Was this email forwarded to you? Sign up here

🗞️ Issue 34 // ⏱️ Read Time: 7 min

Hello 👋

In last week’s newsletter, we explored the often-overlooked water consumption of AI systems and data centres. Now, we turn our attention to another critical resource: Energy. ⚡️

Electricity Usage: The Power Behind the Intelligence

The issue: Compute loves energy

“Google’s greenhouse gas emissions are soaring thanks to AI”. This is one of many similar news stories we’ve seen lately. And yes, training large AI models consumes enormous amounts of electricity. As a result of the booming demand for generative AI and GPUs, AI’s electricity demand is forecast to surge, particularly in data centres worldwide. Let’s start with some numbers.

📚 A 2019 study found that training a single large AI model can emit as much carbon as five cars (5 x 🚗) in their lifetimes.

📈 From 2012 to 2018, the amount of compute used in the largest AI training runs has been doubling every 3.4 months: The energy consumption of AI is growing exponentially.

📊 Research estimates that data centres, cryptocurrencies, and AI consumed about 460 TWh of electricity worldwide in 2022, almost 2% of total global electricity demand, and this is expected to increase by 28% by 2026. 

Even though these viewpoints and analyst projections differ slightly in timelines, all point to the same conclusion: Global electricity demand will increase within the next few years. Drastically. And AI will play a big part in this increase.

Why so much power, and what does it have to do with the weather? 

Why are floods, droughts, tropical cyclones and other types of extreme weather becoming more common, and what does AI have to do with it all? 

Artificial Intelligence, especially deep learning models, requires massive computational power for both training and inference. These models, inspired by the human brain's neural networks, simultaneously perform millions or billions of mathematical operations, each consuming energy. The larger and more complex the model, the more power it needs.

The energy-intensive nature of AI directly translates to a significant carbon footprint:

• Direct Emissions: From fossil fuel-based electricity generation used to power AI systems and data centres.

• Indirect Emissions: From the production and transportation of hardware used in AI systems.

• Lifecycle Emissions: Including the environmental cost of manufacturing and disposal of AI hardware.

Data centres are often run at full power 24 hours a day, seven days a week, regardless of demand. Electricity generation often relies on fossil fuels. When fossil fuels burn they emit carbon dioxide and other gasses. And this is what causes climate change, where extreme weather is one of the consequences. If you want to know more about it, have a look at this brilliant overview by NASA.

The flip side: Energy loves compute

At the same time, AI offers unsurpassed opportunities to deal with the issues we pointed out above. That’s right - the technology is really powerful and can help reduce energy consumption by dealing with challenges such as optimising properties, revolutionising the logistics and transport sectors, and helping innovate new materials. When AI is used to solve climate challenges, there are huge benefits for the planet (and, subsequently, for us as human beings). Here are a few examples:  

• Strong and durable housing materials. Researchers at the University of Illinois Urbana-Champaign partnered with companies to discover better concrete formulas using AI. The early-stage results found the AI-powered formulas reduce the carbon footprint of the concrete by 40% while maintaining strength and durability.

• Food logistics - not too much, not too little. One of the most fundamental applications of AI lies in the development of accurate and reliable forecasts, which is perfect for supply chains. Axfood, a large Scandinavian food retailer, is using these forecasts to better understand the demand for different food products, which leads to a more efficient logistics chain. The use of AI leads to reduced food waste and optimised transport routes. 

Task-Specific Power Consumption

In a first-of-its-kind research effort, Hugging Face and Carnegie Mellon University have provided insights into the energy consumption and carbon emissions of specific AI tasks. The study compares the energy requirements for training, fine-tuning, and inference across various AI models.

Using large generative models for specific tasks (e.g., text classification or question answering) can consume up to 15 times more energy than models fine-tuned for that specific task. So, if you want to translate a text, the emissions will be higher if you use ChatGPT or Claude, instead of using a model focused on translation, like DeepL or Google Translate.

While training large AI models is energy-intensive, most of their carbon footprint comes from their actual use due to the frequency of operations. For very popular models like ChatGPT, usage emissions could exceed training emissions in just a couple of weeks.

The Emissions of the Key Industry Players

Some of the big players have the biggest emissions. They have also been criticised for the growing gap between the real world and market-adjusted carbon footprints, and for using different methods to hide their actual energy consumption, so read the numbers below with a pinch of salt 🤏🏾

Addressing the Challenge

There are plenty of ways to solve the AI-related energy challenges we mentioned above. So, where to start?

1. Create a standard of measurement: A main problem to tackle in reducing AI’s climate impact is to quantify its energy consumption and carbon emission, and to make this information transparent.

2. Invest in renewable energy sources for data centres. 

3. Consider the full lifecycle environmental impact when designing and deploying AI models

4. Develop more energy-efficient algorithms and hardware (e.g., Google's TPUs designed for AI workloads, NVIDIA's energy-efficient GPU architectures)

5. Carbon offset programs for AI companies is something that tends to be suggested as a solution. Unfortunately, these types of programs have proven to be flawed in most cases.

Some of the biggest environmental challenges, including climate change, can be solved through AI. It can shift current practices and catalyse sustainable development. The result could be a future where products and systems are radically improved. Yes, we know it seems complex, and it is. But ultimately, isn’t that what we would like AI to do? Solve real challenges so that we can continue to live our lives and focus on solving other, sometimes smaller, and often better, problems.  

Until next time.
On behalf of Team Lumiera

Emma - Business Strategist
Allegra - Data Specialist

Lumiera has gathered the brightest people from the technology and policy sectors to give you top-quality advice so you can navigate the new AI Era.

Follow the carefully curated Lumiera podcast playlist to stay informed and challenged on all things AI.