The curious climate impact of AI

AI’s paradoxical role in sustainability

You’ve probably seen the headlines that suggest AI is extremely energy-intensive and therefore detrimental to the planet. When The Guardian published an article claiming that data centre emissions were a worrying 7.62x higher than reported by tech companies, it made me question whether my work as a data engineer was as environmentally responsible as I had hoped. But, after diving into the data, I found something interesting.

It’s important, first, to understand that different types of data and AI work have varying CO2 emissions. The lowest emissions typically come from databases and business intelligence (BI) tasks, followed by machine learning (ML) and predictive modelling. Optimisation work tends to be more energy-intensive, and generative AI—particularly model training—is at the high end of the scale in terms of energy consumption. Asking an LLM like ChatGPT or Claude something rather than using a standard Google search consumes between 10 to 25 times more energy per query.

The carbon math

So, to investigate, I analysed our own data—for context, we’re a team of ~40 data scientists and engineers using cloud provider services daily. I collated the CO2 emissions data from the various cloud providers we partner with to develop and run our data and AI products, which falls under Scope 3 emissions in the greenhouse gas protocol. The greenhouse gas protocol classifies emissions into three scopes: Scope 1 (direct emissions from owned sources), Scope 2 (indirect emissions from purchased energy) and Scope 3 (all other indirect emissions throughout the value chain). Our total emissions from cloud computing came to less than 5 tonnes of CO2 for the year. To put that into perspective: although we rarely travel by car or plane (we’re mostly remote), our emissions from travel alone were more than double that.

Alongside the surprisingly (and, frankly, reassuringly) low emissions, I was surprised by the extent to which AI can help reduce emissions. In our own business, we recently developed a route optimisation solution for GXO, a logistics company, that reduces CO2 output by 720 tonnes annually—144 times our 5 ton annual cloud computing footprint—by minimising the distance travelled by their delivery drivers. In another case, we helped a business in the airline industry to tackle food waste; another important source of CO2 emissions. By implementing predictive modelling, we helped find opportunities to reduce food waste by 47%, ensuring better alignment between supply and actual passenger demand.

These examples reveal AI's paradoxical role in sustainability: when created in a considered way, the emissions produced by these solutions are dramatically outweighed by the emissions they prevent. Even if cloud providers have underreported their emissions impact as suggested by recent media, the net environmental benefit remains overwhelmingly positive.

That's not to say we ignore our cloud CO2 emissions. We're actively working to reduce them, but understanding the broader impact of our work has allowed me to reframe how I view sustainability in AI. Here are some practical approaches to developing low-carbon AI:

What I’ve learned about sustainable cloud computing

1. Ensure use cases are valuable and developed products will be used. Assess environmental benefits alongside potential emissions resulting from development and use.
2. Be mindful of the CO2 impact of all your tasks, no matter how big or small. Every action counts.
3. Gather and monitor data on your emissions. Track your CO2 impact and aim to reduce emissions by targeting low-hanging fruits.
4. For fellow developers, here are some specific strategies to reduce your environmental footprint:
   
   
   • Optimise code efficiency: Write efficient code to reduce unnecessary computational load. Use algorithms and data structures that minimise processing power requirements.
   • Be strategic about when you run compute jobs: Use data from carbon intensity forecasts to figure out the best time to run your pipelines. For example, you can refer to data from the Carbon Intensity API, which provides insights specific to the UK.
   • Adopt serverless architectures: Use serverless computing where possible, ensuring resources are only consumed when absolutely necessary. Also, decouple storage from compute to reduce excess usage.
   • Optimise data storage: Regularly clean up unused or obsolete data. Reducing data storage cuts energy consumption significantly.
   • Leverage cloud provider energy optimisations: Many cloud providers offer tools and suggestions for energy-efficient computing, such as using low-carbon regions or identifying underutilised resources that can be turned off.

AI—done right—can be a sustainability multiplier

Sustainability in AI extends far beyond minimising computational emissions. The most significant impact comes from designing AI solutions that address environmental challenges across industries. Our experience demonstrates that thoughtfully applied AI becomes a powerful sustainability multiplier—delivering environmental benefits that exceed its carbon costs many times over.

The key takeaway? AI becomes a powerful tool for sustainability when used thoughtfully. The right AI solutions don't just offset their own emissions—they systematically reduce carbon footprints across entire value chains and industries.

This reframing doesn't minimise the importance of efficient computing—it contextualises it within a broader environmental strategy. As we continue to develop AI systems, the question isn't simply "How much energy does this consume?" but rather "What environmental impact might this enable?". By applying this dual perspective, we can make better use of AI as a force for environmental good while continuing to drive improvements in its energy efficiency.

The climate challenges we face require both innovative thinking and practical solutions. Perhaps counterintuitively, thoughtful application of artificial intelligence might offer one of our most promising paths forward. If you’re curious, get in touch with us for a chat.