18 hours ago
Why your AI questions are a power and water drain
Sam Hawley: How often do you use AI? It's becoming part of our everyday lives. But when you pump in a question into something like ChatGPT, do you ever think about the energy it uses? Today, Gordon Noble from the Institute of Sustainable Futures at UTS on the data centres driving AI and what they're doing to the climate. I'm Sam Hawley on Gadigal land in Sydney. This is ABC News Daily. So Gordon, Googling something or typing a question into ChatGPT, it is so easy and simple for us and very, very useful, of course. But we don't really stop to think about where all that information is coming from, do we?
Gordon Noble: Yeah, that's right. And I think this is crept up on us. When you do a search now with all these AI tools, ChatGPT being one of them, they can be 10 times more energy consuming than, for instance, doing a Google search. And what sits behind that is this massive investment that we've seen in data centres globally. So it's absolutely exploded.
Sam Hawley: Yeah, okay. So we're not thinking about these data centres when we're typing in what we want to know from ChatGPT. And most of us would never have seen one or been in one. What do they look like? Where are they?
Gordon Noble: Yeah, it's a really good question. So data centres, just to put it in visual terms, so the average size of what they call these hyperscale data centres, so they're around about 10,000 square feet. To give you an indication, a Bunnings store is about 8,000 square feet. So they're kind of just big sheds, right? Huge. But what we're seeing now is that we're moving to not just these Bunnings-like sheds, if you like, but we're moving to these massive million square data centres, almost campuses of data centres. Around the world, there are around about 1,100 of these hyperscale data centres. In Sydney, for instance, Sydney is a big centre for data centres here in Australia. We have over 85 data centres. One of the reasons Sydney is such an attractive place for data centres is we have 12 submarine cables that come out of Sydney and basically connect us to the rest of the world. So data centres globally are now around about 1.5% of global energy consumption. The question is what's going to happen in the future?
Sam Hawley: Okay, so Gordon, let's delve a bit further into how these data centres actually work, because while they're enabling us to inform ourselves at lightning speed, they're also using a huge amount of power, massive amount.
Gordon Noble: So roughly at the moment, global energy consumption coming out of data centres is around 1.5% of all the global electricity. The issue is that data centres are highly concentrated. So it's in places in the world, the US, parts of Europe, Ireland is a massive data centre hub, where they're actually causing strain on the energy grid because of how much the energy growth has been. So to give you an idea in Australia, so a research report from Morgan and Stanley, they were projecting that roughly at the moment, data centre energy consumption from the grid is around about 5% of all of our energy consumption. But what they're projecting is that this could grow to between 8% and 15% of all of our electricity consumption here in Australia, depending on some of the decisions that are made as in how much we use AI tools. So what the International Energy Agency is now saying is that by 2030, the energy consumption from data centres will be the size of Japan. So we're talking massive amounts of increase in energy consumption. That's placing strains on the grid, but it's also placing a shift in terms of how the energies come from. So for instance, in the US, we're seeing providers like Microsoft, who are big data centre operators because of the tools that they've got. They're looking to shift to nuclear. And one of the things they're looking to do is to reopen Three Mile Island, which is the nuclear plant that had been mothballed to basically take all that energy from a reopened Three Mile Island. So lots and lots of decisions as a result of this.
Sam Hawley: Why is it, Gordon, that AI takes so much more energy than just Googling?
Gordon Noble: These large language models are effectively trained to look at the whole of the internet, right? So when they're developing these models, they're actually looking at everything in the internet. And then when we ask it to do something, it's churning away from all that work that it's done. Lots and lots of different applications, but I think that common common thread is that it's aggregating across a lot of data rather than just that single data search where it goes to a single source.
Sam Hawley: Do we have a sense already about the sort of strain that it's putting on electricity grids in Australia?
Gordon Noble: Yeah, so at the moment, that's one of the questions. And we don't really have, I think, a good picture of the national demand, right? So the issue at the moment is a lot of the training of these AI tools, they've taken place in the US principally. So they haven't yet really been here in Australia. So that's going to be one of the questions as we increase the size of our data centre industry. Where is it going to start to have implications in terms of energy demand? Will it be, for instance, in Sydney, which is really our data centre capital? What would the impact, if you like, in terms of energy consumption in New South Wales in particular? Other states have the same issue, but because Sydney really is that capital of data centres in Australia, that's where some of the key issues will emerge.
Sam Hawley: And Gordon, every time we use an AI site like ChatGPT, it uses a lot of water, doesn't it?
Gordon Noble: Yeah, look, this is a real sleeper issue, and it's one that we're very concerned with. There's recent research, for instance, that since 2022, all the new data centres that have been developed, you know, two thirds of them are in areas where there's water stress. So it's becoming a big issue. But the way to think about data centres is that they're like the human body, they like to be kept cool, operate efficiently. And one of the ways that that happens is using water. So they consume literally billions of litres of water. The issue as we go forward is how do we actually, in Australia, build a data centre industry that is sustainable, given that we're an arid continent, given that we're going to have challenges from our climate in terms of water. At the moment, one of the opportunities is that both in Sydney and Melbourne, where data centres are going to likely be established, is we actually have surplus water in the form of recycled water. We tip 97% of our recycled water out in Sydney and Melbourne, we actually tip it out into our oceans and bays. So this is an asset, for instance, that could be used if we're smart enough to say, well, how can we actually build, say pipelines of recycled water to use this water, so we're not actually putting stress, environmental stress, on our rivers and creeks and streams, etc. So there's opportunities around this that we could solve.
Sam Hawley: This is all making me start to feel rather bad for using ChatGPT for that recipe last night. I must go back to the old book, the cookbook lives on. Anyway, just tell me about emissions then, because we're meant to be bringing them down and I'm gathering this is not helping. Even the tech companies admit that, don't they?
Gordon Noble: Yeah, this is what happened last year. So I think the surprise to the market is we started having the sustainability reports of the big tech companies and they all started to actually reveal how much their emissions had started to increase over the last four or five years. Each one of them, there are different increases in emissions depending on the way they've structured their operations, whether they build data centres, whether they outsource them, etc. But the picture that was emerging was a very consistent increase in their energy consumption. I think that really woke up a lot of the market in terms of, yeah, this is actually an energy intensive industry. Up until at that time, I think there was a little bit of a lack of understanding of how much energy data centres were creating because it wasn't really being aggregated in a single spot. So as we've been getting what we call these climate related financial disclosures and companies are starting to report on what we call the scope one, two and three emissions, we're starting to get a bigger picture. We're expecting to get more reporting in the next month or so. So what we'll start to see is what's happened since 2024 and 2025 and then we'll start to really have a good understanding of where things go forward. But what the clear picture at the moment is emissions arising in the big tech companies driven by their investments in AI.
Sam Hawley: Gosh, all right. So Gordon, how worried do you think we should be then about this massive energy use and who should actually be taking responsibility for this?
Gordon Noble: You know, I think it's as you mentioned, there's a lot of potential benefits around AI tools. You know, we can use these, for instance, for whether it's the recipe, a lot of environmental applications, a lot of benefits here if we get this right. At the moment from an Australian perspective, what really we haven't seen is a national approach being taken on this. We have, I think, in the Australian government an approach to communications that goes back to the, you know, the days in the early Federation we had a postmaster general. At the moment we need to start to think of, you know, the digital economy as actually moving across a range of different portfolios in the federal government, for instance. So we need a strategy around this to recognise that this has potentially got massive benefits, but also we really need to manage that. What we're seeing in other jurisdictions, for instance, Singapore, have gone down the pathway of establishing a green data centre roadmap. We need something like that in Australia.
Sam Hawley: But what do you think without a new approach, can we keep pumping these questions into ChatGPT and still reach our environmental goals? And can our energy system actually cope with demand that is just going to keep growing?
Gordon Noble: This is the big question. So the reality is if we do have at the higher end of expectations of the growth of AI, the energy that's demanded here just in Australia will actually crowd out other investments that we're making in renewable energy. So whilst we're making progress in decarbonising our grid, you know, there's an assumption that's based on, you know, a certain level of growth of energy demand. If that increases significantly, you start to put pressure on how much we can actually invest in more renewables, in more solar, for instance, more battery technology. It starts to then have that question, do we keep coal-fired power stations longer than we need? So I think there's a broader set of issues that we really need to get our heads around.
Sam Hawley: Gordon Noble is a Research Director with the Institute for Sustainable Futures at the University of Technology, Sydney. This episode was produced by Sydney Pead and Adair Sheppard. Audio production by Cinnamon Nippard. Our supervising producer is David Coady. I'm Sam Hawley. I'm going to take some leave from now for a couple of weeks. Sydney Pead will be with you from tomorrow. Thanks for listening.
