ChatGPT-4 vs DeepSeek: The Environmental Battle of Generative AI

These new-generation models are much more energy-intensive than their predecessors, raising significant environmental concerns. To assert themselves in a competitive sector of the future, players like DeepSeek are claiming AI as powerful as ChatGPT-4, but with significantly reduced energy consumption.

ChatGPT-4: increased power and an exploding environmental cost

It turns out that the more powerful the models become, the greater their environmental impact. Compared to its previous version, which had 175 billion parameters, ChatGPT-4 has surpassed the 1.800 trillion parameter mark, multiplying its energy consumption by a factor of 20. Assuming an organization responded to 1 million emails per month, Greenly estimated that ChatGPT-4 would generate 7,138 tons of CO₂e over a year, split between training and running the model. This is the equivalent of 4.300 round trips by plane between Paris and New York.

This footprint increases further when we consider the impact of its daily use. According to a study conducted by Carnegie Mellon University with Hugging Face, a text query would consume as much energy as 16% of a smartphone charge. Therefore, one million text responses per month would be equivalent to 514 tCO2e in a year. Text-to-image conversion tools, like DALL-E, generate 60 times more CO2e than text on text alone.

DeepSeek: a real, more ecological alternative?

China's DeepSeek recently entered the generative AI race, promising performance equivalent to Western models and less energy-intensive applications. DeepSeek claims to be a more computationally efficient model thanks to the Mixture-of-Experts (MoE) architecture, which only activates the sub-models required for each query. DeepSeek managed to train its model using only one-tenth of the GPU hours required by Meta's Llama 3.1, despite the latter's use of more recent chips. Training used 2.000 NVIDIA H800 chips, compared to 25.000 for ChatGPT-4 and 16.000 for Llama 3.1. DeepSeek's carbon footprint is also reduced, as the H800s are less energy-intensive than other NVIDIA chips used, notably for ChatGPT.

While DeepSeek's more efficient use of resources could reduce its carbon footprint, particularly by limiting server dependency and water consumption, experts warn that these gains could be quickly offset by an exponential increase in consumption as AI models become more widespread.

What future for a more sustainable AI?

The environmental footprint of AI is becoming a central issue. The study highlights several avenues for more responsible development, including optimizing infrastructure by favoring more energy-efficient chips like Google's TPUs, legislative frameworks for environmental standards, and hosting in data centers powered by renewable energy. For example, hosting ChatGPT's training in France rather than the United States would reduce emissions by up to three times, thanks to a higher share of renewable energy in the French energy mix.

For Alexis Normand, CEO & co-founder of Greenly: "With the arrival of DeepSeek, the battle for AI models is no longer just about performance, but also about energy efficiency. The question remains: will all the generative AI giants follow this path, or will they continue to prioritize power over the environment?"

The uses of generative AI are diversifying and even becoming popular trends on social media, such as "starter packs" inspired by Studio Ghibli films. These sets, often composed of aesthetic objects and consumer products, are widely shared and imitated online. However, this consumption, although motivated by a quest for style or identity, leads to excessive consumption of energy and water. The "Starter Pack" trend alone has generated more than 700 million images. To generate one image of this type, it takes 3,5 L of water, 29 hours of light equivalent, and 1 smartphone charge.

Study methodology

Greenly's study is based on a comparative analysis of the carbon emissions generated by training and using AI models, applied to the automatic response of 1 million emails per month over a year. For ChatGPT-4, the impact is calculated based on 25.000 NVIDIA A100 GPUs, operating over a 100-day period in US data centers, at 30% of their capacity. For DeepSeek, the estimate is based on the manufacturer's announcements indicating a number of GPUs eight times lower than that of its competitors.

Access the full study (in English) here.