolate from per-operation improvements alone keep proving too low.
The pattern is not unique to computing. It is a well documented feature of energy economics, and it means that efficiency should be treated as a way to slow growth rather than a way to reverse it.
Where the pressure will show first
The clearest early signals will come from grid operators publishing connection queues, from regulators reviewing large-load tariffs, and from local planning decisions. Anyone tracking the industry should watch those documents rather than press releases, because they record what is actually being built and when.
The broader lesson is that digital growth now depends on physical infrastructure that takes years to deliver. Software can be deployed in minutes. A transmission line cannot, and that asymmetry will shape the next phase of the AI build-out.
Image: Peter Holmes · CC BY-SA 2.0 · via Wikimedia Commons.