Good afternoon everyone and welcome to this panel on sustainable AI. Thank you so much for attending. My name is Julia. I'm a PhD student at the TAI center researching how AI can help us map urban stocks from space and I'm your moderator today. It is so important for us to be having this conversation today, trying to reconcile the seemingly paradoxical natures of AI, both as this giant entity that consumes an enormous and increasing amount of energy while simultaneously being this solution for reducing and responding to climate change. As a PhD student navigating this fast changing field, I'm very grateful to be sitting here today with experts on the field and I'm really looking forward to hearing from your individual perspectives on what the future of AI can look like for a sustainable future. So without further ado, I'll briefly introduce our speakers. Our first speaker is Himanshu Gupta, co-founder and CEO of Climate AI. Himanshu works on climate risk forecasting and adaptation at scale. He helps organizations across food, agriculture and the global supply chain anticipate disruption, strengthen resilience and make better climate aware decisions. Our second speaker is Varun Sivaram, founder and CEO of Emerald AI. Varun's company develops AI solutions to make data centers flexible grid supportive assets unlocking capacity and reducing energy stress. He brings expertise across clean energy innovation, high growth technology leadership and climate diplomacy. Our third speaker is Eric Gelm, independent board adviser and co-author of the net positive AI energy framework. Eric has also co-authored multiple no excuse white papers from the industry net zero accelerator initiative and his work focuses on the operational and governance playbooks that help companies move from commitments to measurable impact. Thank you all for being here with us. Before we begin, please don't wait on me to call on you. Feel free to bounce off of each other ideas. This is a discussion and yeah I'll dive right in and my first question is for Himanshu. We talk about climate change as this future threat but you're seeing the front lines of food systems every day. Where are climate pressures most visible and most disruptive to global supply chains today?
Thanks Julia. You know, when we talk about climate change, we talk about in terms of 2050, 2100, right? And I think that's one of the biggest mistakes we've ever done from a communication standpoint. I'll come back to your pressure point as well about food and agriculture because when we say like by 2050 the world will be 2.5 degrees, it doesn't mean anything. If you, when we travel to the farms, when we travel to talk to food companies and processors, the real way climate change is showing up is increase in volatility. That's one. And two is companies, for any company like CFO or a chief procurement officer, they care about volatility. So what we are saying is like, so how volatility matters, right? So let's say if you are a farmer, you are used to dealing with crop losses once in 10 years because you have balance sheets, you have savings that you can recover from one year of losses or maybe at max two years of losses. Now what is happening is sometimes within a span of 5 years you're seeing three consecutive years of losses and so the adaptive capacity of farmers is not there and as a result many farmers are declaring bankruptcy or leaving farming altogether. Right? That's the real narrative of climate change. We should be having around volatility and then why volatility matters for companies that these farmers supply to. Let's say if a potato farmer is supplying to PepsiCo, right, for that potato gets converted into a Frito-Lay product for producing chips. Now PepsiCo cares about volatility in their supply chains. So if they are not sure if this big supplier or this big farmer is going to be able to supply consistently and reliably this year or next year and next year, now they need to do something about it. And so that's a discussion we are seeing among food companies as well where I was part of a breakfast meeting organized by the CEO of PepsiCo. There were 12 other food company CEOs there. There were three agendas and one agenda was climate volatility and climate resilience as a business case.
Great, thank you. My question is for Varun. I want to ask you where do we stand in terms of grid capacity versus the growing demand of AI compute and what's the impact on energy costs among others.
Thanks. Thanks to AI House for doing this. Hi everybody. I don't know about you, but I had lunch. I'm feeling a little sleepy. So, I think we all need to get a little high energy. I'm going to do my best to pep us all up. Why? Because I'm super excited and anxious about what the AI age holds. We're at the beginning of what I consider an AI super cycle. I went on record two days ago here in Davos with Axios and I said humanity's built less than 1% of the compute we're going to need. It's likely less than a tenth of a percent of the compute we're going to need. So to your question, where are we in terms of grid capacity? We are woefully inadequate. I'll give you some numbers. In the next three years, America, where I'm coming from, America wants to plug in 50 gigawatts worth of AI data centers. We will successfully be able to plug in 25. Around the world, this is playing out. There's only one country in the world that doesn't have this problem. You can guess which one it is. By 2030 in China, there will be 400 gigawatts of spare capacity for China to plug in AI data centers. China does not have a power bottleneck. The rest of the world does. China has a chips bottleneck. America has a power bottleneck. So does Europe. So does India. So does the United Kingdom. That's why we exist because there's 100 gigawatts of spare capacity just lying around stranded on America's grid today. We aren't ready today to unlock that spare capacity. We certainly aren't ready to compete with China in terms of the massive energy generation and grid buildout that they have successfully accomplished and we need new solutions. And the point of this panel as well is to ensure that those solutions are both competitive, keep prices low for communities and are sustainable. And if we want to be sustainable, the marginal new addition of capacity today for AI happens to be for natural gas in the US and coal in India. So there has to be a better way forward. Happy to talk about it in a moment when we get back around. But Emerald AI, the company I lead, tries to enable data centers to be more flexible grid citizens such that they can connect to grids faster, take advantage of that 100 gigawatts of spare capacity in America and hundreds more around the world, and integrate clean sources of energy. You can do all three in a win-win-win scenario.
Thank you everyone. Eric, you co-authored the net positive AI energy framework. What needs to change in design, governance and incentives so AI strengthens the energy systems instead of straining them?
First I would say a problem well defined is already half solved. So it's super important that we actually bring together a shared understanding of the reality and we clearly are at an inflection point. This is a time of the year where not only there is a Davos forum but actually there is an annual publication of the global risks report. I really encourage all of you to peruse that report because it is number one, second to none in terms of breadth and span of risks. And if you did already glance at it probably you would have picked up that indeed as Varun alluded to, there is a huge data center energy demand increase planned by 2030 to 2035 to the point that the estimations are that data centers will consume 20% of the global energy. 20%. In the US alone, the same report highlights that the AI, purely the AI data-driven energy would be multiplied by 30-fold. 30-fold. And all this happens at a time where every day, as Himanshu said, not in 2050, it's every day you see how extreme weather events are actually multiplying, impacting not only people but the energy infrastructure which happen to be quite aging across the world. And on top of that we are seeing a number of increasing grid stresses. So what we need, if we all recognize that the conversation is that now AI has become an energy story and even if you don't care about the planet but you do care about the economy which somehow should be the same, how do you want to make a business in a world where people don't thrive, can't properly live or breathe? So what do we need? We clearly need a systemic intervention and something that allows people to shape a dialogue across not just one element about yeah AI can save 10% energy or help here but the total net effect of AI is positive. A bit like if you are on an executive board nobody cares about whether you do that saving if at the end of the day the bottom line is not where it should be, right? So the net positive AI energy framework is articulated across three simple drivers. The first one is design for efficiency. So clearly designing AI for efficiency is key because number one, you know, it actually helps the hyperscalers, the tech providers, but obviously it helps also the overall infrastructure. You don't need that much energy and that's basically the race that I would trust most of the technology companies are already doing by themselves and Emerald is one of the companies that can actually help them going even faster. That's absolutely critical. Whether you talk about, you know, high efficiency technology or obviously renewable siting, that's the first thing. The second one is deploy for impact. AI can actually be a huge enabler in terms of how we save energy. Just to give you an element, I mean for those of you who follow a little bit about renewable energy, you know that one of the key constraints is what we call renewable curtailment. The time when you can't actually use that energy because the demand doesn't match the supply. With AI we can actually cut by more than 20% the current renewable curtailment. So using AI and deploying it for that impact is the second leg. But the third one is actually equally important, it's shape demand responsibly. And we'll talk that probably later but to shape demand responsibly you need to have a combination of both technology and education. So basically if power has now become the new constraint as you highlighted, resiliency becomes the imperative. I think just to highlight the seriousness of the points that my co-panelists have made, if you remember two years ago there were 8 to 10 retailers in Europe, energy retailers in London, that went bankrupt. And the lack of energy became the affordability and cost of living crisis story in London. And you know the reason why? In the North Sea wind stopped blowing for a week. We call them wind droughts. And most of the European grid was running at 30 to 40% renewable back then. Suddenly, you know, the grid could not handle that much of imbalance between demand and supply. So prices went up and then the retailers could not pass on the cost to the consumers and there were bankruptcies. Now you add another layer of these data centers to that, you're not even talking about like emerging markets. We're talking about Europe as a grid. We're looking at a very serious story here.
Thank you all for setting the scene. Going back to what you were saying, I want to ask Himanshu and Varun. In the intro, I mentioned how AI can help us reduce and mitigate climate change. I wanted to ask you, could you elaborate as to why AI was the key to both of the businesses you built and some of the measurable impacts it enabled?
Yeah, we think of AI as the time and effectiveness multiplier to the solutions of climate change, right? And I won't take too much time here just to demonstrate. Look at a light bulb above you and not the ceiling. Is there a light bulb here? Okay. Can anyone guess why that light bulb, the history of light bulb is also a symbol of the rapid progress AI is making in the solutions to climate change? Any guesses? Height. I'm sorry. The height. Height is much more efficient now than it used to be. You're getting there, but that's not it. Anyone else? Any guesses? Okay. So, light bulbs were invented by the Edison labs, Edison and his lab mates and it took them like four years to go from like a prototype to commercialization. And the reason being they were trying, close to like, they had to conduct 1200 trials before they could actually say that this is the material in the filament that I could use and can withstand 45 degrees or like 60 degrees. Until then they had to keep on experimenting with multiple materials and they couldn't find an answer. It took them three or four years to find that answer. Right now, imagine if the Edison labs had the current AI tools. This would have been done in minutes like you could simulate what materials will perform the best in these heat and temperature environments. So we use the same solution for seed industry, right. It takes 10 to 12 years to launch climate resilient seeds into the market and one of the biggest delays in those launch is for the seed companies to figure out where to conduct trials. So if I want to launch a climate resilient tomato seed for farmers in Zambia, I need to figure out where should I be conducting trials to understand whether this seed will perform the best under Zambia's temperature conditions. So what seed companies do is they literally send teams of four or five each across 10 different locations. These teams will set up camps, collect a lot of soil data, water data for like almost two years, send it back to the headquarters and then headquarters will conduct analysis based on historically that okay, this location seems like this is the best location, right? What we do is like okay can you just draw a circle on the map where you want to launch your seed and then we'll automatically extract and sort of think of this as control C and control V and we can exactly figure out under climate change what are the right locations for you to conduct trials. So instead of conducting 10 trials you can just conduct two or three trials. So the time it takes to launch that climate resilient seeds reduces by 20%. And the cost also reduces by 12%. Right? So now this platform we call that like climate analog is being used by 30% of the seed industry.
This is a key question you've asked, right? You all got, everyone came here to learn hey is AI going to be a climate disaster or is there hope? And so Himanshu just shared with us this really helpful story of how well if there is a climate disaster at least AI is going to help us to be resilient towards some of these damages, right? And AI is also we hope going to help us to develop revolutionary new materials to accelerate clean energy, whether it's nuclear fusion or solid state batteries or carbon capture. But in the near term, we were just talking about how AI is going to explode in terms of its energy demand. You said the figure 30-fold. That's insane, right? In the United States, I think the other figure you said was 20% of global energy. I hadn't even heard that figure. 20% of global energy is going to come from AI. That's insane. And today it's all coming predominantly from fossil fuels. There's no way you get to net zero emissions anytime this century if AI does that right. So Emerald AI I believe is one of these solutions to make AI and clean energy compatible. Here's what we do. We're an NVIDIA startup. We've been around for about a year. And what we do is we take these AI data centers and we flex their power. The reason that's important for an AI data center to reduce its power every so often is because if you can do that, you can fit your data center on an existing grid without overbuilding that grid or that generation to meet a higher and higher peak load. Think about the grid as a superhighway. It's empty most of the time. It faces, bless you, peak load two times a month, right? So most of the time most of the lanes on the road are empty. That's like the grid. Power lines are running at less than full capacity. Power generators are running at less than full capacity. And so you take a data center and it's flexible. And at the worst, bless you, moment of the year or of the month, that data center is going to reduce by 25%. We do it in one of three ways. You can slow down computations like a fine-tuning run on a model that doesn't need to terminate right now. You can move workloads like inference workloads from one location to another. That's called spatial flexibility. So you have temporal flexibility, spatial flexibility. And third, there's resource flexibility. You might have a battery on site that provides you some on-site power and you can draw less from the grid. If you can do that, you can fit your AI onto existing grids. You don't have to expand firm dispatchable generation, which often comes from natural gas or coal. And finally, because your load is flexible, it becomes easier for that grid to integrate clean intermittent sources of energy. Today, the fastest, cheapest source of energy is renewable energy, solar and wind. And it's easier to pull those onto the grid. And so, you can imagine AI going from 0 to 20% of global energy or growing 30-fold could be a disaster if it is inflexible and you only power it with fossil fuels. But it could be a miracle or a savior if it is flexible in such a way that it both protects the integrity of the AI because that's really important but it also makes the grid more resilient and tolerant to intermittent and clean sources of energy. Right? That way we unlock not just abundant clean energy, we actually make it cheaper and easier to bring clean energy. Because today the hardest part about clean energy, I know I come from the clean energy industry. The hardest part about renewables is there's no one to buy your power because it's so volatile. That's what Emerald AI does. We're excited. The last thing I'll say is later this year with Nvidia, we will announce we'll open, we've already announced it, we'll open the world's first power flexible AI factory or data center. It's 100 megawatts in Virginia and we expect that the next thousand data centers then can be built to that reference specification. I'm super excited.
Thank you. I just want to circle back on something you said and maybe moving away from AI solutions to potential barriers to adoption. I want to ask, so you've explained how data centers can support the grid but specifically what barriers are still slowing this down? Are you seeing any resistance?
Totally. Totally. Because this is not the way we've done things for a century, right? For a century the electric power industry has said, 'Hey consumers, you're kind of dumb. I'm going to power you and I'm not going to rely on you to change your behavior.' Right? So if you walk into your bedroom, you turn on your light, the utility is not asking you to change how you turn on your light. Actually, in Europe, they do a better job of this than in America. And on the other side, the data center industry says, 'Hey, I want to be able to use power whenever I want to. I am the cool kid in school. I'm the most important person there is. I deserve my power. Give it to me. And I don't want any compromises.' Both of these sides have to change a little bit. Not much, but just a little bit. The AI industry needs to be a little more tolerant of flexibility where it's able to say to its AI compute customers, look, we can have many times more data centers and they'll be grid friendly and community friendly and drive down costs, but every so often when we hit that peak load moment, a hot summer day in Phoenix, Arizona, where all the air conditioners are running, you've got to be okay with a 25% power cap. Emerald, by the way, makes it possible to do so by gracefully orchestrating those workloads. So the highest priority workloads continue full tilt. And the utilities need to say, 'Hey, if you offer me this capability, I'll connect you faster.' Both sides need to compromise and change the way they've historically operated. That's in my opinion the principal barrier. Regulators can really help by saying new connection methodologies. If you're a flexible data center, you get to go to the front of the line. We'll connect you faster because we want more of you and we want less of the inflexible data centers. That's a policy approach to really drive this change.
Thank you, Eric. In the report you mentioned the Jevons paradox. So I want to ask you when scaling AI, how do we prevent rebound effects from actually canceling out efficiency gains?
This is a super important point because as was alluded earlier, you know, if you isolate efficiency gain you can certainly be quite enthusiastic about the prospect of AI. But I would go back to what working with a board is the typical question at the end of the day: what is the net effect? Whether that's dollar, CO2, water, communities impact. So in that context it's really important to keep in mind two very simple but absolutely critical things for all of you and all of the ecosystem providers in AI to keep in mind. The first one is the Jevons paradox. We've seen that before with the invention of electricity at the beginning of the 20th century or recently with internet and now with AI. What is the Jevons paradox? The cheaper and faster you make a technology available the more we use it. Example: ChatGPT, five days to get to 1 million, two months to get to 100 million, all records broken, right? That's the evidence that the Jevons paradox actually can cancel any efficiency benefit that AI can bring right now because just of the demand explosion. The second element which is not talked well enough and which is not by the way specific to AI but which is exacerbated by AI is dark data. Dark data is basically all the data that we collect and that we don't use. Two years ago, there was a report by Splunk, you can check it out, 2024-25, that actually called out that 55% of the data collected by enterprises are either not leveraged nor analyzed. Think about as a supply chain person, do you want to have 55% of your inventory that is not used at all or not sold? Wow. Okay. So when you think about those two elements, it's really important to keep that in mind because right now and we alluded to that the backlash already started at local level. There was an article in Newsweek back in August talking about the eye-popping price increase of energy in some states in America. There were equal reports about 22% energy bill increase in Ireland. And there is actually a YouTube video by Business Insider about the dark side of the data center which are the consequence of the AI revolution about, you know, the noise, the pollution. So it's really tangible what is happening and how we need to basically adjust that curve because I am totally with Varun and with Himanshu. AI is a major breakthrough that we must embrace to actually go sustainable. So how do we go there? Basically three high-level directions. The first one is right-size. A number of technology providers have actually started to think about how do they route. So let's say that you want to say where do I get to the conference center? Do you need the super heavy one zillion
Parameters, you know, genies to tell you that, or do you just need a Google answer? Right. So the first thing is right-size the model. The second thing is compute at the right time and at the right moment, which, you know, Varun already explained. And the third element is measure and educate. That is super important. How do we send price signals both to the consumers and to companies about what is the bottom-line effect of the AI energy consumption? You have not just, you know, a vanilla type of price where you don't see the cost and the total footprint of the AI consumption you're having. So basically, if data is the new oil, dark data is the leak, and Jevons' paradox is just exacerbating that leak.
Thank you, Eric. Just following up on that, AI efficiency is often opaque. We're seeing initiatives that are trying to promote transparency, but how do we measure it credibly so that efficiency becomes a real lever and not just a slogan?
So we are at the start of it and it's not yet done, but there are a number of conversations happening now. Because to answer your question, basically the first thing you need to have is a standard way of measuring the overall multi-dimensional lifecycle footprint of AI. In the AI Energy Net Positive AI Energy Framework report, if you check it out, Accenture, which has been a partner of that report, a brilliant partner, deployed last summer the Sustainable AI Quotient. That is a simple metric that blends financial, energy consumption, CO2 emission, and water, because water is very important. So we need to start having, you know, a coalition of both tech providers, industry leaders, and obviously governments that agree on, hey, this is now the new way of measuring mileage efficiency for a car and now for AI. So that's the first one. The second one is obviously tied to incentives. And that could be, you know, governments actually incentivizing highly efficient hyperscaler tech providers and obviously efficiency in terms of price signal for the consumers. So that's basically, I think we are at the point where by necessity the entire AI ecosystem is going to need to start measuring that. And there was a report a few months ago about Google who actually proudly displayed, you know, how much they were progressing in terms of energy efficiency and carbon footprint. Nobody forced them to do that, actually. They started talking about it, and I think we need more of that.
Thank you. Continuing on potential barriers, Himanshu, I want to ask you, how do you build trust in predictions that cannot be perfect, right? Specifically, how does your team bridge the gap between a probable climate risk and then a concrete business decision for your clients?
That's the hardest part, building trust. So one of the central values of our company is trust. So what that means is when we go to our customers, like seed companies, we tell them A, what the model can do, B, what the model cannot do, number one. Number two is, you know, the finance industry is the only industry that knows how to act on uncertainty, but the rest of society is not comfortable with it, right? So as an example, if I tell some of you like there's a 30% chance of life-threatening hailstorms happening in the next 15 minutes, you'll ask me like, what do I do with this? Right, tell me, is it happening or not happening, right? So when we go to the industry and we say, look, we cannot tell you that it's happening or not happening, but if you're optimizing your decisions around something, so for example in our case, many food companies when they are buying products from six different regions, in a lot of ways they're acting as mutual fund managers, right? Maybe our model would be wrong in one place, but if you look at the entire portfolio, our projections will be more than right. Right? Similar as good as like you have a mutual fund investment, you know, instead of investing in a stock, you invest in an index or a mutual fund and say like over the next 10 years this is going to give me this much return. So that's the approach we take. And even this approach required a lot of evangelization. We sat down with our users on the fields, on the farms, in their labs, and explained to them this is what it means, this is how you act on uncertainty.
