AI Infrastructure Is Disrupting Energy Demand Forecasts

Energy demand forecasting has long rested on a simple assumption: Demand usually changes in ways that can be observed, measured and reasonably projected. For decades, models have translated economic growth, demographics, policy choices and fuel shifts into long-term estimates. Even major disruptions have often moved through identifiable channels, whether through prices, industrial activity, mobility or investment cycles. Demand was not treated as a mystery, but as something that could be measured and managed. That premise is breaking down as artificial intelligence infrastructure introduces forms of demand that are faster, more concentrated and less visible to traditional models — forcing a rethink of how forecasting is done.

Forecasts always depend on what you’re looking at and when you are doing so. In weather, accuracy declines sharply as uncertainty compounds: A five‑day outlook is right about 90% of the time, a seven‑day about 80%, and beyond 10 days, accuracy drops toward 50%. Energy demand used to be different. It moved slowly, with stable causes. You can see that in how forecasters performed under real stress. Medium-term projections by bodies such as the International Energy Agency (IEA) and US Energy Information Administration, covering five to 20 years, usually land within a few percentage points of actual demand. Even when something breaks the pattern, the break itself tends to follow a recognizable shape. Take Covid‑19. Global energy demand fell by roughly 5%-6% in 2020, the biggest drop in decades, and it moved closely with economic activity and mobility. Demand also began to recover as economies reopened. Even under extreme pressure, the main drivers remained visible, which allowed forecasts to adjust rather than collapse.

When Patterns Break

This logic starts to crack when demand is driven not by one‑time shocks but by steady, accelerating and spiraling technological changes. Covid‑19 had clear bounds: It peaked and then slowly eased. AI progress has no such clear limits: Each improvement enables the next, and speed can pick up with little warning. The result is not just more uncertainty. It is a shorter window for forecasters to adjust.

Moreover, AI is described as a General Purpose Technology, spreading across sectors and enabling new applications. A study done by HAI Stanford University showed that 78% of organizations reported using AI in 2024, and another report from McKinsey & Co. in 2024 showed that AI-related patents now appear across technology sectors far beyond the traditional information and communication domain. Historically, such patterns have preceded large-scale deployment. If this holds, the expansion of AI will continue to drive demand for computational infrastructure, and with it, electricity consumption. In effect, demand is becoming less of a function of steady economic activity and more a function of technological scaling.

Scaling Compute, Scaling Power

Data centers already account for roughly 1%–2% of global electricity consumption, and their share is rising. AI adds pressure because its systems rely on dense computing infrastructure. A single large training cluster can contain tens of thousands of processors, with advanced GPUs drawing hundreds of watts each. Training frontier models has moved from a relatively small electricity load a decade ago to workloads that can require tens or even hundreds of megawatt hours. At that scale, efficiency gains do not automatically reduce total demand. They can also make larger models and wider deployment economically possible.

The difficulty is not only the level of demand but also the forces behind it. Efficiency may reduce the energy needed for each calculation, yet total consumption can still rise as models grow and more computing capacity is installed. This is why forecasts now vary so widely. The IEA notes that “substantial uncertainty both about data center consumption today and in the future” while institutional projections for 2030 range from roughly 200 to more than 1,000 terawatt hours.

The Limits of Forecasting Models

Existing forecasting models are still necessary for reading established links between economic activity, prices, policy and demand. But AI infrastructure changes the role these models can play. They should no longer be treated as tools for producing one confident central pathway, because AI-driven demand does not simply add volatility to familiar patterns. It can create discontinuities. Large increments in electricity use may come from rapid deployment decisions, such as the expansion of data center capacity, rather than from gradual shifts in economic variables. Models calibrated on historical relationships may therefore miss not only the scale of change but also its timing and trigger points. Their role should be less about extending the past into the future and more about testing how demand could shift when technology scales faster than the historical record would suggest.

Forecasting approaches therefore need to pay closer attention to the pathway through which technologies move from development to large-scale use. This includes, in practical terms, tracking the maturity of AI systems, their movement through technology-readiness stages and the speed with which they become commercially deployable. It also means monitoring infrastructure build-out more closely, including data center investment, grid connection requests, chip availability, cooling requirements and corporate commitments to new computing capacity. These are not always captured well by traditional macroeconomic indicators. A national forecast may still track GDP growth, industrial output and energy prices, but AI-related demand can be triggered by decisions taken at the corporate or project level. A single investment decision by a major technology firm can add a concentrated block of future electricity demand before it appears clearly in broader economic data. In this sense, forecasting has to move closer to the micro-foundations of demand: who is building, where they are building, how fast capacity can be connected and whether the underlying technology is ready to scale. The forecasting exercise is therefore no longer only about reading the economy from the top down. It also requires watching the technology and infrastructure pipeline from the bottom up.

A New Geography of Demand

A deeper shift is also taking place in the geography of energy risk. For much of modern energy history, the map was shaped by relatively stable forces: national development paths, industrial bases, population growth, in addition to fixed supply routes. AI infrastructure changes this picture. Large computing facilities can cluster wherever power, land, cooling and connectivity can be secured, and the timing of these clusters is increasingly shaped by corporate investment decisions rather than long national trajectories. This creates a more fluid map of demand. The energy system may no longer be defined only by familiar supply chokepoints, it may also develop multiple localized versions of the Strait of Hormuz with similar sensitivities, scattered across grids and created by decisions to build, expand or concentrate computing capacity. These are places where local demand can rise quickly, strain infrastructure and alter wider planning assumptions.

For forecasters, the central question therefore changes. It is no longer enough to ask how much electricity AI will consume, as the harder question is where that demand will appear, how quickly it will arrive and whose decisions will trigger it. AI infrastructure is therefore not simply adding demand to the energy system, it is changing the way demand becomes visible.

Ayman Akil is the Mena director of academia and government at global analytics company Clarivate (formerly part of Thomson Reuters). He writes in a personal capacity, and the views expressed are his own, not necessarily those of Clarivate.