AI energy consumption in 2026 cannot be reduced to one universal “cost per query” number. The defensible range starts with measured production data: Google reports that a median Gemini Apps text prompt used 0.24 watt-hours in May 2025. That is a useful anchor, not a conversion factor for every ChatGPT, Claude, Gemini, or open-model request. Model size, generated tokens, batching, quantization, hardware utilization, cooling overhead, and regional grid mix can move the result by multiples or orders of magnitude.
The grid-scale story is less ambiguous. The International Energy Agency projects global data-centre electricity consumption rising from about 485 TWh in 2025 to 950 TWh in 2030, with AI-focused facilities tripling their demand. The important practitioner question is therefore not whether one short prompt resembles a few seconds of appliance use. It is which workloads multiply across millions of requests, how efficiently they are served, and where the resulting capacity lands on constrained grids.
The honest answer: there is no single per-query number
“One AI query” is not a stable unit of work. A classification call that emits one label, a retrieval-augmented answer with 20,000 input tokens, and a reasoning run that generates several thousand hidden and visible tokens may all appear as one request in an API log. Their compute paths are materially different.
Measurement boundaries differ too. Some estimates count only accelerator power while tokens are generated. A production-grade total should also include the host CPU and memory, idle capacity reserved for reliability, networking, storage, cooling, and power conversion. Google’s production measurement methodology explicitly includes those elements. Quoting an accelerator-only result beside a facility-level result without stating the boundary creates false precision.
- Input size: long contexts increase prefill computation and memory traffic.
- Output size: autoregressive decoding repeats work for every generated token.
- Model architecture: dense and mixture-of-experts models activate different parameter footprints.
- Serving efficiency: batching and high utilization amortize fixed overhead across requests.
- Hardware: accelerator generation, precision, memory bandwidth, and interconnect affect joules per token.
- Facility: PUE adds cooling and power-distribution overhead to IT energy.
This is the same systems problem practitioners encounter with oversized context windows: an advertised capability is not the same thing as an efficient operating point.
What a measured AI prompt consumes
The best public anchor currently available is Google’s measured median: 0.24 Wh per Gemini Apps text prompt, alongside 0.03 grams of CO2-equivalent and 0.26 millilitres of water. The company says the estimate covers active accelerators, host systems, idle machines and data-centre overhead rather than only the chip’s active draw.
| Workload | Publicly measured result | What it does not prove |
|---|---|---|
| Median Gemini Apps text prompt | 0.24 Wh | It does not establish a universal ChatGPT electricity cost. |
| Text classification, 1,000 inferences | 0.002 kWh mean in the Luccioni study | It does not represent open-ended generation. |
| Image generation, 1,000 inferences | 2.907 kWh mean in the same study | It is not comparable to one short text response. |
The cross-task gap matters more than the headline number. In Power Hungry Processing, Luccioni, Jernite and Strubell measured 88 models and found that multi-purpose generative workloads were substantially more energy-intensive than task-specific inference. Treating every model call as interchangeable obscures the largest optimization opportunity: use the smallest capable system for the task.
For a rough scale check, one billion prompts at 0.24 Wh each would consume 240 MWh. That multiplication is arithmetic, not a traffic estimate for any provider. It assumes every request matches Google’s measured median and therefore should be labelled as a scenario, not reported as observed demand.
From one prompt to grid-scale AI data center power
Per-query accounting can make AI look trivial because a watt-hour is small. Grid planning sees the concurrency, uptime requirement, geographic concentration and growth rate instead. The IEA’s updated base trajectory puts data centres at roughly 950 TWh in 2030, around 3% of global electricity demand. Its earlier detailed analysis reaches a similar 945 TWh figure and estimates roughly 15% annual growth from 2024 to 2030.
That global share needs context. Data centres account for less than 10% of worldwide electricity-demand growth in the IEA base case, behind larger forces such as industrial output, air conditioning, electric vehicles and broader electrification. But data centres are geographically concentrated. A cluster requesting hundreds of megawatts at one interconnection point can be a severe local constraint even when the global percentage looks modest.
The IEA expects the United States, China and Europe to dominate demand, with the United States and China contributing nearly 80% of global data-centre electricity growth through 2030. AI workloads also create fast power swings. That changes the engineering problem from buying annual renewable-energy certificates to securing firm capacity, transmission, storage and flexible load at specific hours and locations.
The site’s broader analysis of the AI hardware race is relevant here: accelerator roadmaps increasingly have to be read as power-density and infrastructure roadmaps, not only throughput roadmaps.
Why token count and model choice dominate
Inference has two main phases. Prefill processes the input context efficiently in parallel. Decoding generates output autoregressively and is difficult to parallelize across time because each token depends on the previous token. That is why output length often has disproportionate energy impact.
Green Prompting finds generated response length strongly associated with inference energy. The Price of Prompting similarly shows why monitoring needs CPU- and GPU-level measurements rather than request counts alone. A useful internal metric is therefore watt-hours per million output tokens, segmented by model, region, hardware and service tier.
Context is still consequential. Large retrieval payloads increase prefill work, memory pressure and latency. Teams operating production RAG systems should measure whether each retrieved chunk changes answer quality. Sending 30 marginally relevant chunks because the context window allows it is an energy, latency and quality failure at the same time.
- Route extraction, tagging and classification to small task-specific models.
- Set output-token budgets by task rather than one global maximum.
- Cache deterministic and semantically equivalent requests where policy permits.
- Batch offline workloads and run them when the grid and serving fleet have capacity.
- Quantize models only after measuring the quality loss on the actual workload.
- Use retrieval filters and compression before expanding context limits.
For teams deciding whether to change the model or the data path, the practical framework in Fine-Tuning vs RAG vs Prompt Engineering helps separate capability problems from retrieval and orchestration problems.
Training energy versus inference energy
Training is a large, bounded event. Inference is a smaller event repeated for the model’s operational life. Which one dominates depends on usage. A rarely deployed research model may spend most of its lifetime energy budget during training. A heavily used consumer service can accumulate more inference energy within weeks or months.
Frontier-training power is nevertheless moving into grid-planning territory. Epoch AI and EPRI estimate that the largest individual runs could draw 4–16 GW by 2030 if scaling trends and infrastructure permit it. The authors explicitly note that feasibility is uncertain, especially at the upper end. This is a scenario about instantaneous power for extreme runs, not a statement that all 2030 models will consume gigawatts.
The distinction between power and energy is load-bearing. A 5 GW training cluster is a power requirement. Its energy depends on duration and utilization. A low-power inference service can consume more annual energy if it runs continuously at global scale. Articles that compare a training run’s megawatt rating with annual household kilowatt-hours without converting units are not making a valid comparison.
Where AI data center power actually goes
Accelerators are only part of the facility. The IEA estimates servers average roughly 60% of electricity demand in modern data centres, with substantial variation by facility type. Storage, networking, cooling, pumps, fans, uninterruptible power systems and conversion losses consume the rest.
Power Usage Effectiveness expresses total facility energy divided by IT-equipment energy. A PUE of 1.17 means every 1 Wh delivered to IT corresponds to approximately 1.17 Wh at the facility boundary. It does not measure useful computation, carbon intensity or water consumption, and a low annual PUE can hide poor accelerator utilization.
Carbon impact adds another variable: time and place. The same 0.24 Wh prompt produces different emissions on a grid dominated by hydro, nuclear, gas or coal. Hourly carbon-aware scheduling can reduce emissions without changing energy. It cannot solve an interconnection queue or a local capacity shortage. Those require transmission, generation, storage or flexible demand.
How practitioners can measure and cut inference energy
The first operational rule is to stop using request count as the denominator. Instrument input tokens, output tokens, latency, batch size, accelerator utilization, cache hits, model version and quality outcome. Join those records to node- or rack-level power telemetry. Without the quality field, an energy optimization can simply produce worse answers more efficiently.
| Metric | Why it matters | Action it enables |
|---|---|---|
| Wh per 1M output tokens | Normalizes decoding work | Compare models and hardware |
| Wh per successful task | Includes quality and retries | Prevent false efficiency wins |
| Accelerator utilization | Reveals idle capacity | Improve batching and scheduling |
| Context tokens per task | Exposes retrieval bloat | Compress or filter context |
| Regional marginal carbon intensity | Captures time-and-place emissions | Shift flexible workloads |
Local inference can also make sense when privacy, predictable utilization or latency outweighs cloud elasticity. It is not automatically greener: an underutilized workstation may deliver worse joules per token than a well-batched data-centre service. Benchmark the actual deployment using the same discipline described in our local LLM setup comparison.
FAQ
How much electricity does one AI query use?
There is no universal value. Google’s measured median Gemini Apps text prompt used 0.24 Wh, but longer outputs, reasoning workloads, larger models and less efficient serving can consume materially more. Always state the model, token counts and measurement boundary.
Does AI training or inference consume more electricity?
It depends on deployment volume. Training is a large bounded event; inference repeats for the model’s lifetime. At high request volume, cumulative inference can surpass training.
Why do estimates for ChatGPT electricity cost differ?
They often measure different models, prompt lengths, output lengths, hardware and system boundaries. Some count only GPU power; others include hosts, idle capacity and facility overhead. OpenAI has not published a universal per-query figure that resolves those variables.
Will AI data centers overwhelm the grid?
Globally, the IEA projects data centres at around 3% of electricity demand in 2030, so “overwhelm” is too broad. Locally, concentrated multi-hundred-megawatt and gigawatt facilities can create serious generation, transmission and interconnection constraints.
The next step: measure energy per successful task
The useful conclusion is not that AI queries are harmless or catastrophic. It is that request counts hide the engineering variables that teams can control. Start with one production workload. Record model, input and output tokens, retries, accelerator utilization, facility overhead and task success. Then test a smaller model, a shorter output budget, tighter retrieval and better batching. The winning metric is not the lowest watt-hours per prompt. It is the lowest energy per task that still meets the required quality, latency and reliability.
{“@context”:”https://schema.org”,”@type”:[“TechArticle”,”Article”],”headline”:”AI Energy Consumption 2026: Cost per Query”,”description”:”Measured AI prompt energy, data-center demand, training versus inference, and a practical framework for reducing energy per successful task.”,”datePublished”:”2026-07-25″,”dateModified”:”2026-07-25″,”author”:{“@type”:”Organization”,”name”:”AI Made”,”url”:”https://aimade.tech/”},”publisher”:{“@type”:”Organization”,”name”:”AI Made”,”url”:”https://aimade.tech/”},”mainEntityOfPage”:{“@type”:”WebPage”,”@id”:”https://aimade.tech/ai-energy-consumption-2026-cost-per-query/”},”keywords”:[“ai energy consumption 2026″,”ai data center power”,”chatgpt electricity cost”]} {“@context”:”https://schema.org”,”@type”:”FAQPage”,”mainEntity”:[{“@type”:”Question”,”name”:”How much electricity does one AI query use?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”There is no universal value. Google’s measured median Gemini Apps text prompt used 0.24 Wh, but model, token count, hardware and measurement boundary materially change the result.”}},{“@type”:”Question”,”name”:”Does AI training or inference consume more electricity?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”It depends on deployment volume. Training is a large bounded event, while inference repeats for the model’s lifetime and can exceed training at high usage.”}},{“@type”:”Question”,”name”:”Why do estimates for ChatGPT electricity cost differ?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Estimates use different models, prompt and output lengths, hardware, utilization and accounting boundaries. Some count only GPU power while others include hosts, idle capacity and facility overhead.”}},{“@type”:”Question”,”name”:”Will AI data centers overwhelm the grid?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”The IEA projects data centres at around 3% of global electricity demand in 2030, but concentrated facilities can still create serious local generation, transmission and interconnection constraints.”}}]} {“@context”:”https://schema.org”,”@type”:”SpeakableSpecification”,”cssSelector”:[“#ai-energy-intro”]} {“@context”:”https://schema.org”,”@type”:”ClaimReview”,”url”:”https://aimade.tech/ai-energy-consumption-2026-cost-per-query/”,”datePublished”:”2026-07-25″,”author”:{“@type”:”Organization”,”name”:”AI Made”},”claimReviewed”:”A median production AI text prompt can use approximately 0.24 watt-hours, but this is not a universal per-query value.”,”reviewRating”:{“@type”:”Rating”,”ratingValue”:”4″,”bestRating”:”5″,”worstRating”:”1″,”alternateName”:”Supported for Google’s measured Gemini median; not universal”},”itemReviewed”:{“@type”:”Claim”,”appearance”:{“@type”:”CreativeWork”,”url”:”https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference”},”author”:{“@type”:”Organization”,”name”:”Google”}}}