AI Energy Consumption is Alarming!!

Ever since its launch, I have used ChatGPT quite frequently. The scope of this use varies across domains, though we won’t get into that here. What we will focus on instead is the environmental impact of this use. And up until recently, just like you, I was quite unaware of how I was killing planet Earth simply by using AI and the related energy consumption.

You see, this wasn’t pure ignorance. My lack of knowledge stemmed from the fact that not many major AI players had made their “per prompt” energy consumption public. Sam Altman, CEO of OpenAI, recently changed that through his blog post. In his blog, Altman cleared that an average ChatGPT query uses about 0.34 watt-hours of energy.

Google has now followed suit, with its recent report titled “Measuring the Environmental Impact of AI Inference.” As the name suggests, the report basically highlights the energy consumed by Google’s Gemini AI upon every prompt that is given to it.

What is contained within the report is soothing yet alarming. How? Read on.

Google Busts the Myths on AI Energy Consumption

Soothing – because Google’s report straight-up disproves the various myths circulating around AI-energy consumption. Like the popular one, that every prompt on Google’s Gemini AI is equivalent to driving a car for 900 miles. Though always unbelievable, it is much better to have the source disprove it with actual figures.

So what are the figures? Google says that a median text prompt given to Gemini, basically one in the middle of this energy demand range, uses 0.10 Wh of energy. For context, this is the same as watching TV for less than nine seconds.

In addition, such a median prompt also emits 0.02g of CO2 and consumes 0.12 mL of water.

The numbers are much more efficient than those shared by Altman for ChatGPT. In addition, the report brings a much-needed respite from the lack of information around the topic. With solid numbers linked to AI-energy usage, that too for each prompt, people and organisations conscious of the environmental impact can measure their carbon footprint. Much like a dark cloud of guilt has lifted.

Not just yet.

AI Energy Consumption: The Volume Killer

Now for the alarming news. Know that Gemini AI hit the 450 million monthly active users (MAU) mark last month. The Daily Active Users (DAU) were estimated to be around 40 to 45 million. These numbers must have only gone up ever since.

Considering these MAUs give a single prompt to Gemini in a month, that is 4.5million Wh of energy used per month.

And if we were to calculate this for the DAUs, assuming they give just 1 prompt a day, this figure scales to 135million Wh of energy every month. In the TV example above, this number is equivalent to running a TV for 3.37 million hours.

Under similar conditions, the daily active users of Gemini tend to produce 27,000kg of CO2 and consume 162,000L of water every month. That is three big swimming pools full of water.

Now ask yourself, how many prompts do you give to Gemini (or your AI tool) every day? Whatever your answer is, I can safely assume it is more than 1 prompt if you are a daily user.

Note that the Google report mentions, the figures mentioned above are a result of “an optimistic scenario at best and substantially underestimates the real operational footprint of AI.” Which basically means that the actual environmental impact is more (and I expect it to be much more) than what is being portrayed here.

But wait, there is another shocker yet to come.

In case in the midst of such big numbers you missed it, know this – this energy consumption is only for text prompts

For prompts around images or videos, this environmental impact will be through the roof in comparison.

Yet Google insists that AI, or more specifically – Gemini, is quite energy efficient.

Here is why.

Google’s Green efforts for AI

For Google, the entire purpose of coming up with this report was to highlight its ongoing efforts towards greener and efficient AI. Through several initiatives it currently employs, Google has managed to bring down the environmental impact of Gemini. At least that is what it claims.

As per the tech giant, these energy requirements and carbon footprint of its AI model have dropped by 33x and 44x over the past 12 months. All this, while the model has managed to deliver higher quality responses than before.

This was possible through a two-pronged approach – a reduction in the energy emissions of Google data centers, and its environmental efforts towards net carbon neutrality and water replenishment. Other than this, the company also credits its advancements in its AI model for the cleaner environmental impact. Here are some of them:

• More efficient model architectures like the Transformer model architecture developed by Google researchers, claiming a 10-100x efficiency boost over the previous state-of-the-art architectures for language modeling.
• Continuous refinement of its algorithms with methods like Accurate Quantized Training (AQT).
• Constant improvement of AI model delivery for responsiveness and efficiency through technologies like speculative decoding.
• Custom-built hardware, like Google’s TPUs built from the ground up to maximize performance per watt.
• Minimised TPU idling through an optimised serving stack.

The Larger (and Scary) AI Picture

Honestly, hats off to Google for coming up with such solid energy consumption figures for its star product at the time. With this, Google has also set a precedent for the rest of the industry to come clean or simply be more transparent in their operations. As the world moves blindingly fast towards AI adoption, such challenges left unaddressed today may just be impossible to handle later.

However, no matter how much Google spins the numbers towards its environmental-conscious efforts, the fact of the matter remains – AI use is having a gargantuan impact on Earth. What’s even scarier is that if Google considers Gemini as one of the most energy-efficient models out there, how much of an impact are other AI tools like Grok, or DeepSeek having?

The real answer can only be imagined for now. At least until other AI firms step up and reveal similar figures. In case they don’t, AI adoption may already be a losing battle for the environment.