AI is getting more efficient. Why the electricity bill keeps growing

The apparent contradiction in AI’s electricity debate is straightforward: an answer can become much cheaper to generate while the industry needs substantially more power. Efficiency is a ratio. Electricity demand is a total. Connecting the two requires knowing how much work is being done—and what kind of work it is.

The International Energy Agency’s April 2026 outlook estimates worldwide data-centre electricity use at 485 terawatt-hours in 2025 and projects about 950 in 2030, roughly 3% of global demand. Those totals cover data centres generally, not AI alone. They are not a meter reading for chatbots.[1]

By contrast, Google’s production study, published in August 2025, reports 0.24 watt-hours for the median Gemini Apps text prompt in its May 2025 snapshot. It reports a 33-fold reduction in that median energy figure from a year earlier. Google explicitly says the findings are neither universal across prompts nor indicative of future performance, and were not independently verified.[2]

Both statements can be true. A vehicle can use less fuel per kilometre while a growing fleet drives far more kilometres. For AI, the complication is that the journeys are changing too: a brief text response, an extended investigation and a generated video are not interchangeable units of useful work.

A median is not a total waiting to be multiplied

The median sits in the middle of a distribution. It is not the arithmetic mean. To estimate total energy by multiplying a request count by a per-request figure, the relevant input is the mean for that same request population and measurement period. A minority of expensive operations can move the mean considerably without moving the median very much.

For an invented example, imagine ninety requests using 0.1 Wh each and ten using 5 Wh each. The median is 0.1 Wh, but the mean is 0.59 Wh: 59 Wh divided by one hundred requests. Multiplying the median by one hundred would produce 10 Wh, substantially understating this example’s total. These numbers illustrate the arithmetic; they are not measurements of Gemini or any other model.

The measurement boundary matters too. Google’s serving method includes host processors, memory, idle capacity and data-centre overhead as well as active accelerators. Counting only active GPU or TPU energy would measure a narrower system. Even the broader serving measure should not be relabelled as the complete lifecycle footprint of training, manufacturing hardware and delivering every related service.[2]

A separate November 2025 research preprint reports more than 32,500 measurements spanning 21 GPU configurations and 155 model architectures, using the vLLM inference engine. Its stated purpose is to examine architectural and operational influences on energy use. It is not a direct measurement of every proprietary service or a reason to assign all requests one universal energy cost.[4]

Ten times more efficient can still mean twice the electricity

For a fixed type of request, total energy is simply request count multiplied by average energy per request. Hold volume steady and an efficiency gain reduces the total. Increase volume faster than the improvement and the total rises. No assumption about a particular AI company is required for either result.

This calculation demonstrates a possibility, not a causal explanation for every new data centre. Cheaper computation might encourage new uses, but observed demand can also rise because products improve, more people adopt them or existing work moves onto the platform. Establishing how much growth was caused by efficiency itself requires more than noticing that the two rose together.

The workload mix adds a second lever. If a service becomes better at easy tasks but its users increasingly request expensive operations, its overall average can rise. That is why “energy per answer” needs a definition of the answer. A completed research report and a one-sentence reply should not be compared as though each were one identical unit of production.

A national forecast is a model, not a reservation at the power station

Berkeley Lab’s 2025 Update, published in June 2026, builds its US outlook from equipment shipments, per-device energy use and cooling performance, including facility types and locations. Its scenarios vary assumptions such as chip shipments, operating lifetime, idle power and utilisation. Those inputs explain why the range is wide: the result depends on how an evolving industry is actually built and operated.[3]

The useful way to read competing forecasts is therefore to ask what each assumes, rather than choose whichever headline number supports a prior belief. Does a proposal describe announced capacity, installed machines or equipment already drawing power? Does it include conventional cloud workloads? What happens if deployment is delayed? A forecast can guide planning without becoming a promise that every projected facility will appear.

Annual energy does not tell a local grid what Tuesday looks like

Power is a rate; energy accumulates over time. A hypothetical facility drawing a constant 100 MW through a non-leap year would consume 876,000 MWh, or 0.876 TWh: 100 multiplied by 8,760 hours. An announced 100 MW connection does not establish that annual consumption. The actual load profile and utilisation are still missing.

The IEA identifies grid connections and constrained supplies of equipment such as transformers and gas turbines as practical limits. Efficiency does not instantly move electricity to the place it is needed, secure a connection or manufacture the equipment that delivers it. The infrastructure challenge is both a quantity problem and a coordination problem.[1]

Emissions introduce another distinction. Google’s carbon-free-energy methodology separates annual renewable matching from matching local consumption with carbon-free supply hour by hour. The same annual electricity use can therefore sit alongside different supply patterns. This is a measurement principle, not a claim that every Google region has already reached hourly carbon-free operation.[5]

For software teams, the productive question is how much resource is required for an accepted result: avoidable retries, oversized models and repeated work belong in that accounting. For infrastructure planners, the questions include location, load shape, equipment delivery and supply. These are connected decisions, but a small per-prompt number cannot settle all of them.

The conclusion is neither that efficiency solves the power problem nor that efficiency is irrelevant. It is the difference between serving expanding demand with less infrastructure and serving it with more. The remaining task is to make demand, workloads and electricity supply visible enough that the gains translate into a system that can actually support the work.

Report an error or suggest a correction