Can We Have AI and a Liveable Planet?
Boris Gamazaychikov, who measured AI's footprint inside Salesforce and co-created the AI Energy Score, a star rating for AI models, on what is true and what is exaggerated in claims about AI's water, energy and carbon use. He now works at Sustainable AI Group, which estimates the environmental impact of closed AI models and pushes AI companies to be more open. He discusses data centers and local communities, electricity bills, air pollution, and why smaller models can often do the same job as frontier models for far less energy.
Transcript
By 2030, the world's data centers will burn through more electricity than the country of Japan is currently using today.
I feel like we're just constantly justifying the energy use because the potential achievement is just so great.
Sam Altman says claims around AI's water usage are "totally fake." Yet some people say one chatbot conversation uses up an entire bottle of water.
I think it's a question that the AI industry is not caught up with.
It always comes back to us to ask the questions with the smallest amount of information to go on.
Our guest today says a lot of these numbers are wrong and that AI energy story isn't so straightforward. So is AI actually cooking the planet or could it help save it?
It's not like a Hummer versus a Prius. It's like a jet plane versus a bicycle. There hasn't really been a good way to understand that difference.
Boris Gamazaychikov measured AI's footprint inside Salesforce and then co-created its first energy label, something like the sticker that you would find on your fridge except for chatbots.
So how do the various AI models rank on energy usage?
Sure, there's some inconsistency because there's a information vacuum due to these very AI companies.
Today with Boris, we will look at the real numbers, the fake ones, and whether super intelligence is worth the energy bill.
This isn't that hard for those companies to provide more transparency, so we feel confident it can happen.
All right, Boris, so we just want to set the record straight today. We've heard a lot of numbers being thrown around, especially when it comes to water usage with regard to data centers and how much they utilize with just one ChatGPT conversation. What can you tell us?
There have been some headline numbers that have been picked up and shared widely. The answer is actually much more nuanced and it really depends on the type of cooling system and where that's being deployed. And actually it matters more the context of the data center. So is it in a water stressed region? Is it in a desert region as a lot of them are being built? Or is it an area with plentiful water? So in those cases it might not be so bad, and actually water use is something that reduces the energy of the data center.
So it's a bit of a nuanced question. It's a bit of a generalization to say that, you know, one ChatGPT query uses this much water. But ultimately we don't have a lot of data and so it's important to push on the data providers to provide some more information there.
So for residents who are maybe in those neighborhoods where the data centers are being proposed, are they rightfully worried about their water usage being diminished or maybe their water bills going up or their energy bills going up? Is that a fair concern of theirs?
It's definitely a fair concern and I think these concerns should be used to demand more transparency from the developers as well as, you know, the implementation of better water systems if they're truly going to be using a significant amount of water, better energy systems. So bringing clean new energy that's not just bringing gas power energy like a lot of these data centers are doing, and that's going to drive pollution in the local neighborhood. So there is a lot that communities can do in demanding a better future with AI.
There's currently a drought in the UK, Boris, right? We've got a hosepipe ban, but there's also data centers supposedly being built and I'm still using ChatGPT. So in that case, what is actually taking priority when it comes to legislation? Are those data centers using that water?
Yeah, and it really varies region to region, but oftentimes the data center loads are going to be able to use the water. We should say there are technologies out there that use less water for cooling. That doesn't solve the entire problem because upstream of that, the power that's generated for the data center could also use water. So it's kind of an interconnected system. And I think that the main takeaway is we need more information, and communities and people should push on the developers and on governments to push on the developers themselves to get more information.
So how did that happen? Because at the moment I feel like we've got an impasse. On the one hand, we've got people who are sort of saying don't build data centers. On the other hand, we've got the technology companies wanting to build the data centers. So how do we get the processes in place now for that conversation, the communication you are talking about, to happen?
Big tech and data centers have kind of steamrolled a lot of these communities and approached this development process not as they have with other industries, where you're looking to bring community in, but actually hiding behind shell corporations coming in there and surprising the community with these projects. Now there are some good examples of organizations, Google's a really good one, that are trying to work with a local community, that are working with a utility to manage any increase on price, to actually reduce the carbon intensity of the electric grid, to add battery storage and things like that. So this isn't to say that all data centers are bad, but we do have some really bad examples out there and now it's kind of on the data center developers to prove that they are not some of the bad ones and to show a better way forward.
The example that you just gave with Google, I'm sure that's on Google's initiative. So you can't really allow this to be a question of just allowing the company to be a good or bad actor.
Definitely and I think that's what some of the community pressure has done. I mean, the data center pushback in the US is more bipartisan than almost any other issue right now.
Anything. We finally found our common ground.
That's forcing these companies to rethink what they're doing, and I think that pressure should continue. Ultimately, I wish we had more government intervention here, but I don't think in the short term. Well, we are seeing some action at the local level with the state of New York, for example, and some other moratoriums proposed, but I don't see a wholehearted regulatory approach in the near term. We should push for that, but in the meantime, the communities really do have the power.
But do you think that data center build-outs are inevitable like the factories were, and is the only thing that can stop them is just the lack of resource on earth?
So the AI we're being pushed these days is generative AI, that's with models that are getting bigger and bigger, and the philosophy behind them is you need more scale. Scale in terms of the compute and thus the energy, the data, the model size. So the current paradigm of AI development is that we need more and more and more data centers basically. We're, our organization, pushing against that. And we see a lot of big companies that are questioning that as well. If we can take a smaller model and have the same business outcomes as these ever-growing models, do we really need multiple gigawatt plus data centers going up? I think that's an open question.
We think that the data center build out is significantly overbuilt, especially as you have open source models coming in, starting to commodify things. I do think a level of data center build out will be needed, but not to the magnitude that's currently being proposed.
Aside from energy and water usage, is there an environmental impact that we're overlooking, maybe perhaps one that might be coming down the road? And how transparent do you feel the companies are genuinely being with regards to environmental impact?
The companies are not being transparent at all. And so one of the things we're working on at my organization, Sustainable AI Group, is how do we estimate the energy and carbon and water use of these closed models so that organizations can better understand how their choices impact these effects, while also pushing for more transparency.
In terms of environmental impacts here, I think one that's under-reported is the air pollution to the local communities. So a lot of these data centers, especially the bigger ones, they're not able to wait to connect to the local electricity grid. So what they're doing is they're bringing their own power, and that power is very often fossil fuel, natural gas turbines that are often less efficient than even the gas power plant that's on the grid. And less efficient means more pollution, not just carbon emissions, but nitrous and so forth. And that is often happening in areas where residents are already disenfranchised and already have high levels of asthma and cancer. And so these environmental impacts, they're not just on the planet, but they're on people too when you have your neighborhoods with increased air pollution.
Data centers drive 55% of US electricity demand growth. Are my electricity bills going to go up as a result of this?
The US president has come out there and brought a lot of tech companies together and put together a pledge to not drive rates up, but ultimately energy experts do say that kind of voluntary pledge is not going to actually prevent what's happening. And what we are seeing is you have a lot more demand for energy on the grid, and the way these utilities are structured is that additional cost is spread amongst all rate payers, including residents. So generally speaking, yes. There are some solutions that some areas are using, like making the data centers pay more. And so there are smarter ways to approach this, but generally speaking, yes, this will raise electricity rates.
Boris, I would love your take on the fact that Sam Altman recently said that with regard to the amount of water usage being consumed when it comes to AI, that the numbers that are being floated are totally fake. What do you say to that?
I think this space has become really polarized and really controversial. I think on one side, you have folks like Sam Altman that are completely downplaying the issue. And on the other side, I think there is some alarmism. There are some numbers that have been taken out of context that have been taken to apply to all of AI usage. Those types of numbers that are taken out of context let Sam Altman say that the numbers are all fake because, sure, there's some inconsistency because there's a information vacuum due to these very AI companies.
I think there are very real numbers that we're seeing. Bloomberg just put out a study saying that all of this data center development could increase US power emissions by 20%, which is massive. This is a huge country with a lot of emissions already. So rather than nitpick at some of these kind of smaller edge cases that people are trying to do the best they can with the information they have, you can actually see this development of the data centers is potentially going to increase emissions in a big way.
All right. Perfect place for us to take a quick break. Coming up with Boris Gamazaychikov, we're going to be talking about an AI Energy Score, what that on other side the break.
So Boris, your group created the AI Energy Score. What is that and how are companies ranking so far?
Yeah, so the idea is when you get a car at the lot, you have a miles per gallon rating or a A through F rating in Europe, you want to understand the relative environmental impact of what you're buying. And so within AI, there's actually an even bigger difference than just at the car lot. It's not like a Hummer versus a Prius. It's like a jet plane versus a bicycle, where the frontier models are using so much energy and then you have small models that can do the same thing for a lot of your use cases at orders of magnitude less. And so there hasn't really been a good way to understand that difference.
And so we worked on how do you measure and disclose this information in a comparable way? And so we started with open source models because they're the most straightforward to measure and we created a rating, a one to five star rating. We're working now to also integrate closed models. So most of the models folks are using are closed, Claude, ChatGPT, Gemini models, for example. We're working on how do you bring those models into the same rating system so that anybody can go. We have a label, you can see this model has this star rating, and that could influence how somebody individually, but mostly we think enterprises that are adopting this technology, they're going to find this the most useful. And we've already done some pilots with some organizations on, as they're procuring these models, as their engineers are using different models, how do they see these star ratings and how does that influence their behavior?
So how do the various AI models rank on energy usage?
We haven't released the full results yet, but we will-
I got to know.
We will pretty soon.
So who's the jet engine? Who's the bicycle?
A general sense is some of the newer models are getting bigger, the Fable class models, the GPT-5.6 Sol. At the same time, small models are getting more powerful. So Google released a set of open source models, the Gemma 4 series. Those can actually run locally on your laptop and even on your phone, and they're actually the same capability in a lot of metrics as GPT-4o was, which was the frontier about two years ago. So we're able to do things with some of these small local models that we thought was completely groundbreaking two years ago, but we're kind of here distracted by the biggest shiny object on the horizon with things like Fable. So I think it's a mentality shift that needs to happen.
So what does good look like, Boris? Because what you said right at the end before we went to the break was kind of mind-boggling, right, that US emission's going to go up by 20% when we're trying to tackle climate action, climate change. So what does good look like? How do we navigate this?
I think in the near term, businesses have the biggest role to play in lieu of governments really stepping up. I think I've seen personally businesses really interested in understanding their impact of their use of AI and how they can push the AI industry towards sustainability. And what that looks like is transparency, first and foremost, really understanding the difference between the models that are being used, also understanding the difference in the actual data center that's powering the model. Is it a renewable energy powered data center which uses little water, or is it this data center that's massive, that's powered by fossil fuels? Very different environmental impact in those different scenarios, and both actually exist.
But that's kind of easy, or maybe not easy, for Google or for the big companies, but a smaller company, how on earth would it even start trying to navigate that?
A small company that is kind of procuring AI and wants to make an impact, I mean, I think the company should understand, even though they're small, they're joining this chorus of voices that are all asking for the same thing and are part of this bigger movement right now.
You mentioned though early on that these companies are not being transparent right now. They're not really wanting to play nicely. So how is that going to come about then? Is it going to be hitting them in the wallet? Is it going to be government regulations? How are we going to get them to play with us?
Exactly, yeah, hitting them in the wallet is the main thing that's kind of behind all of this.
And they've got infinite money.
Well, that's actually kind of how things are shifting. So to date, the AI industry has been focused on kind of the venture capital investments and a lot of these circular investments. Now, we're entering a new phase where they're going to have to show real revenue from real customers. So with that, you have a shift in dynamics, where the customers are going to be able to demand more. And ultimately, this isn't that hard for those companies to provide more transparency. So we really think they can get there. And there's other examples within the tech industry over the last 10 years where shifts like this have happened, so we feel confident it can happen.
I think you make another really interesting point about the different size of models and the fact that different models are perfectly good for different tasks. And it feels to me at the moment that what we are doing is always using Fable to do whatever it is that we might want to do. So how do we start educating people on that, that there are different models available for different tasks?
Ultimately, it does come down to the developer of whatever product has AI integrated into it. So we work with companies that are integrating AI into more traditional software, and we tell them, default to the smallest model possible. And if there's optionality, have some sort of symbol there that helps the user understand what the impact is. There is a lot on individual awareness, but ultimately it is about the designers of these systems and these products to build that in, and individuals can demand that, for example.
But I feel like we are just constantly justifying the energy use because the potential achievement is just so great. So Dario Amodei recently said, "You know what? We could cure cancer in five years." And when I hear that, I say, "Okay, if we could cure cancer, well, maybe it's-"
You don't mind if your water bill goes up?
Maybe it's worth being in a drought for 10 weeks in the UK, right?
Totally.
Yeah. How do you deal with that?
Yeah, I think folks have purposefully conflated what they're pushing forward, this frontier generative AI, with the AI for science. The frontier generative AI, let's be clear, is being pushed and is being developed primarily to automate knowledge work. That's the primary way they're doing reinforcement learning and fine-tuning on these models. It's for coding, it's for legal work, that type of thing. AI for science is very much a growing and dynamic field, but very rarely are those models the same type of frontier LLMs.
It's so frustrating because recently someone said to me on X that, "Oh, we haven't priced in the fact," some Silicon Valley person, "that we haven't priced in the fact that people are going to live two times as long." And it's like AlphaFold is not the same as Fable 5, and it's so frustrating.
AlphaFold takes very little energy, that's a great example. And a lot of AI models that are good for climate, like optimizing traffic or the electric grid, those are really small kind of domain-specific models too.
Yeah, I was going to ask, do you feel that there is a justification for, okay, yes, AI in the future of AI, we're going to have all of this extra taxing of environmental resources, but do you feel in the end it'll be worth it?
I think we can have the good AI without the trajectory that we're on right now, which is ever bigger fossil fuel powered model development. So I think we can have the good without the bad.
Well, that's great.
Yeah, I will take that.
We'll take, it feels to me, Boris, when... I mean, we've obviously had a lot of guests on the show, and it always comes back to the individual. It always comes back to the member of the public. It always comes back to us to find the information. It always comes back to us to ask the questions. It always comes back to us, and it always comes back to us with the smallest amount of information to go on. And the people with the power, it's nothing to do with them. It all comes down to us. In all honesty, how's that going to work?
My background is working on climate, and within climate, there's some great examples of how that's happened and that's not worked. I mean, even with the concept of recycling, that was something that the manufacturers of plastics put on the public rather than taking care of the issue themselves. So here, the individual can do things, does have power in terms of systemic action, voting, spreading the word, pressuring businesses, but ultimately, those in power, which is why I mentioned targeting businesses, the large businesses that can actually impact the spend for these AI companies, that's where the real power and the real pressure should be.
Yeah, I think it's just so frustrating that there's discourse out there that says don't use ChatGPT because you're contributing to the climate impact, whilst Goldman Sachs is rolling it out to 300,000 people five times a day, Fable 5, yeah.
I agree. I definitely agree.
Quick fire.
Where do you think people will draw the line on bringing AI into their personal lives?
I think it's been interesting to see pushback, growing pushback on tech within the data center layer, but also within the Meta glasses side of things.
Good point.
Yeah, I think that's a great example of, people are having a hard time getting off social media. They know it's bad for them, but they're having a hard time getting off. But here's this new thing, and we can push back on it. And with Flock cameras and surveillance being pushed back on, I do think that people have hit their limit, and I hope any new intrusion like that will be thought about more holistically and we push back on.
What's something that's universally accepted in your field that you actually disagree with?
I think there's a lot of folks that do think that you can't use AI at all if you care about the environmental impacts of it. I don't think that's accurate. I actually think understanding how these systems work makes you a better advocate for more sustainable AI.
What are you optimistic about?
I've been surprised and optimistic about how this has become a topic of conversation. I'll talk about working on the environmental impacts of AI these days, and a lot of people will immediately nod their heads and understand that it's an issue, whereas a couple years ago, people didn't really understand that. And just the pushback, I think we can really take this energy and this momentum and create a better future.
Okay, so not all doom and gloom from Boris, which I thought was maybe what we were going to hear, but he's quite optimistic that there can be both AI and an environment for us to enjoy in the future.
Yeah, reassuring. I think what he did for me, which I really needed, was just untangle this all and untangle the confusion with AI for science, the energy demands of that versus frontier general AI, and just how that's all being confounded to push an agenda. And I now feel like I have some clarity. I didn't realize that there was such a discrepancy between, I guess, the AI, AlphaFold versus ChatGPT.
Yeah, no, I think that's really interesting, and the Fable versus the kind of open source model and the different types of electricity usage. So I found that really interesting. I found him really inspiring. I'd say the thing that gets me though is everything's dumped on us, right? The regulation is dumped on us. The big companies, government can wash their hands on it, right? It has nothing to do with them. It's down to us to put the pressure. It's down to us to sort it. That's just not going, if you pardon the pun, wash.
No, but I mean, it is what you're seeing though. You're seeing your everyday citizens and the people within these communities where they're proposing these data centers, they are the ones fighting back, which is not right, but at least they're taking action.
What's so painful is that we've seen this before, what you were saying about recycling. And I mean, I'm not an anti-climate person, but I never kind of recycled because I'm exhausted by it. Like what am I going to do?
So we've seen this movie before and it doesn't end well.
Yeah, and here we are with a drought in London.
All over again.
All right, well, our thanks to Boris, and I will be looking forward to the data that they release on how well the companies score when they finally put that out. That should be interesting. Maybe we'll have to have him back. Yeah. All right, thanks so much for joining us here on Agents of Tech today. Be sure to weigh in in the comments below. Let us know what you thought about our conversation with Boris. And of course, always like and subscribe while you're there. We'll see you next week on Agents of Tech.