AI Energy Use Calculator

Understanding the AI Energy Use Calculator

An AI energy calculator estimates annual electricity and emissions from daily query volume and energy per query. The per-query figure is where nearly all the uncertainty sits, and published estimates for it vary by more than an order of magnitude.

How it actually works

Enter queries per day and estimated watt-hours per query. The calculator multiplies for daily kilowatt-hours, annualises, and applies a grid emission factor of 0.4 kg CO2 per kWh. Five hundred queries a day at 0.3 Wh each gives 0.15 kWh daily, about 55 kWh yearly, and 22 kg of CO2.

The deeper context most people miss

Comparisons like these are widely circulated and rest on shaky foundations. Most per-query figures are estimates derived from limited disclosure, they vary enormously with model size and response length, and image and video generation consume far more than text. Treat any single number as an order-of-magnitude indication rather than a measurement.

Why per-query energy figures are so uncertain

Providers publish little about inference energy, so most circulating figures are inferred from hardware specifications, estimated utilisation, and assumptions about model size, and different analysts making reasonable assumptions arrive at estimates differing by a factor of ten or more. Several things drive genuine variation rather than estimation error. Model size matters directly, since generating each token requires reading the model's weights from memory, so a model with ten times the parameters consumes roughly proportionally more per token. Response length matters as much, since energy scales with tokens generated and a long answer costs many times a short one. Input length affects the prefill phase. Batching matters enormously in the other direction: serving many requests together amortises the weight reads across them, so a busy endpoint is far more efficient per query than an idle one, which means utilisation is a first-order variable. Hardware generation matters, with newer accelerators delivering substantially more inference per watt. Reasoning models change the picture again, generating extensive internal reasoning before answering and consuming correspondingly more. Quantisation, distillation, and other efficiency techniques reduce it. Given all this, a single per-query figure necessarily averages across enormous variation, and the honest position is that precise personal footprint calculations for AI use are not currently well supported by available data.

A worked example: putting the number in context

Fifty-five kilowatt-hours a year for 500 daily queries is roughly what a modern fridge uses in a month or two, and the 22 kg of CO2 is comparable to driving a petrol car somewhere around 60 to 100 miles. For an individual, this is a small share of a personal footprint that typically runs several tonnes annually, and it sits far below flights, driving, home heating, and diet. That framing matters because individual AI use has attracted attention disproportionate to its share of personal emissions, while the aggregate picture is genuinely significant and operates at a different level. Data centre electricity demand has grown substantially and AI workloads are a meaningful and rapidly growing part of it, with several national grid operators and the International Energy Agency identifying it as a material factor in electricity planning. Some utilities have delayed coal retirements and approved new gas generation citing data centre demand, which is where the environmental consequence actually bites. The distinction is between an individual's marginal use, which is small, and infrastructure buildout driven by aggregate demand, which is not. Both things are true and conflating them produces bad reasoning in either direction: individual abstention achieves little, and the sector's aggregate trajectory is a legitimate concern that is addressed through grid decarbonisation, efficiency, and siting decisions rather than through individual restraint.

Deciding what actually reduces AI energy use

For individuals, the levers are modest and worth knowing anyway. Using a smaller model where it suffices reduces energy roughly in proportion to parameter count, and many tasks genuinely do not need a frontier model. Requesting concise output reduces generated tokens, which is the main driver. Avoiding unnecessary regeneration and refining a prompt rather than repeatedly rerunning helps. Image and especially video generation consume far more than text, so casual generation of media is where individual usage becomes non-trivial. For organisations the levers are larger: routing simple requests to smaller models, caching responses to repeated queries, batching where latency permits, and choosing providers and regions with cleaner grids all matter, and several providers publish regional carbon intensity to enable that choice. For the sector, the decisive factors are hardware efficiency improvements, which have been rapid, model efficiency improvements, which have also been rapid, and the carbon intensity of the electricity supplying data centres, which is the variable with the largest leverage. Several major operators have made substantial renewable procurement commitments, and the credibility of these varies between genuine additional generation and certificate purchasing, which is worth distinguishing when assessing claims.

Water, hardware, and the parts that get less attention

Electricity dominates discussion and two other factors deserve mention. Water consumption for data centre cooling has become a recognised issue, particularly where facilities are sited in water-stressed regions, and evaporative cooling consumes water directly rather than merely circulating it. Reported figures vary and disclosure is limited, and several communities have raised objections to facility siting on this basis. Water use trades against energy use, since less water-intensive cooling generally consumes more electricity, so the two cannot be minimised simultaneously. Embodied emissions from manufacturing accelerators and building facilities are substantial and are frequently omitted from operational figures, and given how rapidly hardware generations turn over, the embodied share is not negligible. Electronic waste from decommissioned hardware follows. On the other side of the ledger, applications of the technology may reduce emissions elsewhere through better grid management, materials discovery, and efficiency optimisation, and several analyses have argued the net effect could be positive, though these estimates are speculative and depend on deployment choices rather than following automatically. The honest summary is that the sector's environmental accounting is incomplete, disclosure is limited, and confident claims in either direction, whether that AI is an environmental catastrophe or that it will solve climate change, are running ahead of the evidence.

Variations: modality, model type, and grid factors

Text generation is the least energy-intensive common workload. Image generation consumes substantially more per output, with published estimates suggesting a large multiple of a text query. Video generation is far more intensive again. Speech synthesis and recognition sit between. Embedding generation is very cheap per item. Reasoning models consuming extended internal reasoning use considerably more than their nominal size suggests. Training and inference are separate accounts, with inference dominating over a deployed model's life at any real traffic. On grid factors, the 0.4 kg per kWh used here approximates a global or US average, and actual intensity ranges from under 0.05 in hydro and nuclear heavy grids to over 0.7 in coal-dependent ones, so the emissions figure can be wrong by more than an order of magnitude depending on where the compute physically runs. Time of day matters too where grids vary hourly. Several providers now publish region-level carbon data and some support carbon-aware scheduling for flexible workloads, which is more effective for batch processing than for interactive queries.

Thinking about AI energy use sensibly

Treat any per-query energy figure as an order-of-magnitude indication, since published estimates vary by more than tenfold and depend on model size, response length, batching, and hardware generation. Recognise that individual usage is small relative to a personal footprint, sitting well below flights, driving, home heating, and diet, so individual abstention achieves little. Attend instead to the aggregate picture, where data centre demand growth is a material factor in electricity planning and has influenced generation decisions in several regions. Use smaller models where they suffice and request concise output, since energy scales with model size and tokens generated. Note that image and especially video generation consume far more than text, which is where personal usage becomes non-trivial. Adjust the grid factor for your region if emissions matter, since intensity ranges from under 0.05 to over 0.7 kg per kWh. And treat confident claims in either direction sceptically, since disclosure is limited and the sector's environmental accounting remains incomplete.

What people get wrong

• Treating a per-query energy figure as a measurement, when published estimates vary by more than an order of magnitude and depend heavily on model size and response length.
• Applying a single grid emission factor, when carbon intensity ranges from under 0.05 to over 0.7 kg per kWh depending on where the compute physically runs.
• Conflating individual usage with aggregate impact, when personal AI energy sits well below flights and heating while sector-wide demand growth is a genuine planning concern.
• Assuming text and image generation are comparable, when image generation consumes a large multiple per output and video generation far more again.

Where the math comes from

Daily kWh = (Queries per Day × Watt-hours per Query) / 1,000. Yearly = Daily × 365. CO₂ = Yearly kWh × 0.4 kg, using an approximate average grid emission factor. Both inputs carry large uncertainty: per-query energy varies by more than an order of magnitude with model size, response length, and batching, and grid intensity varies similarly by region.

Questions and answers

Related calculators

Embedding Cost · AI Chatbot Total Cost · AI Inference Cost · Fine-Tuning Cost · AI Image Generation Cost