What is the «weight» of an AI query: how much traffic do AI-powered chatbots consume?

Text query: the easiest category

To understand how AI processes text queries, it's worth first explaining the basic unit of measurement the model works with — the token. A token isn't a letter or a whole word, but a fragment of words: sometimes a whole short word, sometimes just part of one. It's into these fragments that both your query and the model's response are broken down before processing. Understanding tokens in AI explains why longer queries «cost» more computationally — regardless of how many kilobytes of data they actually represent.

Here's how a chatbot works from the network perspective: communicating with a chatbot like ChatGPT is similar in data volume to a regular instant messaging conversation — both the request and the response are text, and even a detailed response of several paragraphs weighs in at just a few kilobytes — much smaller than the average web page (the same 2.5+ MB). The bandwidth load is low, and even a slow internet connection can handle it without issue.

Interestingly, the model's «thinking» complexity during request processing doesn't always correspond to the volume of data transferred over the channel. According to estimates by the research organization Epoch AI, the average request to ChatGPT (model GPT-4o) consumes approximately 0.3 watt-hours of computing power — less than some household appliances. This means that the main «load» of such a request falls not on your Internet connection, but on the computing power of the data center where the model is running — and along with electricity, water is also consumed to cool the servers, as discussed below.

Image: medium size, noticeable difference

When it comes to image generation — for example, in Midjourney or similar services — the situation is different. The text request (prompt) itself is still small, but the result — the finished image — is a file that needs to be downloaded to your device. A typical generated image by default has a resolution of approximately 1024x1024 pixels (approximately 1 megapixel), and after quality enhancement, it can reach 2048x2048 pixels. In terms of file size, this equates to somewhere between 1 and 3 megabytes, depending on the level of detail and format — roughly equivalent to one or two typical photos from a modern smartphone.

The difference arises in scale: if a designer generates dozens of image variants in a workday, searching for the desired result, the total volume of downloaded files accumulates faster than with regular browsing of websites or social media. The question of how many megabytes image generation consumes over a full workday depends on the number of attempts: for example, 50 generated and downloaded images equates to approximately 50-150 MB, which is comparable to downloading several dozen regular web pages in a row.

Video: the heaviest category

AI-powered video generation places the heaviest load on the bandwidth. Even a short video in standard HD quality (1080p) with typical web compression weighs approximately 20-50 megabytes per minute, depending on the compression level and image detail. This means that a single short, 15-20-second generated video can consume as much bandwidth as several dozen generated images combined. If someone on your team regularly uses such tools, this should be taken into account when planning internet speed, similar to how bandwidth is allocated for video conferencing or streaming.

Power consumption and water footprint of artificial intelligence

Besides the load on the internet channel, there's another, less obvious, dimension of «weight» in AI work: the resource consumption in the data centers where the models physically run. Power consumption in AI data centers has long been a topic the industry has largely ignored, so Google's data is particularly noteworthy.

In the summer of 2025, the company released rare industry figures for its Gemini model: one request consumes approximately 0.0003 kWh of electricity — only about a tenth of what a typical lightbulb requires to operate for an hour. When converted to water consumption for cooling the AI ​​servers, one request «costs» approximately 0.002 liters — this is the answer to the question of how much water is consumed to cool the AI ​​servers per interaction with the model.

The numbers seem paltry, but the scale changes the picture: a million requests to Gemini consumes approximately 300 kWh of electricity and 2,000 liters of water, while a billion requests consume approximately 300,000 kWh and 2 million liters, respectively. These figures are directly related to the carbon footprint of artificial intelligence: the more electricity a data center consumes, the greater the carbon footprint of a single request to a neural network, depending on how clean the power source of a particular data center is.

By comparison, previous studies showed that early versions of ChatGPT consumed even more power per query, suggesting that the models' efficiency is gradually improving. Other large companies, particularly Microsoft and OpenAI, haven't yet published such detailed statistics, so Google is a rare exception, adding transparency to the discussion of the environmental footprint of a single AI query.

For the average user or small business, these figures are more interesting than practically significant — no office will notice a difference in their water bill due to the use of a chatbot. But for companies planning their own AI infrastructure (for example, by housing powerful servers in a data center), the issue of energy efficiency and equipment cooling becomes quite practical — and this is yet another argument in favor of housing equipment in a specialized data center with sophisticated cooling, rather than in a typical office server room.

How much does a typical «AI day» in the office «weigh»?

If you add up the typical workday of a team that actively uses AI tools — text queries to chatbots, dozens of generated images for marketing materials, perhaps short video clips — the total bandwidth load can significantly exceed what an office that only uses email and a browser is used to.

AI agents — services that perform a number of actions automatically, without constant human intervention (for example, independently analyzing documents or monitoring data) — are also worth considering. They can generate dozens of requests per minute without a click, meaning they can seamlessly accumulate bandwidth significantly faster than a single employee manually interacting with a chatbot. Understanding how chatbots consume bandwidth compared to such automated scenarios helps you more accurately plan your bandwidth load in advance. This doesn't mean you need some fancy plan, but it's precisely for this reason that it's important to allow for a speed buffer rather than just relying on a bare minimum.

Maxnet offers a plan with a reserve specifically for these scenarios — from moderate speeds for targeted use of AI tools to more powerful channels if image or video generation becomes a regular part of the team's workflow. The key is to consider not only the number of people in the office but also the specific content they exchange online daily.