AI uses a lot more energy than you might initially think, but exactly how much, and what can AI companies do to mitigate their impact on the environment?
As the world sees more and more effects of climate change, being environmentally conscious is something that every company needs to take into account. AI also uses its fair share of resources; clean, fresh water and electricity consumption has skyrocketed in order to keep the data centers running. But AI is here to stay, so finding a solution to the energy consumption problem is essential. How can AI be more sustainable while still keeping up with its usage demands?
According to a study from researchers from the University of Colorado Riverside and the University of Texas Arlington published in October of 2023, ChatGPT-3 needs to consume one 500ml bottle of fresh water per 10-50 responses, and ChatGPT-4 likely consumes even more. As freshwater scarcity is one of the looming threats as the world moves further into the 2020s (about 1.1 billion people worldwide don’t have access to freshwater), realizing that the total demand of global AI for freshwater could be as much as 6.6 billion cubic meters in 2027 can be concerning. But what are these data centers doing with all this freshwater?
The study categorizes water usage into three areas: scopes 1, 2, and 3. Scope-1 water usage encompasses on-site water used in cooling towers and air cooling for servers. Servers are extremely energy-intensive, especially when they use high-end GPUs, and they can easily overheat. Scope-2 water usage comes from off-site water used to generate electricity (which we’ll cover in the next section). Finally, scope 3 is water used in the processes for AI chip and server manufacturing, and this category uses an enormous amount of water.
All three of these usage categories require freshwater, and wafer fabrication in scope-3 usage needs ultrapure water. Wafers are thin discs of silicon, and the water used to keep them clean during fabrication is so pure that it even has salt ions filtered out. Water used for cooling also needs to be fresh so it doesn’t corrode the machinery.
Water recycling would help this incredible drain on resources, but often efforts are not efficient enough. The study notes that much of the data around scope-3 water usage isn’t available, but, in Singapore, wafer plants recycle about 45% of their water while semiconductor plants only recycle about 23%. While there’s no getting around that AI data centers need water to keep them cool, more effort should be put into using water sustainably and recycling as much as possible.
Researchers have estimated that training a GPT-3 model emits 305% more carbon dioxide than a passenger jet flight from San Francisco to New York, and another study found that while the average American produces 36,156lbs of carbon dioxide a year, training one model with an NLP pipeline with experimentation produces 78,468lbs. However, much data about AI training and data centers isn’t public, so it’s difficult to know exactly how much CO2 is being produced. The lack of data sounds off certain alarms that training AI models may not be as sustainable as the companies training them would like us to believe.
AI is consuming a significant amount of electricity as well. According to the International Energy Agency, data centers use between 1 and 1.5% of the global electricity consumption, and AI is taking up a significant portion of that. Nailing down exactly how much electricity an AI model uses per response is tricky, but a study reports that in 2020, Chat GPT-3, when training at a Microsoft data center, used 1,287 megawatt hours of electricity. In comparison, the average home in the UK uses about 3.5 megawatt hours per year, as of September 2023.
Unfortunately, most companies haven’t had a good track record using renewable energy (hydro, wind, solar, etc.). In 2017, Microsoft sourced 32% of their energy from renewable sources while Amazon-AWS only sourced 17%. Google was the notable outlier with 56% of their energy consumption coming from renewable sources. But it looks like this is going to change.
Mircosoft has the right idea. They have recently announced “the single largest corporate power purchase agreement ever signed,” in which they have agreed to invest more than $10 billion into the development of 10.5 gigawatts of renewable energy with Brookfield Asset Management between 2026 and 2030. In comparison, 10.5 gigawatts is about three times more than the 3.5 gigawatts consumed by data centers in Virginia, which is the largest data center market in the world.
Google is also determined to start operating fully on carbon-free energy by 2030. According to Google, they “reached 64% carbon-free energy globally on an hourly basis.” This goal was announced in 2020, presumably after they realized how much electricity it takes to run and train AI models.
While Amazon claims to have been the largest corporate buyer of renewable energy in 2023, there’s more that they could be doing. Their lack of transparency around their other sustainable efforts has been met with criticism and discussions about how truly sustainable Amazon can be. After all, same-day delivery has changed the way consumers shop and buy (I myself got way too excited about next-day delivery for a new rice cooker), but it’s now become the expected standard. Unfortunately, a supply chain so robust will always have a darker side. Because of a lack of strict return policies, many of the items returned to Amazon sellers are disposed of. In 2021, $761 billion of merchandise was returned in the US retail industry, both online and in-store, and in 2019, an Amazon facility in Europe was reported to have sent 293,000 items to the landfill. While Amazon has made significant strides in the renewable energy sector, it’s clear they have a long way to go in other areas.
We build our own LLMs for our platform, and they’re hosted on OpenAI, so we were quite excited to hear about Microsoft’s new energy deal! Plus, when a software is high-quality, it requires less testing and debugging. By investing in a robust software, you reduce the amount you need to test with self-healing tests that don’t break and need to be rewritten or rerun. If you’re curious about how you can run fewer tests and still retain full coverage and peace of mind, set up a time to chat with us!
