# Measurement is the bottleneck for green AI

> Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.

Source: https://greenai.guide/articles/measurement-is-the-bottleneck/  
Updated: 2026-11-19

## In short

Most AI energy figures are either reported by the companies themselves, using methods that are not published in full, or estimated from assumptions rather than measured on real hardware. Different studies also draw the boundary of "the system" in different places. Until measurement is consistent, green AI claims are hard to check and comparisons between models or providers can mislead.

## Key points

- The best-known per-prompt figures come from the providers themselves and use different methods, so they cannot be compared directly.
- Changing the measurement boundary can more than double a result; Google's own figure for a Gemini prompt falls from 0.24 Wh to 0.10 Wh if only the AI chips are counted.
- Many tools estimate energy from hardware specifications and grid averages rather than metering it.
- Transparency about AI's environmental costs has been declining, not improving, according to a 2025 analysis.
- A good figure states what was measured, how, over what boundary, on what hardware and for what task.

You cannot manage what you cannot measure, and AI's energy use is still poorly measured. This is not a detail: it
decides whether green AI claims can be checked, whether buyers can compare products, and whether policy can target
the right things. Without trustworthy numbers, every green claim is a press release.

## Why are AI energy figures hard to trust?

**Because most come from the companies selling the AI, use methods that are not fully published, and cannot be
checked independently.** That does not make them wrong, but it makes them hard to compare.

The two most quoted figures for a text prompt are both self-reported:

| Figure | Who | What we know about the method |
|---|---|---|
| 0.24 Wh for a median Gemini Apps text prompt | Google, 2025 | A published paper describes the method, including data-centre overheads and idle capacity. Data from May 2025, not independently verified. |
| About 0.34 Wh for an average ChatGPT query | OpenAI's chief executive, 2025 | A blog post; no methodology published. |
| Roughly 0.3 Wh for a typical GPT-4o query | Epoch AI, 2025 | An independent estimate from public information about hardware and model size. |

The numbers look similar, but they are not like-for-like: a median is not an average, and the boundaries differ. They
also cover only short text prompts, not image or video generation, long reasoning or agents.

## How much does the boundary matter?

**A great deal.** Google's paper makes the point itself. Counting only the active AI accelerators (TPUs and GPUs)
gives 0.10 Wh per median prompt. Adding the host servers, idle capacity kept ready for demand, and data-centre
overheads such as cooling gives 0.24 Wh, more than twice as much.

Where you draw the line decides the answer. A full boundary for an AI service might include:

1. The AI chips doing the work.
2. The rest of the server: processors, memory, storage.
3. Idle and spare capacity held for peaks.
4. Data-centre cooling and power conversion.
5. The network carrying requests and answers.
6. The user's device.
7. Training the model, shared across its uses.
8. Manufacturing the hardware (embodied carbon).

The same applies to bold comparisons. A claim that AI uses many times more, or less, energy than a person doing the
same task depends on what is counted on each side: the person's computer and time, the model's training, the work
needed to check and correct the AI's output. A 2026 UK green AI sandpit of mathematicians, computer scientists and
industry stressed that system boundaries must be stated before any such comparison means anything.

## What is the difference between estimating and metering?

**Metering reads actual power from hardware while the work runs; estimating calculates it from assumptions.**
Both have a place, but they are often confused.

- **Metering** uses hardware counters (such as those built into modern processors and GPUs), power meters on
  servers, or data-centre meters. It is accurate for what it covers but needs access to the hardware.
- **Estimating** multiplies assumed power ratings by run time and utilisation, then by an average grid carbon
  intensity. Tools such as CodeCarbon estimate the power of processors, GPUs and memory and apply regional grid
  intensity. They are easy to use, but they are estimators, not meters.

Estimates can be far out, especially for shared cloud hardware where you cannot see what else is running. Simple
rules of thumb are also risky: Mytton, Lundén and Malmodin showed in 2024 that network energy is mostly a fixed
baseline, so "kilowatt-hours per gigabyte" factors misrepresent the effect of sending more or less data.

The International Telecommunication Union's 2025 report on measuring AI's environmental impact identified heavy
reliance on indirect estimates, without empirical measurement, as a pressing problem.

## Is transparency getting better?

**Not by most measures.** Luccioni, Gamazaychikov, Alves da Costa and Strubell analysed environmental disclosure for
AI models in 2025 and concluded that transparency about the costs and impacts of these models has decreased, even as
their use has grown.

Regulation is starting to push the other way. Under the EU AI Act, providers of general-purpose AI models must
document the known or estimated energy consumption of their models (Annex XI). That documentation goes to regulators
and downstream providers, not to the public, and estimates are allowed where measurement is not available.

## What would good measurement look like?

**A figure you can reproduce: what was measured, how, over what boundary, on what hardware and for what task.**
Useful building blocks already exist:

- **Open benchmarks** such as the ML.ENERGY Leaderboard, which measures the inference energy of open models on
  known hardware, and the [energy labels and benchmarks](/articles/energy-labels-for-ai/) now emerging.
- **Standards** such as ISO/IEC 21031:2024, the Software Carbon Intensity specification, which defines a carbon
  rate per functional unit from energy, grid intensity and embodied emissions.
- **Accuracy alongside energy.** A cheap answer that is wrong is not efficient. Energy figures should be reported with
  the quality of the result, and with how much rework was needed.

Measurement also needs to reach finer levels than "the data centre": a request, a model, or a line of code. That
kind of attribution is what lets engineers find and fix the expensive parts.

## What to do next

- When you see an AI energy figure, ask: who measured it, over what boundary, and was it metered or estimated?
- Measure your own workloads on the hardware you use; see [measuring AI energy](/practice/measure-ai-energy/).
- Report the boundary and method with any figure you publish; see [making honest green claims](/practice/honest-green-claims/).
- Ask suppliers for measured energy per task, not marketing ratios; see [sustainable AI procurement](/practice/sustainable-ai-procurement/).
- Look for tools that help in the [products directory](/products/) and expert help in the [consultants directory](/consultants/).

## Frequently asked questions

### Why can't I compare the energy of ChatGPT and Gemini?

Because the published figures are self-reported, cover different things and use methods that are not fully published. One may include data-centre overheads and idle capacity, another may not. Until both are measured the same way, a direct comparison is not reliable.

### What is the difference between estimating and measuring AI energy?

Measuring reads power from hardware sensors or meters while the work runs. Estimating calculates energy from assumptions such as chip power ratings, utilisation and run time. Estimates are useful when measurement is impossible, but they can be far out.

### What is a system boundary?

The line that decides what is counted. A narrow boundary counts only the AI chip; a wider one adds the server, cooling, idle capacity, the network, the user's device and the manufacturing of the hardware.

## Sources

- [Measuring the environmental impact of delivering AI at Google Scale](https://arxiv.org/abs/2508.15734), Elsworth et al., Google (2025)
- [The gentle singularity](https://blog.samaltman.com/the-gentle-singularity), Sam Altman (OpenAI) (2025)
- [How much energy does ChatGPT use?](https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use), Epoch AI (2025)
- [Misinformation by omission: the need for more environmental transparency in AI](https://arxiv.org/abs/2506.15572), Luccioni, Gamazaychikov, Alves da Costa and Strubell (2025)
- [Measuring what matters: how to assess AI's environmental impact](https://www.itu.int/hub/publication/s-gen-gda-001-2025/), International Telecommunication Union (2025)
- [CodeCarbon](https://codecarbon.io/), CodeCarbon project (2020)
- [The ML.ENERGY benchmark: toward automated inference energy measurement and optimization](https://arxiv.org/abs/2505.06371), Chung et al., NeurIPS Datasets and Benchmarks (2025)
- [ISO/IEC 21031:2024 Software Carbon Intensity (SCI) specification](https://www.iso.org/standard/86612.html), ISO/IEC and Green Software Foundation (2024)
- [Network energy use not directly proportional to data volume](https://doi.org/10.1111/jiec.13512), Mytton, Lundén and Malmodin, Journal of Industrial Ecology (2024)
- [Regulation (EU) 2024/1689 (AI Act), Annex XI](https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-11), European Union (2024)
