How much electricity does one AI task use? TokWatt exposes why published estimates disagree
The open database harmonises 134 estimates across 20 task types and shows a six-order-of-magnitude spread. Most reference values still rest on one source, and the boundary excludes idle capacity, hardware manufacture and user devices.
The short version
- What happened
- TokWatt version 0.1.0 curates 134 estimates from 13 public sources covering 20 task types and records which hardware, server and facility components each estimate counts.
- Who it affects
- A person choosing between a short answer and a long-running agent cannot infer the energy difference from the word 'AI' alone. Transparent task and boundary labels help organisations estimate demand without turning a partial operational figure into a full environmental claim.
- What remains unproven
- Unreviewed preprint
Keep in mind
Unreviewed preprint
Curated rather than systematic review
English-language and web-visible sources
91 of 134 rows from one modelling preprint
12 of 19 reference values depend on one source
Sparse first-party disclosures
Agent values based on two individuals' logs
No frontier closed-video measurement
Idle capacity and embodied energy excluded
Networks and user devices excluded
No audited product ranking
No separate funding or competing-interests statement
Source analysis · 2 linked sources. Primary analysed source date: 29 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
A short answer and an agent task are not the same unit of work
TokWatt, a new open database, tries to make public claims about AI electricity use comparable before they are repeated as universal facts. Its first release contains 134 estimates from 13 sources published between 2023 and 2026, covering 20 task types. After mapping estimates to a common server-plus-facility boundary, the paper reports 0.34 watt-hours for a short chat answer, 4.74 Wh for a reasoning answer, 124 Wh for a short video and 150 Wh for one task delegated to a coding agent.
Those values span six orders of magnitude across the wider task set. That is partly a real difference in workload: generating video or running many agent calls is not equivalent to classifying one text. It is also a measurement problem. Published estimates variously count accelerators, host servers, facility overhead, idle machines or none of those explicitly. TokWatt records five boundary components and the functional unit so readers can see whether two numbers describe the same thing.[1][2]
The harmonisation is reproducible, but the evidence base is lopsided
The inclusion rules required a first-hand public measurement, model or disclosure tied to an identifiable task, convertible to watt-hours, with enough method detail to classify the boundary. Secondary accounts could be stored but not used in a reference set. The target boundary includes accelerators, host equipment and facility overhead; it excludes idle capacity and embodied hardware energy. Sensitivity tests changed five of seven calculator values by nothing and the largest by 12%, suggesting that source disagreement matters more than the chosen conversion factors.
Yet the source mix constrains the precision. One modelling preprint supplies 91 of 134 rows and nine of ten estimates behind the reasoning-answer value. Twelve of 19 reference values rely on one source. The agent numbers come from two individuals' logs, frontier closed video systems have no provider measurements, and only Google, OpenAI and Mistral supplied first-party per-query environmental figures in the reviewed set. The study is a curated English-language collection, not a systematic review.[1][2]
The distinction that matters
Why two AI energy numbers can both be honest
The answer changes with both the task and the system boundary.
Functional unit
Answer, image, video or agent task
Different amounts of computation
Operational boundary
Accelerator + host + facility
TokWatt's harmonised target
Excluded here
Idle capacity, hardware manufacture, network, user device
Reference value below a full footprint
Evidence
134 estimates from 13 sources
Many tasks still rely on one source
Reference values are useful for scenarios, not product labels
Boundary reporting remains sparse: of 132 estimates with an energy value, 127 count accelerator energy and 112 count the host, but only one fully counts idle capacity and none includes embodied hardware energy. Seven estimates from two sources cover all four operational components. Networks and user devices also sit outside the target. That means the values are deliberately narrower than a complete life-cycle footprint and cannot establish the electricity use or emissions of a named data centre.
For product teams and policymakers, TokWatt's strongest contribution is traceability: every estimate has a source locator, functional unit, method and boundary flags, while validation scripts rebuild the derived tables. Its values can support transparent scenarios if the uncertainties travel with them. They should not become claims that every prompt consumes 0.34 Wh or every agent task 150 Wh. Audited provider disclosures, repeated measurements under real serving loads, systematic searches and life-cycle accounting would materially change the assessment. The preprint lists no separate funding or competing-interests statement.[1][2]