The AI Dilemma: Can Society Have Its Code and Compute Too? Do We Need to Go Nuclear?

Maybe to set the scene to start with in terms of there's a lot of different question in terms of how much energy they are actually using.

I think Microsoft has said that its energy use would be six times greater in two thousand and thirty compared to its two thousand and thirty targets in twenty twenty because of greater AI usage. So maybe I want to kind of, you know, set the scene and ask the question because what we prepared at panel, we got kind of slightly different perspective. So, Philippe, if I want to start with you because you said that the use of AI was a crisis in the making. So maybe if you want to say a few words in terms of what you're seeing in terms of the need for energy for AI.

Yeah. Thank you. I mean, this is obviously a really important question and one that is somewhat imponderable at this point. Clearly, with the advent of Gen AI, with the massive acceleration in the usage of AI, Data centers have seen very strong growth in energy consumption and really, it depends who you believe.

And even over the last couple of years, we've seen lots and lots of variation in what people think could be the increase in energy consumption. A bunch of forecasters broadly believe it's going to be 3x in 3x increase in the next ten years. Could it be 6x? I mean, Microsoft are clearly there.

And it will be interesting to see what people like DeepSeek, which is obviously widely trailed, what that what that means. And, obviously, my esteemed companion here from NVIDIA probably has a better idea than I do on that one.

Do you wanna answer that?

Yeah. I can I can I can jump straight in that?

Just so so, I'm in video with the, AI, infrastructure company. So we build the kind of design the hardware, behind a lot of the kind of AI revolution and think, keenly about energy efficiency. Because if you think at a very basic level about what AI is, it's kind of tasks out one side, but energy, energy in the other side. So energy efficiency is is really important in that, and we've seen kind of dramatic improvements in energy efficiency, over the last ten years, kind of a hundred thousand times improvement in energy efficiency.

And I think DeepSeek is is an example of, of where we've seen the kind of, researchers or or actually not researchers kind of, hedge fund wants in, producing a new kind of reasoning model and a way of doing AI that is kind of more efficient.

And I think that's a kind of a good example of where AI is going in general, but we've seen just a huge increase in efficiency of, training of AI models, and the inference. So, like, using those AI models with live data, over the last ten over the last ten years. The the stat I kind of always use, which is if we'd seen the same efficiency in motor vehicles as we have, in, semiconductors, in the lot since twenty sixteen, we'd now be able to drive to the moon and back on a single tank of petrol. So we are in kind of one of the most, innovative efficiency seeking, industries here, and DeepSeek is a good example of that.

I'm gonna reuse that stat. I'm gonna study from you because I think it's a good one. Lucy, you mentioned that it was a bottleneck but done in Akilesia. So maybe you you said in terms of energy usage of AI that it just kind of was a bottleneck but not very much in Achilles. Yeah.

So maybe if you want to kind of go a bit deeper on that.

Yeah.

So I I agree with the previous comments that how much energy AI is is currently using or how much increase in demand it's driving is a bit of an unknown. And, as they both said very well, we can expect to see efficiencies in both on both the software side and the hardware side. So improved algorithms driving efficiencies, but also improved chip technologies also driving efficiencies.

Having said that, I think in the near term, we will see more demand from AI, not least because there are more people building more models at the moment. So even if these become more efficient over time, at the moment, we're seeing an increase in demand.

I think your question was about does it does it need to be, is it an Achilles heel or just a short term bottleneck?

Well, I think absolutely there are solutions to address this increase in demand in the short term that might actually also be beneficial for the wider energy system. So, I think there are ways that we can be thinking about the increased power demands for data centers in particular that actually might also help to integrate more renewable sources into the energy system and actually reduce costs, in particular things like costs associated with constraints in the energy system. So I by no means think it it is, an Achilles' heel. In fact, I I actually think the activities of some of the hyperscalers, one of the, maybe one of the biggest favors they will do us all is actually speeding up some of the the regulatory challenges So I don't think real the real challenge at the moment actually is about the supply, to power these data centers. And I think it remains around, the regulation, things like permitting, the speed at which, grid connections can be obtained.

Thank you. Rob?

Yeah. That's thank you.

Agree with what Lisa says about the supply side constraints. I think it's also worth considering the the demand side in AI is globally a very small fraction of of energy generation that's expected to grow hugely, but no one's quite precise about that.

The what the way the generative AI industry is set up at the moment is is that the AI businesses are still working out their business models. You know, this is this is fueled by private investment. It is not it it we as the consumers of it are not paying the the bills, essentially. It's it's an investment.

But we will need to see where that settles down, which will, at that point when that happens, we'll see what are actually the capabilities and molds out there, what's the what's the pricing of them. And then us as kind of consumers of these things, it's commonly thrown around that it'll take, you know, years and years, maybe decades to get even more with what we have today, work out the right uses for it. It's very easy to say, AI can do this, AI can do that. But the the real question is, what is AI good for?

Because AI can very easily become an inefficient and costly way to do something that can be solved by a more, like, bespoke approach correctly and cleanly and cheaper. Like, AI is getting more efficient. Inference cost is, you know, adding some numbers together is still always gonna be a grossly inefficient use of an AI model even if you can do it very, very well. So we'll we'll also see the demand side kind of stabilize, and people will work out what AI is really good for.

Because I don't expect in ten years the conversation will be, sort of the same level of hype as we see today. It'll be more focused on real real use cases.

And I think the timeline you're saying is quite interesting because, as I'd like to remind everyone, ChargeGPT was in November twenty twenty two. So it wasn't that long ago, if you think about it. So you both mentioned in terms of what is AI good for and what can AI do for the energy sector. And we also got two of present and looking at the more specific nuclear sectors. And maybe, Jeff, if I start with you, in terms of, you know, AI, or do you think AI can help the nuclear sector?

Well, I mean, I think, we can see already that there's a there's a number of, applications that are emerging and and start up companies that are out there working on them to, to basically address, issues in the nuclear project life cycle that, could help out the nuclear sector in a in a big way. I mean, AI can, is is looking to to sooner or later be able to, you know, greatly facilitate the design, licensing, construction, supply chain management, decommissioning that's already having a big impact, I would say, including through robotics and, you know, digital twins and and that sort of thing. So, you know, I think in in short order, we're we're gonna start seeing, some of these solutions appear on the horizon.

The IEA is convening a a a symposium on AI and nuclear in December, and, we're gonna be inviting a lot of a lot of these, well, all of the big tech companies, but some of these start up companies that are working in particular in particularly on these applications. I think in the US, there's already three or four that are starting to to, kinda make a name for themselves, and, also in Europe and, of course, in Asia. So, the the hope is that, you know, a lot of the bottlenecks that exist in the nuclear project life cycle right now could really be, addressed, by by AI, you know, hopefully bringing down, costs and and and, you know, cutting some of these timelines that that have been an issue for nuclear.

And, Andrew, I saw you nodding a couple of times. Obviously, you're a nuclear nuclear company. So what's your perspective in terms of what AI can do, what AI will do as well? Yeah.

Thanks very much. I mean, obviously, I agree with the all of the, the ways in which AI can help us as a as an advanced nuclear technology company.

I mean, I think that's clear, and it applies in the same way as it does to any, industrial user of AI.

I think the other angle of it is, you know, how how is the synergy between AI data centers and their energy demand, and how does that actually work very strongly with, advanced nuclear technologies coming online, you know, in the in the next, ten years or so. And, you know, we are a technology developer, but we're also effectively an energy plant developer. You know? And we've been around the UK looking at different siting options.

And and as we walk around and talk to landowners, so they have all been approached by data centers. And And so we're actually we're all competing for the same land, but that's actually a perfect synergy because we all need similar things. Right? You you we all need a grid connection, but there's a demand and a supply side.

And, you know, the other interesting point which you you started to mention, I think, a little bit around, you know, how does, how do you modulate the different, sources of power as part of this transition that we all are enjoying?

And and, I was at a a conference a data center conference in in Antibes back in, the summer last year, and I think it was Google was talking about how you could at the you know, instantaneously, AI driven and controlled, switched, capacity or demand from one country to another, which, of course, you can't transmit power like that, but you could transmit demand like that. And so you've got this kind of perfect ability to modulate demand to match supply, which is what the energy transition needs, not only to maximize the benefit of firm power like nuclear, but also to integrate, nondispatchable power like renewables. So, actually, if the three, I say, traditional renewables, nuclear, and and and data center users work together, what a brilliant system solution that that could be there.

Andrew, how do you think about the you mentioned the ten year interval until Yep. Realistically AMRs, SMRs are going to start appearing. How do you think about the gap between them or the interval and what what data centers might do in the meantime?

Yeah. I mean, you know, that's the challenge we've all got. Right? And, you know, people have talked about the growth, what is it, six times or something that we're gonna be seeing, until twenty thirty. And then beyond that, who knows what. Right? So, we've all gotta have a long term view and a short term view.

Grid, to some extent, can meet short term views, and I know that the sort of net zero team in in current government are considering short term, but we also need to be looking at long term. And we need to be thinking, what is the long term sustainable solution that best deals with rapid rapid upscaling? And I think that's where nuclear really can come up come along.

But, actually, yeah, of course, it can't solve the next five years, but it can solve ten years. So for me, that's not a reason not to adopt a nuclear solution, but actually, yes, we need to look at a at a at a short term. And, actually, sites, projects, opportunities, countries that can solve that interim are the ones that are gonna get the, I guess, the greatest adoption of, of the hyperscalers coming to that area.

I think the timeline is quite interesting because, obviously, it can't happen now. Very interestingly, when France announced their big deal, and they basically kept saying that, you know, we've got nuclear in France, so therefore come to us because we're far more renewables and sustainable than some of our competitors, you know, as a country.

Tom, I kinda want to ask in terms of the relationship with kind of public policy. What important is it to get that right? Not only, I guess, from nuclear, but also in terms of renewables going forward.

Anybody wants to take that, or do I just speak on something?

I'm just gonna do that.

I can I can I can start us off on on that one?

It probably is is one of the least specialist people on that area. I'll give a quick intro.

So I think the they need to play together. Right? The the technologies that we're using for for generation now, we need market designs that that allow those to actually be deployed effectively.

And so we've talked a lot already today, this morning, and and many of the the sessions we had, a lot of discussions about some demand side flexibility, about storage.

We need markets that are gonna incentivize the right behavior, and that could be in terms of scheduling industrial activity for times of low carbon intensity or the right price signals reaching them. It could be it or or excess renewable generation. I think, you know, in the UK, we we we we pay a lot to turn off wind type turbines when there's too much wind. The price of curtailment is not what you'd ideally ideally have.

But we also need to look at kind of business process, kind of what's going on scheduling in cloud data centers, all the way down to kind of household level, how people can smartly schedule things going on in their house or ideally have, like, autonomous systems doing that for them.

But we need we need policy and regulations that create the markets that incentivize that activity.

One that I think here that I'm I'm very happy to hear people tell me, tell me differently. But I I feel that to get to get market balancing kind of flexibility outside of specialist people operating the energy industry, you need to just have markets to get really clear price signals. That's how kind of households will will respond. That's how you can easily get other businesses on board without specific subsidy schemes and offsets, etcetera.

So so the things that that whether in the works that will work towards that is we have kind of half hourly reforms on the way in the UK markets right now, GB market.

So that's gonna incentivize time of use tariffs, which will provide more ways to send those price signals to to, like, small businesses, to domestic users. And then the we could talk about policy of, say, regional markets, zonal markets, so we can then create incentives to collocate high energy demand. It could be data centers, it could be anything from an industrial perspective close to sites of renewable, which, say, in the UK is not that's not really a a system that exists yet on in terms of wholesale markets. The pricing is unified.

Anybody else want to comment on that? Because this collaboration, obviously, between, I guess, energy companies and governments, but there's also gonna be, I guess, greater collaboration between energy companies and technology companies in there. So maybe Andrew and let me remind everyone, you can just slide those to ask question. With a bit of luck, it will come on one of the screen, and you can see the QR code. So please do.

So maybe, Andrew Oh, sorry.

Go ahead.

Oh, Philip. Yeah.

As I was concerned, the zonal piece is quite interesting.

Mhmm.

I mean, even now, you can observe, say, in Ireland, about twenty five percent of all that energy gets consumed by data centers. You can see the same thing in Virginia and East Coast of the US, and quite high levels in, say, along the a three in the UK. And actually getting that right, which I think might come back to Andrew's point about planning and both looking for the same provider of land. If you can get co located either hubs set up to provide power for a range of possibly data centers or other engine intensive users, that could be quite compelling. So I think it linked back to Lucy's point about, well, actually the grid are going to be the net recipients of some quite large challenges at the moment.

And the it's been designed for a fossil fuel based world. And depending on where you site your nuclear plants, for example, that might massively help solve some of those problems.

I mean, you mentioned France, and that, you know, the French government is is putting pressure on EDF to offer favorable, tariffs for, people who locate data centers next to, nuclear power plants. There's a similar, Biden era end of the Biden era, executive order that, requires the, the US, nuclear plants to do something similar and make space available for, for data centers on or around, US nuclear reactors. So, you know, that's obviously a really positive signal. But then, you know, basic economics will will drive the fact that a nuclear plant in the UK or anywhere else for that matter, collocated with a data center will drive the power purchase price behind the wire, which is far lower than would be available on the off the grid, and far lower than would be available on a kind of firm basis from any other provider.

And so there should be fundamental economics that drives data center developers, nuclear power plant developers to get together to talk, to collaborate, to do these deals, and to have a longer term vision, you know, that allows both of those projects to move forward even in a kind of a scaling factor applying on an individual site.

Max, if you look at what the government's doing around AI growth zones at the moment, which is kind of facilitating of the back of that because, sorry, using this conversation to get kind of, separated from the actual kind of output of what you're talking about with AI.

And AI is not just chat GPT, but kind of drug discovery. It's about building digital twins and nuclear reactors and things like things like that as well. So what the government is doing with AI growth zones is a very interesting way of, connecting up, those those sites, and I think they've they've just kind of closed the first round of expression of interest. So a lot of places in the UK have had an opportunity for kind of regional growth, particularly in, in the those postal industrial areas that might have, excess energy. And by the way, AI AI, like, data centers are really good to put on the, back of unreliable supply of energy as well because you're if you can do something with that excess energy rather than turn off turning off kind of wind turbines using it to kind of, train models or kind of do inferencing in in data centers is a good way to kind of actually optimize the output in a in a in a long tail.

And maybe kind of moving away a bit from from nuclear in terms of renewables, is it gonna be either a stronger also going forward, or is it kind of an even longer time frame in terms of kind of using renewables for AI and data centers? Lucy, do you wanna take that?

Yeah. I mean, I I would agree with the previous comments that I think nuclear is a is a ten plus year play, particularly for sort of traditional, large scale new nuclear developments. I think we're unlikely being very realistic, we're unlikely to see anything in the West sooner than a ten year time horizon.

I know there's a lot of interest in small modular reactors.

I think that there needs to be more evidence around the commercial viability of those. Because they're smaller, the output is is lower, and therefore, the cost per megawatt hour is not necessarily currently as, commercially competitive. But, like I said, I think a ten plus year time frame, these are very interesting, and they will play a role.

In the shorter term, I think, solar, wind, and battery, battery storage, will be important. So we talked earlier about how we're already seeing increased demand for AI for data centers right now. So this is, at a pace of urgency which is outstripping the pace at which we can get new nuclear on stream right now.

For things like off grid micro solar micro grids, these are the types of things that we have the potential to complete things like site acquisition and permitting and have a full operational site maybe within a few years. So that's something that certainly could be, a useful means of powering data centers and AI, in the nearer term.

There have been some interesting studies, done looking into this. One that, is in my mind was one recent study from Duke University.

This looked at headroom on the grid in the US, and I think that came up with a figure of something like a hundred gigawatts that could be connected, could be used to add additional, data centers potentially with pretty much immediate effect, with only a short small amount of flexibility from those data centers needed to support.

And just to put that into perspective, that hundred gigawatts is more than the current total installed, nuclear capacity in the US, so it's not, trivial.

So I I think there's a lot of hype about nuclear at the moment.

I I certainly hope some of it is justified, particularly hype about fusion.

But I think we need do need to put this into perspective, and recognize that solar wind and battery certainly has an important role, in the nearer term.

Is it justified? Jeffrey, are you nodding?

Well, I mean, I think it is. I mean, the companies themselves are clearly signaling that they that this is something that they're they're interested in or that that that they want and that they're investing in. I mean, we've all heard the we've all heard the news about the major tech companies in the United States making investments in in, advanced nuclear companies. Those are those are investments that will help, you know, get those, those technologies eventually commercialized. It's not gonna happen tomorrow. So that's important for the the development and the deployment of those technologies, not for meeting an immediate need.

But, you know, there I mean, a couple of things. I mean, at the IEA, we we help countries, that don't have nuclear power develop the the the the infrastructure to to support a nuclear power program. So, you know, setting up a a regulatory body, human resources development, a a number of different issues. And we're getting calls from countries that, that host data centers, that have a have a have a big need, and they see this obviously growing in the near future. And they wanna know, hey, how can we set up a nuclear power program? Because this is something that, you know, we we we envision a a pretty big need. So I think that right there is is is is a signal.

In terms of, like, using the existing nuclear assets out there, we've also seen in, you know, like in the United States, some some companies directly, you know, doing PPAs with to restart, already retired assets like at Three Mile Island or whatnot. But also you mentioned you mentioned France. I mean, you know, that's an interesting case where, you know, they have a huge nuclear fleet, but, there's there's a lot of there's a lot of room there to increase the the capacity factor of those of that fleet. I think it's maybe at about seventy percent or something.

I might don't don't quote me exactly. So, you know, that that could be, you know, through through, through, you know, taking taking, a certain, you know, measures on those reactors. You could get it up to, I don't know, you know, get it up to eighty five to ninety percent capacity factor, which could, you know, which could could support a a a a, you know, a lot of demand. In addition, a lot of those reactors could be upgraded, I understand, by a few megawatts.

So, you know, and then there's the existing technologies that aren't that are, you know, already licensed nuclear technologies, the large reactors, that that that we know about.

You know, I think we we heard about Stargate, this, one announcement in the United States, you know, where the I think it was five hundred billion dollars in in in energy, infrastructure.

You know, I was recently in San Francisco talking to a big, tech company, and and, you know, they were very clear that, you know, this is, this is gonna they're gonna need natural gas for this right now. And, you know, in in some places, you know, coal assets will be extended to to to to, you know, support the grid.

But that, nuclear still for them is very much something that they want in the medium to long term, and they need to start working on that now to make it happen. And, I wouldn't be surprised if some of them are looking at at also their already existing, designs that, that, where the supply chains are more or less there. And, you know, if they get working now on those, you know, in in five to ten years, they could they could have them perhaps online.

I think one of the other observations links to things that various people have said, is just around what the what the hyperscalers are doing.

So Microsoft have got creative with the deal with Constellation, which is really good news. But Microsoft, Amazon, Meta, all have been traditionally massive users of PPAs with renewables, and they've all, I think, signaled a direct a direction of travel, which is that, actually, having been unpicked by various people, that isn't quite as green as they might have hoped It's where they see an opportunity with nuclear. And I guess this probably was when nuclear come in. The AMR, SMR proposition is actually of a size where you can more directly collocate those, and you can take yourself out of the planning nightmare that is getting permission for grid access, etcetera.

I think those sort of direction travels clear in their minds at least.

Yeah. I mean, what I would add is I don't think nuclear is and should be competing with renewables. I think they're complementary. I think we need all of them. And I think if you look at the scale of the the growth in in energy to deal with net zero generally, not just data centers, but also, you know, the hard to abate sectors.

We we should be looking at everything we've got, and I think we need all of it. So, for me, let's not debate which we should have. We should get on and and and meet the energy needs of the future.

Actually, the one issue nuclear is dealing with is is dealing with the the kind of legacy of the industry, part of which is kind of fear, part of which is political challenges, part of which is stop start. So if anyone's seen Nuclear Now by Oliver Stone, it's a good a good watch to give some history of of how nuclear is, where it is. But I think, you know, it's for the developers like like mass like, you go to to demonstrate that we can build out that we can do it cost effectively, that we deliver a competitive l c LCOE, and we we need to get on and do it. But, again, we're the economy of many, not the economy of scale.

And so to deliver that, you have to start. And and again, just to use another simple analogy, if you look at we're in the UK. Right? You look at Hinkley Point and and and a significant amount of of expenditure on one sorry, on one site with two reactors.

Right? If the first one goes fifty percent late and fifty percent over budget, it's a massive headline, huge, you know, possibly tens of billions of of pounds. You build twenty reactors, they reach a billion, and the first one is fifty percent over cost and fifty percent late. It's about half a billion, and no one's talking about it in a program of twenty billion.

So you actually just need to believe, get on, do, learn, and deliver. And and that's really what, what SMRs and AMRs are about.

Thank you. I mean, let let's maybe move on a bit in terms of we've touched on it a little bit already in terms of what can AI help us do better. I would like to I'm not quite sure what I believe it is, but I'm still gonna gonna say it. Apparently, the American Council for an Energy Efficient Economy estimated that for every kilowatt of kilowatt of energy consumed by the IT sector, ten kilowatts were saved in other sectors. So we might see more energy used by AI, but everybody else is gonna be using this as a it's a net positive. So maybe if I start with you, Rob, because you've we talked a bit about before in terms of what do you think in terms of kind of the right framework to optimize that kind of energy demand going forward?

There's a lot of ways lots of ways to to slice that one in terms of the the numbers quoted. I'm not I'm not entirely sure of the context, but, I would think that that could well be true of of kind of business and and office operations.

AI has a lot of possibilities to automate some some very manual human workflows, which, yeah, frees up time. And, you know, I'm over I'm aspirational about that. If AI can save time from sort of human toil, maybe that leaves us all with with more capacity to think about important big picture things if the AI AI hasn't also automated automated that away.

But my my view on on on on the broader topic there is is from the end user perspective, and I guess the nuclear perspective gives off the other end the the the chain here. But, in that world of of someone renewables heavy dynamic grid, there's a question of let's assume that AI will be doing kind of scheduling flexibility, moving a lot of load around to just kind of smooth out the, the output profile of renewables generation.

Those systems are going to be very important. They're gonna be critical infrastructure.

And I think what anyone who's who's done a done an AI, or machine learning kind of project development knows that the AI and one thing, but the data sources you're feeding it with, your training within are are critically important. So I do think, as well as the the sort of market side changes I mentioned earlier in terms of, you know, time of use tariffs are a great way to incentivize this, assuming that, you know, careful work goes in to get the right price signals. You need meaningfully kind of split up zone on market, so you're you're capturing kind of the values for those renewables, for businesses that want to want to go there but can't just, like, commission their own power plant to to co locate and kind of bypass the wholesale market. I think that's important for for medium scale and kind of widespread adoption.

But I I do think what we what we should expect for in some ways, this is the easy bit, but it's it's bit not to forget is we need to make sure that the the platforms that are producing data on grid generation mix, on pricing under whatever tariffs and regimes and regulations exist in certain markets through to kind of more technical details like, transmission network frequency, the kind of things you're aiming to optimize on or respond to in terms of flexibility service.

I think we'll we'll we'll start seeing those more as kind of critical infrastructure in a way that we don't really think of it that way when most flexibility is controlled from a control control room environment to kind of centrally dispatch. So I think there'll be a change to more open distribution network there. We need to think about the kind of the data connections between all the all the pieces.

Yeah. Max?

One of the things that that's been most interesting to me is, we've mentioned this a couple of times about digital twins. So we talk about AI and the kind of efficiency from a kind of a, a workforce level. But if we look at it in terms of the place that people work, both in kind of offices, but in, industrial kind of factories, digital twins have an enormous potential to help design better processes. Like, you look around this room, there's, like, ventilation systems, there's there's lighting, and that was probably when there was nobody in here that was, the same systems were running.

And what digital twin of this room could do is kind of use actually Gen AI to put you all in or kind of versions of you and, and then kind of model more effectively, like, the energy use. And you can see the kind of then you can start to see the efficiency potential, across a range of things, and you transfer that small use case to, industrial processes. And, actually, NVIDIA has done this in a a factory with, in Mexico with Foxconn, where we built a digital twin of the factory, and we do it with other areas as well. And we saved, thirty percent of, the kind of annual energy use, and that's gonna increase all the time as you optimize the processes.

And what AI allows you to do, this kind of greater scale of use of, computation allows you to make more accurate predictions, bring in kind of weather models as well to understand how that might change. And so therefore, this kind of base level energy use that we might use, you cannot start to see it kind of optimizing it. So there was a great potential. We talk about the day a data center use of, AI as being about two percent growing to six percent, but there's a there's a kind of ninety four to ninety six percent of energy use in there, which is very carbon intensive, which has the there's a huge potential in there.

Lucy?

Yeah.

So just just to add to that, I mean, I I totally agree with the sentiment in your question. So, we have to be really careful not to make AI the the bogeyman or the bogeywoman because some of the downstream applications, you know, may actually have the potential to, as you suggest, really, return orders of magnitude in terms of emission savings.

I think one good example is DeepMind's AlphaFold. So, DeepMind estimate that this their AI model that looks at how proteins fold has saved something like a billion years of, of of of of alternative research.

And we have to remember that the the counterfactual for a lot of this stuff is not something that would probably generate no emissions at all. It would still be something that's generating some emissions. So you have to keep that in mind. And just to build on Max's point about the digital twin, one of the things my organization, has developed is a predictive AI model for energy demand.

And, so it's called Faraday.

This is a model it is a generative AI model, but for anyone technical, it's slightly older school gen AI. So it's a variational autoencoder combined with a Gaussian mixture model.

And, we this is important because we need very good models of what's happening on the demand side, good models of what's happening on the supply side so that we can optimize these at the grid or the energy system level.

Both of these things, supply and demand, are becoming more complex and changing more because of our changing energy system. So, even just a small improvement in the accuracy of a predictive forecasting model, is is very powerful and important when you put it in a real world grid system operator context.