The price of a prompt is not a single number but a complex ecological transaction. For an image generated with NanoBanana Pro, nature pays with the extraction of gallium and copper, the evaporation of several milliliters of water, and a carbon release that requires several hours of a tree's life to remediate. For an essay from ChatGPT, the price is significantly higher—the electrical equivalent of a microwave running for over a minute and the hydrological loss of several liters of water.

Nano Banana logo on a smartphone in Suqian, Jiangsu Province, China, September 17, 2025. (AFP)

The Energetic Price of Image Generation

Unlike text generation, which is sequential and token-based, image diffusion involves predicting pixel values across a high-dimensional latent space. Research suggests that video diffusion is 30 times more costly than image generation, but image generation itself is approximately 60 times more energy-intensive than simple text requests.

For NanoBanana, the energy cost per prompt is a function of the resolution and the number of steps in the diffusion process. Google’s internal reporting estimates a median Gemini text prompt at 0.24 Wh, equivalent to watching television for less than nine seconds. However, when moving to the NanoBanana Pro model for 4K image synthesis, the energy requirements spike dramatically.

For a high-resolution image generated in 30 to 42 seconds, the consumption can range from 0.25 Wh to 0.84 Wh on highly optimized hardware, though less efficient deployments can reach as high as 2.9 Wh per image.

Water is another area of concerns when it comes to the environmental costs of AI. Google reports that it replenished 64% of its freshwater consumption in 2024, up from 18% the year prior. Yet this comes as water consumption rose 28% in just one year, from 24 to 30 billion liters. Withdrawals, including from water-stressed areas, surpassed 41 billion liters, over three-quarters of which was potable water.

For NanoBanana Pro image generation, the water cost per image is estimated between 0.27 mL and 0.91 mL. Google’s data centres alone consumed 27 billion liters of potable water in 2024. Furthermore, 28% of total withdrawals occurred in regions with medium or high-water stress, exacerbating local vulnerability even as Google celebrates replenishment programs.

Chatgpt Logo on a Mobile. (AFP)

The Price of an Essay

The generation of an essay using ChatGPT involves a fundamentally different architectural mechanism: transformer-based token prediction. Each token (approximately four characters or 0.75 words) requires a pass through the model's parameters. Consequently, the "price" of an essay is not fixed but scales with the length of the output and the complexity of the prompt.

For a 1,000-word essay (roughly 1,333 tokens), the energy consumption has been calculated at approximately 1.4 kWh based on the methodology that a 100-word response consumes. To put this in perspective, generating a single academic essay draws as much power as running 14 LED light bulbs for an entire hour.

As OpenAI transitions from GPT-4 to GPT-5 and reasoning-enabled models like o1 or o3, this energy cost is projected to rise significantly. Reasoning models create "Thinking" tokens—internal monologue that the user does not see but which requires substantial compute power. These models can use up to 100 times more energy than standard models for the same task. Producing a 1,000-token response with GPT-5 can consume up to 40 Wh, while the most intensive models like o3 or DeepSeek-R1 can exceed 33 Wh per "long prompt".

For a standard ChatGPT prompt, the water footprint is officially reported at 0.32 mL, or about one-fifteenth of a teaspoon. However, this figure is a median for simple queries. For a 1,000-word essay, the consumption rises to approximately 5.2 liters, more than ten standard 500mL water bottles.

Environmental impact of AI. (AFP)

How Many Trees Are We Losing?

To be precise, prompting an AI does not directly result in cutting down trees in the way that printing a physical book does.

But… Generating a single image using high-power models like those in the Nano Banana suite is estimated to emit roughly 2 to 5.6 grams of CO2e. For perspective, it would take a mature tree an entire year to absorb the carbon produced by roughly 4 to 10 such images.

When prompting Chatgpt, on the other hand, a medium essay has an estimated carbon cost of 10 to 50 grams CO₂e, and this eliminates 1/4 to 1 full day of one tree's oxygen production. When training the model, the estimated carbon cost is about 500 Metric Tons of CO₂e, which eliminate 25,000 to 30,000 tree's oxygen production. Estimates vary based on the model's size (GPT-4 v/s GPT-3.5) and the energy grid used by the data center.

An Equinix employee checks servers at a data center in Paris. (AFP)

Conclusion

Collectively, the "price nature pays" for our reliance on generative AI in 2025 is equivalent to the environmental footprint of one of the world's largest cities.

While industry roadmaps suggest that smart siting and grid decarbonization could cut these impacts by 73% for carbon and 86% for water, the current trajectory places the industry’s net-zero targets in significant jeopardy.

Every refined, thoughtful prompt saves resources, but the ultimate solution lies in a fundamental shift toward digital sustainability—a purposeful use of these "resource-ravenous" tools to ensure that the intelligence they generate is worth the ecological price paid by the planet.