AI Sustainability: Greenly’s Findings on DeepSeek & ChatGPT
A study conducted by Greenly, a company specialising in enterprise carbon accounting, has compared the environmental impact of two prominent AI platforms: OpenAI's ChatGPT-4 and DeepSeek.
The research finds pressing concerns about climate impact and sustainability, as these next-generation AI models continue to expand.
Gen AI, particularly large language models (LLMs), requires substantial computing resources – and the training and operation of these systems demands significant electricity and water consumption, resulting in carbon emissions and contributing to electronic waste accumulation.
Electronic waste refers to discarded electrical or electronic devices that often contain hazardous materials requiring special handling and recycling.
Additionally, the environmental impact becomes more apparent in models such as ChatGPT-4, which contains 1.8 trillion parameters, twenty times more than its predecessor. As a result, the research demonstrates that as model complexity increases, the associated climate footprint expands too.
The bigger picture on AI’s environmental impact
In a hypothetical business scenario where an organisation employs ChatGPT-4 to respond to one million emails monthly, Greenly found that AI could generate 7,138 tonnes of carbon dioxide equivalent (tCO₂e) annually – comparable to 4,300 round-trip flights between Paris and New York.
Additionally, research from Carnegie Mellon University and Hugging Face, indicates that a single text-based request consumes energy equivalent to charging a smartphone to 16%.
Furthermore, the production of AI hardware components, including processors, graphics processing units (GPUs) and AI chips, necessitates the mining of rare earth minerals, which can result in environmental degradation such as soil erosion and pollution.
DeepSeek vs ChatPGT
However, DeepSeek presents a potentially more environmentally sustainable approach to AI development.
This Gen AI model utilises a Mixture-of-Experts (MoE) architecture, which activates only the relevant sub-models for each specific task – thereby reducing the computing power requirements.
Furthermore, DeepSeek was trained using 2,000 Nvidia H800 chips, compared to the reported 25,000 for ChatGPT-4 and 16,000 for Meta's Llama 3.1.
Nvidia H800 chips are processors designed specifically for AI workloads that offer high performance with optimised energy consumption.
These H800 chips also consume less energy during operation.
Overall, DeepSeek requires one-tenth of the GPU hours (the time that GPUs operate at full capacity during model training),used by Meta's model, resulting in a reduced carbon footprint, decreased server usage and lower water requirements for cooling systems.
Yet despite these efficiency gains, Greenly cautions that such improvements might be short-lived as global AI usage continues to increase, potentially offsetting efficiency improvements through sheer volume of use.
“DeepSeek's emergence has put energy efficiency at the heart of the battle between AI models,” says Alexis Normand, CEO and Co-Founder of Greenly.
“But it remains to be seen if other players will follow this path, or continue to prioritise raw processing power at the expense of the environment.”
Sustainable AI progression
As AI adoption scales, regulatory bodies are implementing ethical and sustainability boundaries.
The EU’s recently introduced AI Act is such a step in this direction.
“AI has the potential to change the way we work and live and promises enormous benefits for citizens, our society and the European economy,” says Margrethe Vestager, Executive Vice President for a Europe Fit for the Digital Age.
“The European approach to technology puts people first and ensures that everyone's rights are preserved. With the AI Act, the EU has taken an important step to ensure that AI technology uptake respects EU rules in Europe.”
However, despite the environmental challenges, there are indications of progress in the sector.
AI is being applied to accelerate decarbonisation, improve energy efficiency and support sustainable development goals – and if deployed strategically, it could contribute to reducing global emissions by 1.5-4% by 2030, according to the study.
Furthermore, energy-efficient design, renewable-powered data centres, edge computing and open-source model reuse are among the strategies being considered to reduce the sector's environmental impact.
A 2020 study published in the scientific journal Nature found that training a single LLM can generate emissions equivalent to approximately 300,000 kilograms of carbon dioxide.
As a result, several approaches are being developed to address these concerns.
Now, AI is increasingly being used to optimise energy consumption across various industries, potentially offsetting its own footprint through efficiency gains elsewhere.
Additionally, companies are exploring the use of renewable energy sources for data centres and more efficient hardware designs.
“This act marks a major milestone in Europe's leadership in trustworthy AI,” comments Thierry Breton, Commissioner for Internal Market.
“With the entry into force of the AI Act, European democracy has delivered an effective, proportionate and world-first framework for AI, tackling risks and serving as a launchpad for European AI startups.”
Explore the latest edition of AI Magazine and be part of the conversation at our global conference series, Tech & AI LIVE.
Discover all our upcoming events and secure your tickets today.
AI Magazine is a BizClik brand
