How much energy does AI use? Training vs inference

Few questions about AI get asked more often, or answered worse, than how much energy it uses. The figures in circulation are striking, widely repeated, and frequently years out of date, and they are deployed with equal confidence by people arguing that AI is an environmental catastrophe and by people arguing it is a rounding error. Both cases can be built from real numbers, which should tell you something about the state of the evidence. The energy story of AI is usually told in the wrong unit: the per-query figures everyone quotes are stale and vary by an order of magnitude depending on what you ask, the load that actually matters is inference running continuously rather than training running occasionally, and the binding constraint is not generating electricity but delivering it to a specific site, which is why efficiency gains keep getting absorbed instead of banked.

This guide explains where AI's energy actually goes, why the balance between training and inference flipped and what that changed, the constraint that limits AI infrastructure in practice, why models getting dramatically more efficient has not reduced total consumption, and why the numbers in this area deserve more scepticism than they usually receive. The aim is a picture accurate enough to reason with, including honesty about how much remains uncertain.

The number everyone quotes, and why it misleads

The most repeated statistic about AI energy is that a single chatbot query uses roughly ten times the electricity of a conventional web search. The underlying figure, around 2.9 watt-hours per query against roughly 0.3 for a search, comes from estimates made in 2024, and it has been quoted continuously ever since.

There are three problems with leaning on it. The first is that it is probably stale: more recent measurements put the median energy of a short text query an order of magnitude lower, in the range of a few tenths of a watt-hour, as models and serving infrastructure became more efficient. The second is that a median hides enormous variation. Energy per query depends on the size of the model, the length of your input and its output, and above all on whether the model is doing extended reasoning, so a brief factual answer and a long chain of deliberation over a large document can differ by orders of magnitude. Quoting one number for "an AI query" is like quoting one number for "a car journey." The third is scale intuition. Even the higher estimate is small in isolation, comparable to running a household bulb for a short while, which is why per-query framing invites the conclusion that individual use is trivial.

That conclusion is correct and beside the point, and this is the heart of the matter. AI's energy significance was never about what one query costs. It is about the aggregate: a small number multiplied by billions of daily interactions, embedded into search results, office software, phones, and development tools, running continuously and growing. The right question is not what a query costs but what the total load looks like and, more importantly, where it is going.

The flip: training was the story, inference is the story now

For years the standard framing of AI's energy cost focused on training, and the headline numbers were dramatic. Training a large frontier model consumes on the order of hundreds of megawatt-hours to more than a gigawatt-hour of electricity over weeks or months of continuous computation across thousands of accelerators, and estimates for training a model of the GPT-3 generation land around 1,300 megawatt-hours. Those figures are real and they capture attention because they are concentrated and easy to picture.

But training has a property that makes it the less important half of the story: it happens once. A model is trained, and then the job is finished. Inference, the work of actually running the model to answer requests, happens every time anyone uses it, forever, at whatever scale the product reaches. And once hundreds of millions of people are using AI daily, that continuous load overtakes the periodic spike.

That crossover has now clearly happened. Current estimates put inference at roughly eighty to ninety percent of AI computation and a clear majority of its energy consumption, having risen from around a third of compute a few years ago. The exact share depends on who is counting and how, but the direction is not in dispute. This inverts the intuition most people carry, which is that training is the expensive part and using the model is cheap. Per unit, that is true. In aggregate, it stopped being true, and the practical consequence is that AI's energy footprint is now driven by adoption rather than by the scaling laws that govern training. Every new integration of AI into a product that millions of people touch adds permanent load, in a way that one more training run does not.

Where the energy actually goes

Inside a facility, the electricity divides into a few broad categories. The largest share goes to the accelerators themselves, the GPUs and specialized chips doing the matrix arithmetic, which are power-hungry by design and are run as close to continuously as possible because idle hardware is wasted capital. A substantial further share goes to cooling, because dense racks of accelerators generate heat that must be removed, and modern high-density AI deployments have pushed operators toward liquid cooling precisely because air can no longer keep up. The remainder covers networking, storage, and power conversion losses.

The industry's traditional efficiency metric, power usage effectiveness, measures how much of a facility's total electricity reaches the computing equipment rather than being spent on overhead. It is useful but limited, because a facility can score well while running inefficient models that waste energy per unit of useful output. That is why attention has moved toward measures of useful work per unit of energy, such as tokens generated per watt, which capture what a system actually produces rather than only how efficiently it is powered.

The constraint that actually binds: delivery, not generation

Here is the part that most coverage misses, and it is the most practically important fact in this area. The limiting factor on AI infrastructure is generally not whether enough electricity can be generated in aggregate. It is whether power can be delivered to a particular site, on a particular timeline, through the local grid.

A large modern AI facility can require power on the order of hundreds of megawatts, which is the scale of a small city concentrated in one location. Connecting a load like that requires transmission capacity, substations, and grid interconnection approvals, and those take years to build and permit, considerably longer than it takes to buy and install accelerators. The result is a bottleneck that has little to do with chip supply: operators can acquire hardware faster than they can secure places to plug it in. This is why siting decisions have become dominated by where power is available rather than where land or talent is, why operators have pursued long-term arrangements with generators including nuclear, and why several jurisdictions have begun reviewing whether planned facilities are compatible with existing infrastructure. It also explains the local dimension of the issue, since concentrated demand affects regional grids and prices in a way that a global percentage figure completely obscures. The distinction matters because "can the world generate enough power for AI" and "can this county deliver five hundred megawatts to this parcel by next year" are different questions with different answers, and only the second is currently binding.

Why efficiency gains keep getting absorbed

A reasonable expectation is that as models become more efficient, total energy consumption should fall. Efficiency has improved substantially, through better chips, better serving software, quantization that shrinks the memory and computation each parameter requires, mixture-of-experts architectures that activate only a fraction of a model per token, and distillation that compresses capability into smaller models. Some recent models have achieved large reductions in the energy required per unit of output.

Total consumption has risen anyway, for two reasons that both deserve stating plainly. The first is the classic rebound effect: making something cheaper causes more of it to be used. Cheaper inference means AI gets embedded in more products, invoked more often, and applied to tasks that were previously not worth the cost, so the savings per query are consumed by growth in queries. The second is more specific and more recent. The frontier has shifted toward reasoning models that deliberately spend far more computation at inference time, generating long internal chains of deliberation before answering, because doing so improves accuracy on hard problems. This is a direct trade of energy for capability, made on purpose. So the same period that produced remarkable per-token efficiency gains also produced a class of models designed to use many more tokens per answer. Efficiency and consumption are moving in opposite directions, and treating improvements in the former as a solution to the latter misreads the dynamic.

The measurement problem

Anyone trying to pin down AI's energy use quickly runs into a harder obstacle: the data is poor. The major operators do not publish detailed, workload-level breakdowns of energy consumption, so nearly every widely quoted figure is an estimate constructed from partial disclosures, hardware specifications, and assumptions about utilization. Different methodologies produce results that differ by large factors, projections for the same year vary by hundreds of terawatt-hours across sources, and figures are frequently repeated long after the systems they described have been superseded.

This is a familiar problem in a new place. Much as with AI benchmarks, the numbers are treated as measurements when they are closer to estimates with wide error bars, and they are then quoted with a confidence the underlying evidence does not support. Because the range is so wide, both alarming and reassuring conclusions can be constructed honestly from published figures by selecting among them, which is roughly what has happened in public discussion. The appropriate response is not to dismiss the question but to hold the numbers loosely: the direction of travel, sharply upward, is well established, while precise magnitudes and forecasts are not, and any source presenting a single confident figure for AI's energy use is overstating what is known.

What can be said with reasonable confidence

Stripping out the contested parts leaves a picture that is still useful. Data centres as a whole account for a low single-digit percentage of global electricity consumption, with AI workloads the fastest-growing component within that. AI's total energy use is rising quickly and is expected to keep rising, though estimates of how far and how fast vary widely enough that specific 2030 figures should be treated as scenarios rather than predictions. Inference dominates and its share is growing, which ties consumption to adoption. Efficiency per unit of work is improving substantially and is being outpaced by growth in the amount of work. And the practical constraint in the near term is grid delivery and local infrastructure rather than global generating capacity.

For an individual, the implication is that personal AI use is a small part of a personal energy footprint, and choosing not to use a chatbot is not a meaningful environmental action, which is worth saying because the per-query framing implies otherwise. The significant decisions are made at the level of infrastructure: where facilities are built, what generation they are connected to, how efficiently models are served, and whether the energy comes from sources that make the growth tolerable. AI's energy question is an infrastructure and policy question wearing the costume of a personal one.

The short version

AI's energy use is widely discussed and poorly measured. The commonly quoted figure that a chatbot query uses about ten times a web search comes from 2024 estimates, and newer measurements put a median short text query an order of magnitude lower, while long or reasoning-heavy queries can cost far more, so a single per-query number is close to meaningless. The per-query framing also misses the point, because AI's significance comes from aggregate load rather than individual use. The central shift is that inference has overtaken training: training a frontier model consumes hundreds of megawatt-hours but happens once, whereas inference runs continuously for every user, and now accounts for roughly eighty to ninety percent of AI compute and most of its energy. That ties AI's footprint to adoption rather than to model scale. Inside facilities, most power goes to accelerators and cooling, and the limiting constraint is usually not generating electricity but delivering hundreds of megawatts to a specific site through a local grid, which takes years longer than buying hardware. Efficiency has improved through better chips, quantization, mixture-of-experts, and distillation, but total consumption has risen anyway, because cheaper inference drives more use and because reasoning models deliberately spend more computation per answer. Underlying all of this, the measurement is weak, since operators do not publish detailed breakdowns and estimates vary by large factors.

The idea to hold onto is that AI's energy story is told in the wrong unit: per-query figures are stale and wildly variable, the load that matters is continuous inference rather than periodic training, and the constraint that actually bites is delivering power to a particular place rather than generating it in aggregate, which is why real efficiency gains keep being absorbed by growth rather than reducing total demand. Treat any single confident number in this area, in either direction, as a sign that someone is arguing rather than measuring.

Common questions

How much energy does an AI query use? There is no single reliable figure, and this is the most misunderstood part of the topic. The widely quoted estimate of roughly 2.9 watt-hours per query, about ten times a conventional web search, dates from 2024, while more recent measurements put the median short text query an order of magnitude lower, in the range of a few tenths of a watt-hour. Actual consumption varies enormously with model size, input and output length, and especially whether the model performs extended reasoning, so a brief answer and a long deliberation over a large document can differ by orders of magnitude. Any single number quoted for "an AI query" conceals that range.

Does AI training or inference use more energy? Inference, by a wide margin, and this reverses a common assumption. Training a frontier model is dramatic but periodic, consuming hundreds of megawatt-hours to over a gigawatt-hour across weeks or months, and then it is done. Inference is the work of running the model for every user request, and it continues indefinitely at whatever scale the product reaches. Current estimates put inference at roughly eighty to ninety percent of AI computation and a clear majority of its energy, up from around a third of compute a few years ago. The practical consequence is that AI's energy footprint is now driven mainly by adoption rather than by how large models are.

How much of the world's electricity does AI use? Data centres as a whole account for a low single-digit percentage of global electricity consumption, commonly estimated at around one and a half percent in recent years, with AI workloads the fastest-growing part of that total rather than all of it. Projections for the rest of the decade vary widely between sources, sometimes by hundreds of terawatt-hours for the same year, which reflects genuine uncertainty about adoption rates, efficiency gains, and how the accounting is done. The upward direction is well established; the specific magnitudes should be treated as scenarios rather than forecasts, and forecasts in this area have a mixed track record.

Why is AI energy use hard to measure? Because the operators with the data do not publish detailed workload-level breakdowns, so almost every public figure is an estimate assembled from partial disclosures, hardware specifications, and assumptions about how heavily equipment is utilized. Different methodologies yield results that differ by large factors, and figures are often repeated years after the systems they described were replaced by more efficient ones. The result resembles the situation with AI benchmarks: numbers that are treated as measurements are closer to estimates with wide error bars, which is why both alarming and reassuring conclusions can be honestly constructed by selecting among published sources.

What limits how much AI infrastructure can be built? Usually power delivery rather than power generation, or chip supply. A large AI facility can require hundreds of megawatts at a single site, comparable to a small city, and connecting that load requires transmission capacity, substations, and grid interconnection approvals that take years to permit and build. Accelerators can be purchased and installed far faster than the electrical infrastructure to run them can be provisioned. This is why siting is now driven by where power is available, why operators pursue long-term generation arrangements, and why some jurisdictions have begun reviewing whether planned facilities are compatible with local grid capacity.

Are AI models becoming more energy efficient? Yes, substantially. Efficiency has improved through better accelerators, improved serving software, quantization that reduces the memory and computation each parameter needs, mixture-of-experts designs that activate only a small fraction of a model per token, and distillation that compresses capability into smaller models. But total energy consumption has continued rising regardless, for two reasons. Cheaper inference causes more inference, as AI is embedded in more products and invoked more often, a classic rebound effect. And reasoning models deliberately spend far more computation per answer to improve accuracy, trading energy for capability on purpose. Efficiency per unit of work and total work are moving in opposite directions.

Should I use AI less to save energy? Individual AI use is a very small part of a personal energy footprint, and the per-query framing that dominates coverage tends to overstate the significance of personal choices here. The consequential decisions happen at the infrastructure level: where facilities are sited, what generation they connect to, how efficiently models are served, and whether growth is matched by low-carbon supply. That does not make the aggregate question unimportant, since aggregate load is exactly what matters and it is rising quickly, but it does mean the lever is policy and infrastructure rather than individual restraint. AI's energy question is an infrastructure issue that is often discussed as a personal one.