How Much Energy Does ChatGPT Use?

Why Is There No Single Number?

A watt-hour (Wh) is the energy a 1-watt device uses in one hour. Estimates for one ChatGPT prompt run from a few tenths of a watt-hour to more than 30 Wh, because they describe different models, different prompt lengths and different slices of a data center.

The closest thing to an official figure came in June 2025, when OpenAI CEO Sam Altman wrote that "the average query uses about 0.34 watt-hours."^[1] That is a company statement. The post names no model and gives no method or scope. Microsoft researchers, listing it alongside other estimates, mark it as unclear whether it was measured or estimated.^[4]

How Have the Estimates Changed Over Time?

The widely repeated older figure is about 3 Wh, from a 2023 estimate by Alex de Vries. An analysis by Josh You at Epoch AI says it assumed 4,000 input tokens and 2,000 output tokens per query, older A100 chips and maximum server power. (A token is a chunk of text, often part of a word.)^[2] Microsoft's table lists that estimate at 2.9 Wh and calls it outdated.^[4]

Newer estimates are roughly ten times lower:

• Epoch AI, February 2025: about 0.3 Wh for a GPT-4o text query, assuming 500 output tokens and about one second of an H100 chip. This is a back-of-the-envelope analysis, not peer reviewed, and the author calls it relatively pessimistic.^[2]
• Jegham and colleagues, 2025 preprint: 0.42 Wh for a short GPT-4o prompt of 100 input and 300 output tokens, modeled from public speed benchmarks and inferred hardware.^[5]
• Google, 2025 preprint: 0.24 Wh for the median Gemini Apps text prompt in May 2025, measured in Google's own production systems. This is self-reported and is not ChatGPT.^[3]
• Microsoft researchers, 2026 in the journal Joule: a median of 0.31 Wh, with a middle range of 0.16 to 0.60 Wh, modeled for large open models on H100 servers. The typical query has 500 input tokens and a median of 300 output tokens. This is also not ChatGPT.^[4]

The drop does not show that ChatGPT became ten times more efficient. Much of the gap comes from different assumptions about tokens, chips and how busy the servers are.^[2] Real efficiency gains also exist: Google reports a 33-fold drop in energy per median Gemini text prompt from May 2024 to May 2025, credited to software, but that says nothing direct about OpenAI's systems.^[3]

What Changes the Number?

The model and the prompt length. The Jegham preprint shows both at once. These are modeled averages, in Wh per query, not measurements.^[5]

Outside estimates disagree here too. Microsoft's long-query scenario, with a median of 5,000 output tokens, comes to 3.91 Wh.^[4]

Pasted text. Epoch estimates that a 10,000-token input, about a long magazine article, raises a query to about 2.5 Wh, and that 100,000 tokens takes it to almost 40 Wh. That is mostly a one-time cost, since follow-up messages in the same chat do not repeat it.^[2]

What gets counted. Google's own example is the clearest. Counting only active AI chips in its most efficient data centers gives 0.10 Wh. Adding host processors and memory, idle standby machines and building overhead gives 0.24 Wh. Training, external networks and your own device are left out of both.^[3] The Jegham figures also leave out idle GPUs and assume a batch size of 8, and larger batches would lower them.^[5]

Average or median. Google uses the median because a few huge prompts skew the average.^[3] Altman's figure is an average, so the two are not directly comparable.^[1]

Images and video. None of these figures cover them. The Microsoft authors say image and video generation have higher energy use.^[4]

How Does Training Compare With Everyday Use?

Training is a large cost paid up front, while everyday use, called inference, keeps running.

The best-documented example is old: Patterson and colleagues reported 1,287 MWh to train GPT-3, using data OpenAI measured.^[6] Epoch estimated in 2025 that models comparable to GPT-4o drew about 20 to 25 megawatts for about three months.^[2] Multiplying that out, with three months taken as 2,160 hours:

• 20 MW × 2,160 h = 43,200 MWh
• 25 MW × 2,160 h = 54,000 MWh

For the everyday side, a working paper using OpenAI's message counts reports 2,627 million messages per day on consumer plans in June 2025.^[7] If each message used Altman's 0.34 Wh, that gives:

• 2,627,000,000 × 0.34 Wh ≈ 893 MWh per day
• GPT-3's training: 1,287 ÷ 893 ≈ 1.4 days of that use
• A GPT-4o-class training run: 43,200 to 54,000 ÷ 893 ≈ 48 to 60 days of that use

This is a rough cross-check using figures from different years, and it leaves out business plans and images.

What Does That Look Like at Home?

These comparisons are our own arithmetic from the sourced inputs. The EIA reports that US residential customers averaged 865 kWh a month in 2024, which is 10,380 kWh a year or about 28.4 kWh a day.^[8] Google's paper uses about 100 watts for a modern TV.^[3] The EPA uses 19 Wh for a smartphone charge.^[9]

Per-person math does not settle total demand. The IEA estimates all data centers, not just AI, used about 415 TWh in 2024, and projects about 945 TWh by 2030.^[10] The Microsoft authors note that lower energy per query can encourage more use, so total energy can still rise.^[4]

Myths vs Facts

What We Still Can't Tell You

We can't give you ChatGPT's actual energy use today. We ran no measurements of our own, and we did not find a published OpenAI per-prompt figure for GPT-5 or later models, or a peer-reviewed direct measurement of ChatGPT's production systems. The GPT-5 numbers are modeled from outside, the Microsoft numbers come from open models, and the Google numbers come from Gemini. Most of these figures describe 2025 systems, and models have changed since. We also cannot tell you what any one prompt of yours used, because nothing here reports that.