AI has changed the data-center electricity story. Demand is rising much faster than it did through much of the 2000s and 2010s, hyperscale facilities are being proposed and built at extraordinary scale, and utilities in several US regions are confronting loads large enough to affect generation planning, substations, transformers and transmission. Anyone still arguing that AI will have little effect on electricity demand has missed what has happened over the past few years. The harder question is how much confidence to place in forecasts that extend today’s growth rates through 2030. That requires separating two claims that are often bundled together. The first is that AI is already creating a significant new electricity load, especially in the United States and especially in a handful of concentrated data-center regions. The evidence supports that. The second is that today’s relationships among workloads, chips, models, cooling systems, facility utilization and electricity consumption can be projected forward for another four or five years with only modest change. The history of computing gives much less reason for confidence in that assumption. In 2006, US data centers consumed about 60 billion kWh of electricity, roughly 1.5% of national consumption. The Environmental Protection Agency reported that their electricity use had doubled over the previous five years and warned that it could almost double again over the next five. That concern was understandable. The dot-com expansion had been followed by rapid growth in enterprise computing, online services and server infrastructure, and the electricity curve looked steep enough to support alarming extrapolations. By 2014, however, US data-center electricity consumption was about 70 billion kWh, roughly 1.8% of national electricity use. Digital activity had exploded during those eight years, but electricity demand had not followed anything close to the same trajectory. Berkeley Lab later found that growth slowed sharply after 2010 as servers became more efficient, virtualization improved utilization, workloads shifted into more efficient hyperscale facilities and operators squeezed more useful computation out of each unit of infrastructure. The same broad pattern appeared again with cloud computing. The shift to hyperscale facilities initially raised concerns about ever-larger power consumption, but consolidation and better utilization displaced large amounts of inefficient enterprise infrastructure. Cryptocurrency created a genuine new electricity load, especially through proof-of-work Bitcoin mining, but much of the broader blockchain industry moved toward less energy-intensive mechanisms. The COVID period produced an enormous surge in digital activity, yet operators again responded with optimization, capacity management and more efficient infrastructure. AI is different in an important way because its growth has been strong enough to overwhelm some of those efficiency gains. The electricity curve has clearly turned upward again. Large model training, inference at scale and the rapid deployment of specialized accelerators have created real demand, and data-center developers are asking utilities for gigawatt-scale connections in some regions. That should change our assessment of absolute electricity consumption without encouraging us to abandon scale. The International Energy Agency estimates that all data centers globally consumed about 415 TWh of electricity in 2024, around 1.5% of world electricity consumption. Its central scenario has that rising to roughly 945 TWh by 2030, just under 3% of global electricity. A doubling of consumption in six years is substantial. It implies significant investment in generation, grid connections and supporting infrastructure. But less than 3% of world electricity is still a very different proposition from rhetoric suggesting that AI is on course to become the dominant global electricity problem. The United States is the important exception because the load is both larger and much more geographically concentrated. The IEA estimates that the country accounts for roughly 45% of global data-center electricity consumption, and capacity is concentrated further inside a limited number of regional clusters. Berkeley Lab’s latest bottom-up modelling produces a central estimate of roughly 11.8% of US electricity consumption by 2030, with a wide range around it. If the central estimate is realized, that is a major change in the US electricity system rather than a rounding error. The regional implications can be much larger than the national percentage suggests. A cluster of hyperscale facilities can create very large requirements around a relatively small number of substations and transmission corridors, while transformers, switchgear and transmission projects have their own long lead times. Utilities cannot assume that efficiency improvements will make those loads disappear, and regulators should not wait until every forecast uncertainty is resolved before planning infrastructure. The relevant question is how much of the cost and construction risk should be committed against loads that are still partly speculative. Berkeley Lab’s model is considerably more sophisticated than the cruder forecasts built by multiplying the electricity consumption of one AI query by a speculative number of future queries. It starts with expected shipments of computing equipment, estimates annual energy use by device, models different facility types and cooling requirements, and incorporates information about where capacity is planned. That makes it useful for grid planning, but it does not make the 2030 endpoint an observed fact. The result still depends on assumptions about equipment shipments, utilization, facility construction, cooling, model efficiency and the proportion of announced capacity that actually gets built. Five years is an unusually long interval in computing. Between now and 2030 there will be several generations of AI hardware, substantial changes in model architectures, and continued work on compression, quantization, caching, batching and inference routing as companies try to reduce both capital and electricity costs. Cooling systems will change as well, and utilization will matter more as the industry shifts from the current land-grab phase toward normal financial scrutiny of expensive infrastructure. This is familiar from large technology systems. I spent decades helping architect, rearchitect and deploy workloads into enterprise data centers and later public and private clouds. One of the rules repeatedly reinforced in software engineering is to optimize late. Premature optimization wastes time and money because engineers often guess the wrong future bottleneck. Once a constraint becomes materially expensive, however, the engineering priority changes quickly. Compute, memory, network capacity and storage have all gone through that cycle. Electricity is now becoming one of AI’s expensive constraints, so substantially more engineering effort will be directed at reducing the amount required per useful unit of output. Rebound effects mean that efficiency will not necessarily reduce total electricity consumption. Cheaper computation tends to produce more computation, and AI is likely to follow that pattern. A model that requires much less energy per inference may simply be used far more often. But rebound does not make efficiency irrelevant to forecasting. If useful inference becomes ten times less computationally expensive while demand increases fivefold, electricity consumption will still be much lower than a projection based on the earlier energy intensity. Forecasts therefore depend on two uncertain curves at once: how much AI society chooses to use and how much computation and electricity are required for each unit of useful output. Infrastructure constraints create another source of divergence between announced capacity and operating demand. Data-center proposals require financing, chips, customers, transformers, substations, transmission, cooling equipment and enough electricity in the right location. Some proposed facilities will be delayed, downsized or cancelled. Others will move to regions where power can be delivered sooner. Some developers will secure grid capacity for projects that never reach full utilization. Those forms of attrition are normal in large infrastructure pipelines and matter when forecasts are converted into generation and transmission commitments. They also create strong incentives for optimization. When interconnection queues stretch for years, transformers become scarce and utilities start asking developers to pay more of the infrastructure cost, reducing electricity intensity becomes economically valuable in a way it was not when power and grid capacity were abundant. The AI industry does not need to become environmentally virtuous for that response to occur; electricity simply has to become expensive enough to influence capital allocation and engineering priorities. The appropriate planning response is therefore more nuanced than either dismissal or panic. Utilities in major US data-center regions should plan against demanding scenarios because the infrastructure lead times are long and some of the load is already real. Transmission and substation projects should move faster where customer commitments are credible, and regulators should pay close attention to who bears the cost if forecast demand does not materialize. Very large speculative interconnection requests should not automatically become infrastructure investments socialized across all electricity customers. The global discussion needs a different denominator. All of the world’s data centers consumed about 1.5% of electricity in 2024. Even under the IEA’s strong-growth central case, they remain below 3% in 2030. AI can therefore be one of the fastest-growing new electricity loads in the world while still remaining a relatively small share of global electricity consumption. Those statements are entirely compatible. The United States requires more caution because both the scale and concentration are materially different. If Berkeley Lab’s central case occurs, data centers will become one of the largest new sources of electricity demand in the country, and some regions will experience much more pressure than the national average suggests. That deserves serious grid planning. It does not require treating the precise 2030 endpoint as settled. Two decades of data-center history do not demonstrate that the current forecasts are wrong. They show why long-range electricity projections for computing should remain planning scenarios rather than being treated as observations from the future. AI is already creating material electricity requirements in concentrated US markets, and utilities should plan accordingly, but the final load will be shaped by several years of hardware improvement, software optimization, capital discipline, interconnection constraints and project attrition. The electricity demand increase is real. The size of the 2030 endpoint is still being determined. For the deeper analysis of data-center electricity history, AI efficiency, US concentration risk and why 2030 projections deserve a denominator, read AI Electricity Demand Is Rising. The Hype Still Outruns The Load at TFIE Strategy Briefing.
+54 others
