How generative AI model training and deployment affects sustainability

At PwC, we’re studying this sustainability issue closely as we undertake our own rapid rollout and scaling of GenAI and help clients do the same. We’ve considered both ongoing use in enterprise applications as well as the impact of designing, building and training the foundation models that we and many businesses license. Our analysis also builds upon PwC’s prior work on blockchain's sustainability impact and its potential to accelerate decarbonization.

Here are some of our initial findings.

• GenAI’s biggest contribution to increasing emissions will come from its routine use. While the energy to train models has been a primary focus, for corporate users, model inferencing, the process of user prompting and model responses, will have a greater impact. (This finding aligns with recent third-party research on the topic.)
• This increase could be limited. In our analysis for PwC, these annual emissions are projected to be a fraction of business travel.
• GenAI could help drive decarbonization. GenAI-driven transformation has the potential to make operations more carbon efficient.

Assessing GenAI’s impact on emissions

To meet existing and proposed regulations, make progress on net-zero goals and provide greater stakeholder transparency, your company may be working to disclose direct (Scope 1), indirect (Scope 2) and value chain (Scope 3) emissions. Here are some GenAI dimensions to consider.

• Building and training foundation models. Foundation models (also called large language models or LLMs) are meant for general use. Vendors offer them pre-trained on vast amounts of data. Their design and build requires compute power, which requires energy. In our study, these emissions were allocated across each corporate entity that licensed the model, essentially sharing the emissions burden equally.
  
• Building and designing domain-specific models. These models, pre-trained to deliver value more quickly in specific domains (such as software development, tax or legal) also require compute power and energy to be designed and built.
  
• Customization. A leading practice for businesses is to license versions of foundation and domain-specific models and then securely embed proprietary data and expertise to enhance performance. Customization also requires compute power and energy as well.
  
• Model inferencing. Also called user prompting, this aspect covers much of the day-to-day use of foundation and domain-specific GenAI models.
  
• Enterprise applications. Many enterprise applications (such as ERP) now have GenAI capabilities built in — and their use may also involve additional carbon emissions.
  

In conducting our analysis on PwC’s initial use of GenAI, we estimated labor hours, number of processors used, power per processor and an emissions factor for metric tons of carbon dioxide equivalent.

What’s next for GenAI’s emissions impact: Transforming work, transforming tech

Our analysis was a first step. Now we’re studying how GenAI is likely to further impact emissions going forward. As GenAI makes workflows and business processes more efficient — reducing both manual activities and non-GenAI compute workloads — the drop in emissions could be significant. This is true across enterprise operations, including those activities directly related to carbon emissions. For example, we can use GenAI and analytics to help clients in prioritizing decarbonization efforts. The solution can run through millions of permutations of potential decarbonization levers, helping business stakeholders better understand those options to drive decision-making. 

GenAI and related technologies are also advancing in other ways that could affect emissions in a wide range of ways.

• More compact GenAI models. Improved architectures are resulting in models that require less compute power and produce fewer emissions.
  
• More complex GenAI applications. As companies look to automate and augment increasingly complex use cases, they’re incorporating multi-step interactions instead of single “calls.” These can increase inference costs and emissions.  
  
• GenAI proliferation. General-purpose GenAI models may fall short at highly specialized domains and tasks. The rise of specialized models, which require extra training and maintenance, might increase compute needs and related emissions. 
  
• Specialized processors. Specialized processors such as FPGAs and TPUs are taking the place of traditional GPUs for GenAI models. These processors can often support more energy-efficient compute for model training, predictions and applications. 
  
• Quantum computing. Quantum computing can enable AI systems such as the transformer networks that underlie GenAI models to perform certain complex calculations more quickly. Quantum systems may also use just a fraction of the energy required by traditional computational techniques. 

Your next steps

GenAI’s relationship to emissions is complex — with both positive and negative direct and indirect impacts — and it’s evolving quickly. To better manage GenAI’s impact, we recommend these important actions.  

Measure it: Based on our study, we believe that it’s possible — with the help of a life-cycle assessment process — to estimate the emissions impact of potential GenAI solutions and then categorize these across your emissions budget.

Be thorough: A rigorous analysis should consider the emissions impact of the pre-trained models, of your customization of these models and of everyday, application-level use.

Start early: If you conduct this assessment early — when evaluating potential GenAI models and their business case and use cases — you can make choices that align with your sustainability goals. This includes considering whether GenAI is the right solution for your business application.

Consider convergence: Look at how GenAI is converging with other technologies driving transformation and decarbonization to better understand its potential and collective impact.