AI vs Google Search energy: is AI really 10 times worse?
The famous 10-times comparison mixes an old search figure with changing AI estimates. A fair comparison depends on the task, system and year.
There is no dependable rule that one AI message uses exactly ten times the energy of one Google search. The famous ratio combines a Google estimate from 2009 with early assumptions about ChatGPT, then treats both as timeless constants. Search infrastructure, AI hardware and the products themselves have changed.
Recent public estimates place an ordinary ChatGPT text query at around 0.3–0.34 watt-hours. That is close to the old 0.3 Wh Google search estimate often used as the denominator, not ten times it. But this does not prove that modern AI and modern search use the same energy. Google does not publish a current universal per-search figure, and many searches now include AI-generated summaries. The honest comparison is a range tied to a defined task and boundary.
Where the ten-times claim came from
Google wrote in 2009 that a search used about 0.0003 kWh, or 0.3 Wh. That was a company estimate for the search system of its time. It predates today’s data centres, ranking models, rich result pages and generated summaries.
The AI side was often set at about 3 Wh per query. That estimate assumed older accelerator hardware, a GPT-3.5-scale model and a long output. Epoch AI’s review of the older calculation explains why it considers those assumptions unrepresentative of typical current text use.
Divide 3 Wh by 0.3 Wh and the memorable ratio appears. The arithmetic is easy; the inputs are not comparable.
What newer AI estimates say
OpenAI chief executive Sam Altman stated in 2025 that an average ChatGPT query uses about 0.34 Wh. Epoch AI separately modelled a typical GPT-4o request at roughly 0.3 Wh. Neither figure is a meter reading for every message.
Epoch’s result depends on assumptions about model architecture, hardware, utilisation and token counts. Altman did not publish a supporting calculation. Their agreement is a useful signal, but it cannot fill the missing modern search denominator.
Long-context, reasoning and tool-heavy AI requests can use much more than the typical text case. One AI message’s energy cost varies because “message” describes an interface action, not a fixed amount of computation.
Search and chat do different jobs
A conventional web search ranks and returns links. A chatbot generates an answer token by token. If the user only needs a website, opening a search result is likely the more proportionate tool. Asking a model to write an essay is not meaningfully comparable with retrieving ten blue links.
The boundary gets blurrier when search includes generated summaries, or when a chatbot searches the web before answering. One visible request can combine ranking, retrieval, page processing and one or more model calls. Counting that as either “a search” or “an AI query” hides most of the system.
A fair test would give both products the same information task, record whether the answer succeeds and count every backend step and retry. Public data rarely supports that experiment across commercial services.
A low per-query figure does not settle total demand
Even if ordinary text generation approaches the old search estimate, AI can still increase total electricity use. It creates new uses, generates longer outputs and is being embedded across products. Training models and constructing accelerated computing capacity also sit outside most per-query comparisons.
The International Energy Agency estimates that all data centres consumed about 415 TWh in 2024 and projects about 945 TWh in 2030 in its base case. It describes AI as the most important growth driver alongside other digital services. Those totals include much more than search or chat and should not be divided by a guessed number of queries.
System growth and an efficient individual answer can both be true. They operate at different scales.
Carbon can reverse an energy comparison
Electricity is not carbon. Two tasks using the same watt-hours can have different emissions when one runs on a fossil-heavy grid and another uses lower-carbon electricity. Time matters on grids with changing wind, solar and demand. Annual renewable-energy matching may also produce a different accounting result from the physical generation available during the request.
That means “which emits more?” needs infrastructure data beyond the per-query energy estimate. It should also distinguish operational emissions from embodied emissions in chips and data-centre construction.
Our AI carbon footprint explainer sets out those boundaries. A simple energy ratio cannot answer all of them.
Choose the tool by the job
Use ordinary search when you want a known page, current sources or several viewpoints to inspect yourself. Use a small AI model when synthesis, rewriting or conversation adds value. Use a larger reasoning or research system when the complexity justifies it. Avoid repeating either process because the first request was vague.
Treechat estimates the operational energy of its responses from model and token use, with data-centre overhead and grid carbon assumptions stated on our methodology page. It does not claim a universal saving against Google Search because a current like-for-like public baseline is not available.
The useful conclusion is not “AI is ten searches” or “AI is now the same as search”. It is that the popular ratio has expired. Compare a defined task using contemporary evidence, and be willing to say the public data does not support a neat number.