Skip to main content Skip to footer


Making generative AI green


October 18, 2023

Generative AI applications like ChatGPT, DALL-E and Stable Diffusion have captured the public’s imagination and are among the fastest-growing consumer applications in history. And they’re proving to be remarkable tools in addressing some of the most complex global challenges: Take Insilico Medicine. The multinational biotechnology company used generative AI to discover a COVID-19 therapeutic drug that entered clinical trials in China. Generative AI has huge potential to drive the ESG and sustainability agenda of businesses by improving climate action, driving responsible production and protecting sensitive data. That’s the upside.

The downside: On top of well-known social implications, such as fears around the impact on jobs and the ability to generate disinformation, developing, training and operating GenAI can use a tremendous amount of energy and create a significant carbon footprint. The Generative Pre-trained Transformer 3 (GPT-3) used in the drug trials alone may have required up to 1.287 gigawatt hours of electricity to train, which is equivalent to the annual energy consumption of 120 US homes. The process also likely generated 502 tons of carbon emissions, which is comparable to the annual emissions of 110 US cars.

Researchers at Facebook estimate that everyday usage of large language models burns an even larger carbon footprint. If current trends continue, machine learning systems could consume nearly all of the world’s energy production by 2040.

The greening of generative AI is the next frontier to retain the huge positive impact it can have in driving the broader sustainability agenda.

Embed sustainability. Digital to the core.
Embed sustainability. Digital to the core.

What steps can companies take to enhance energy efficiency and reduce carbon emissions throughout the AI development and operations cycle?

We suggest three:

  1. Minimize the computational cost of generative AI models by using more efficient algorithms, architectures and hardware – As a first step, companies can develop or adopt an accurate energy estimation approach that is non-invasive, vendor-agnostic and has comprehensive hardware coverage. This can help reduce the time and energy required for training and deployment, resulting in reduced emissions and costs.

    More energy-efficient approaches to AI don’t always require significant compromises to be made on the quality of AI models. For example, researchers from Google and University of California, Berkeley have shown that the carbon footprint of large language models can be reduced by 100 to 1,000 times with the appropriate choice of algorithms, customized hardware and energy efficient cloud data centers. Researchers at Accenture Labs found that training an AI model on 70% of the full dataset reduced its accuracy by less than 1% but cut energy consumption by a staggering 47%.

  2. Adapt pre-trained models to new tasks or domains instead of collecting new data or training new models from scratch – Where possible, businesses can fine-tune pre-trained large models for private use at a fraction of the cost and emissions required to build a new model. By reducing the size and complexity of these models, companies can shrink their energy consumption and carbon footprint. Stability AI has launched its open-source StabilityLM suite of language models that can generate text and code and can be freely adapted for commercial and research purposes. One experiment by Accenture shows that training a much smaller “student” model, just 6% in size of the original “teacher” model, achieved the same level of accuracy (99%) but consumed 2.7 times less energy.

  3. Apply generative AI to accelerate the energy transition by optimizing renewable energy generation, storage and distribution – Applying generative AI to the energy sector can help reduce greenhouse gases, increase energy efficiency, create new low-carbon innovations and solutions. How exactly? Generative AI can be used to predict energy demand or boost the production of renewable energy by optimizing design (like locating solar panels or designing blades for wind turbines) based on weather patterns. However, not all applications require generative AI. For many, a simpler diagnostic AI technique will be greener and more appropriate than a generative AI approach.

How can companies get started?

Short-term (6 months–1 year)

  • Conduct a rapid assessment of the Green AI maturity of projects and make recommendations to improve their maturity.
  • Guide AI engineers and managers toward the best practices to build greener AI models
  • Compare and contrast key model and training choices for their energy implications

Medium-term (1–2 years)

  • Help ensure clear governance structures that define principles, practices and metrics related to generative AI
  • Establish measurement criteria to start reporting the benefits
  • Optimize AI workloads in the cloud to reduce emissions while complying with constraints

Long-term (3–5 years)

  • Create experiential learning to help developers understand the energy implications of training generative AI models
  • Participate in ecosystems to share data and best practices
  • Certify AI models for energy efficiency, carbon emissions and water impact using agreed standards and specifications for the sustainable use of generative AI across industries

As generative AI becomes more prevalent, it is crucial for companies and communities to have a reliable network of individuals, standards, tools, and practices. The Green Software Foundation (GSF) exemplifies how collaboration can establish such a network.

The GSF, founded by Accenture, Microsoft, GitHub, and Thoughtworks, is dedicated to supporting the Information and Communications Technology (ICT) sector in reducing its greenhouse gas emissions by 45% by 2030. The foundation focuses on various initiatives, including assessing and reporting the carbon footprint of applications, discovering energy-saving techniques for AI, and developing tools and training for green software engineering. One of their notable contributions is the Software Carbon Intensity technical specification, which provides guidance on measuring total carbon emissions. By addressing the issues of ambiguous metrics and inconsistent standards across industries, the GSF aims to find solutions for these challenges.

As generative AI continues to evolve, it presents both new opportunities and risks for sustainability. While we rely on this innovative technology to address the world's most pressing challenges, it is crucial that we proactively implement measures to mitigate any potential drawbacks.

Embedding sustainability

This blog is part of a series discussing how leaders can embed sustainability into different aspects of their organizations to create value and impact. The other topics are:


  • Sanjay Podder, Global Lead – Technology Sustainability Innovation