With Laurent's Segeleen from London and Gerard read from Berlin.
This is redefining energy today.
On redefinding energy Jihad, it's a very hot topic. We're going to talk about AI data centers and the cost of speed. Yeah.
I like the way I say that the cost of speeder on because it definitely does say that we're in an AI race and the winner seems to be the one that's the fastest. At least that's what the industry thinks anyway.
The first award from our new partner. Today's show is supported by the BMW Foundation, Harvard quant The BMW Foundation unites leaders from diverse sectors to develop solutions that foster an innovative economy and a future proof society. A key focus is the energy transition and climate change, where the
Foundation drives international collaboration to accelerate the energy transition. With rising energy demands from AI and data centers, new partnerships, effective collaboration, and the exchange of science based solutions and strategies are essential. That's why the BMW Foundation supports this podcast and bring these discussions to global stages by hosting the Energy Security Hub at the Munich Security conference in twenty twenty six, streaming live from February twelve to the fourteenth.
You can learn more at the BMW Foundation dot org.
Back to the show. Yeah, and of course the greed being on decade long planning, and all of a sudden you get three mega what's coming in one giga? What's coming there? And the grid simply cannot follow, which means that it's all in on deck and there's a lot of things happening behind the meta. And you know, we're going to talk in details about those OCGT open cycle gastar binds or maybe fuel cells and batteries. Basically they're throwing everything they can just to get some energy.
It's a really offsy topic to talk about, so and it's great to actually bring an expert on that knows the space very well. And we ended up finding Andrew Perry, who's the director of the Energy Transition and Environmental Business Unit at Faculty AI, which is leading consultancy businesses in the AI space in Europe.
Yeah, and the interesting thing is that it's coming from the AI world and it's not coming from the nuts and boards world. So he show us the moon, we look at the finger. We have a very good conversation. So let's bring on on de Perry on the show and the welcome to the show.
Pleasure to be here, thank you for having me.
Andy. Maybe let's jump straight in. I mean, what we're really going to do is talk about AI and energy, try to get a better understanding of that. But maybe just talk a little bit about what's going on in AI. I think that'd be a great starting point.
The technology development is, if anything, progressing quicker than what people have been expecting.
Recently.
There's something called the International Mass Olympiad, which is like
the premiere high school competition for math. And if you looked at the bedding markets, so like the kind of predicted markets, the prediction was abound twenty percent that AI could could win that competition in this year, so twenty twenty five, And actually it won that competition in July, and then a month later it then achieved a perfect score in the ICPC, which is the International Collegiate Programming Contest, which is basically where top university teams come together and
solve like complex algorithmic problems. This was GPT five. I thinkam GEM and I both were able to achieve a perfect score of.
Out twelve, which none of the human teams managed that.
Pace of change listening incredibly quick and so it's beating benchmark every which way. I think the benchmarks effectively meaningless now because because all the benchmarks effectively and decent, And so when you look at the rate of progress and you look at it over time, it is possible actually to understand it in a way that doesn't always seem possible. So it sometimes feels like it's just this exponential thing,
like how do we get our hands around it? But if you look at the last of six years, there's a really interesting study by a Meter, which is a nonprofit research organization, which tried to look at how long it took humans to complete tasks that AI could now do and how that's developed over time.
If you look way back.
Sort of six years ago, with GPT two, we were talking about tasks that took a few seconds that AI was able to complete, and we were quite impressed by that back in twenty nineteen, and then GVT three a couple of years later, was able to do sort of ten second task. Then GVT four, which you guys will remember that was in twenty and twenty three, you know, was able to do sort of five minute tasks, but we're now up to the point where we're able to do sort of one to two hour tasks with AI
with a fifty percent success rate. So what you then see when you look at it that way is that you have this sort of doubling of ability every seven months or so.
What you're actually saying is that AI what could take the human two hours AI is doing in five se That's what you're sort of saying.
Yeah, fairly instantly, it's a now.
So what's interesting about that is like you've basically had a model back in twenty nineteen that was like a preschooler in its intelligence, that then developed into more like a primary schooler and then became like a high schooler with GVC four and with GVD five is kind of like a genius high schooler that's now winning the Mass Olympiad and is able to complete these coding and programming
tasks that the university level graduates are struggling with. You're seeing this evolution of the models that it's kind of progressing at about three times the rate of what a child would in its development, which is really impressive looking backwards when you start looking forwards and you think, well, okay, if you take that progression that has been pretty consistent for the last six years and play it forwards, you start getting into some very interesting, expiking gary types of
scenarios of what can happen in the.
Next three or five years.
And it helps to explain a bit some of the what seemed like fairly crazy investments and predictions that being made right now in the market of like where things could.
Go, because at the end of the day, it's probably a bit of a winner take hold. Yeah, and I know that there are probably five or ten big guys literally putting everything they have hoping to be or do not know the Microsoft or the Google of tomorrow factor ten, But of course there's more applicants than final winner. How do you see even the structure of the industry like new player, more of the same or some which you throw the towel and say, okay, fine, what's the feeling right now?
I think you're right that through the winner takes All dynamics is definitely a play part of that is because there's a perception there's a sort of a three to five year window looking to like twenty twenty seven, eight twenty thirty, where if you get the same levels of development that we've seen so far, we will achieve some sort of AGI, so artificial general intelligence effectively AI being on a general basis at the level of the smartest humans. If you get to that point, then you start moving
into a position of potential platform for superintelligence. So one of the reasons why you see this dynamic right now is because everyone's chasing this dragon, racing to be the first to get there or to be one of the first to get there, and that explains probably two things. Explains the massive investments in compute, and it explains the massive investment into talent, and METSA is probably the best
example of people who are doing this. So the hundreds of billions going into computational infrastructure and hundreds of millions
and billions being spent on talent. The reason for that is because if you look at what drives development and what will drive these improvements over the coming years, and the sort of order of magnitude improvements that we've seen over the last six is primarily compute and investment into that, and we've seen as sort of a five X on Moore's law in terms of computational improvement and algorithmic efficiency, and that's where the talent comes in, that's where the
brain power comes in. Where you have a similar impact to computes. Probably it's probably underestimated in the main, but it's actually probably simlar to compute where algorithmic improvements are improving the performance of the models, improving the efficiency of the models, improving their ability to use the compute, and
therefore improving their performance. So right now you have this fight for those two things to move as quickly as possible, and because of the trend line, if we can pinue on that trend line, then it's a three to five year window to get something that people would expect to look like an AGI level. That's really interesting in itself as well, because then you get this platform twowards well.
If you're at that level, and especially if AI is good enough to do AI research as well as the best human then AI can start becoming self replicating, it can start producing the next AI model that will be beyond a level of human capability or sort of understanding. So I'll say it's very exciting and very scaryly very quickly when you go down that road. It could also pluitow, and there's reasons why it might not develop at that pace.
I can see that your head is in the stars, our boots are in the mud, and we are the nuts and balls guy. And a lot of our listeners are from the infrastructure world, and of course they're going to invest into those data centers, providing equity or that, and hopefully the principle of infrastructures is like they like to see their money back now. Of course I hear, it's fine with German paddocks. We build all of that center, we'll have use for them. And by the way, the railways,
the internet, YadA, YadA, YadA. But when you would build a railway, the alway will be there for the next twenty thirty years. So even if the trains the first three four years were not there, they would come over time. The second problem I have is the obsolescence, because I see the chips moving so fast in terms of improvement that I hear that the data center with chips which are three or four years old, you can literally scrape
it and almost go back from zero. So you've got the obsolescence factor in the technology that you didn't have in the previous let's say, crazy cycle of bubbles of infrastructure. What's your take on my thinking right now?
I think you're right when you look at the GPU side of this, there's definite obsolescence risk. These things get replaced every two or three years, so hard to reteep
that capital spend if you're not using it. What's interesting though, right now, I think with ai and I'm not a financial expert, but when you look at the annual run rates that Opening Eye is operating, that Anthropic is operating, and the speed of improvement of those run rates, you do see some level of justification even for the numbers that they are being valued at and are investing. So I think open Aiye was at a twelve billion ARR quite recently and was increasing at about a billion a month.
Anthropic was similar, so they'd achieved like a five billion ARR and again was improving about a billion a month.
So when you compare to say.
When people talk about terms of the AI bubble competulate the dot com bubble, it's like there's very real revenues being associated with a lot of the activity that we're seeing, which I think it makes it fairly unique. The pace of development that's being evident in actual usage. When you look at then the levels of investment that they're putting in.
If you believe in the basic kind of optimistic scenario, or you think that optimistic scenario is likely enough that there's a decent chance it could happen, then if it's someone like Mark Zuckerberg, it makes sense to, as you said, drop potentially two hundred billion in investment, even if that may not come through, because the risk of not being the winner in it is too great. There's all kinds
of financial shannagans going on around this. There's the fact that in Vidia is investing in companies that then use this infrastructure, and you've got this kind of circular capex economy. You've got things that you guys will understand. But the me to be honest, I think in terms of the investment vehicles being used, but there does seem to be a stronger base to this than other capical investment programs
that we've seen. What's really interesting for me, I guess for this podcast is like the implications through energy and the fact that energy really is a bottleneck to this development and will be absolutely key to the success of being able to build these eighty centers and the way that these set companies are now becoming major energy players and also quite independent energy players, maybe even separate from the grid.
And it's good to talk about energy. So what I'd like to get your view on and just explain what's going on here. If I take my iPhone, it uses maybe two killer one hours of power over a course of a year, and I'm charging it every day, right, and I use it over the whole cost of the year. And then I go and look at Video's new blackweld chip and this uses fifteen kile of what hours of power per day and actually the size of it is
the size of a credit card. So just explain to me what's going on there, because obviously if I think in terms of relation there, I just can't fathom that. And obviously that explains then why there's massive need for power going for Explain that, right, what's going on there? Why suddenly these chips need so much power.
That's rather than focusing on the chips themselves, maybe thinking about like the size of the data centers that are containing them. Obviously, the more power of the chips, the less you need for your data center and the more efficient you can operate. But the reason that I think the data centers are so big, there are free factors
that really drive AI progress. When is this computational capacity and there are order of magnitude imprevements that we've seen so far and order of magnitude improvements that can be unlocked through these investments. So it's just a huge driver of the power that drives then your ability to move quickly in AI development. The second is around algorithmic efficiency
and so the application of brain power to it. And the third is the way that these models operate and how they become unhobbled by being able to either have memory, or to be able to use tools, or to be able to operate on an agentic basis, or to reason deeply and think the wait that basically the model can enhance its own capability through that kind of unhobbling of
the way that it operates. When you look at the power usage of these chips and the power usage of these data centers, it's purely a reflection of the fact that there is this raw muscle power that can drive AI progress. It's the seed at which you're able to develop and train models, and then the inference power you need to be able to run them, which is also massively increasing.
Myself and Ron probably very much agreed. There seems to be just this crazy frenzy in and or around the view that we're going to lead not even just gigawats power, but one hundreds of gigawats of power to fuel this AI race that we're seeing. And we're obviously seeing CEOs or businesses, and we're even seeing financial institutions say things like, well,
we'll never meet this AI gap in our lifetime type stuff. Right, So, how do you view this and how do you try and help us get to sort of think through and understand what's going to happen.
The way energy is becoming so key to powering these data centers is just like transformational, and it's just increased so quickly in the last couple of years. And so you say, we're now seeing I think of a move into the realm of like ken gigawatt plus sort of data center clusters as we go into the next few years.
We already have like one gigawatts in places like Abelene in Texas where Stargate's getting developed, or the Colossus cluster in Memphis that the XAI is developing, where you've just got immense amounts of energy usage that's in the hundreds of hundreds of megawatz, just incredibly quickly that's been sput up.
The difference in these times between tech progress and what we know in the energy industry is kind of energy development progress, where in tech world you're moving quarter by quarter, and energy infrastructure you're moving in many years or decades and to get stuff built, and these two things clearly don't compute when you try and make these two things together. So we now see companies just taking it into their own hands to develop their own generation on site completely islanded.
I'm interested to understand maybe what's happening with permitting there and how they've managed to get around rules around this. But when you look at the different types of generation that they have with options click, clearly there's a big investment in nuclear and SMR development that feels like there's massive hurdles to cross from a regulatory point of view to get new generation SMRs developed that will take ten
plus years. Renewables are complicated, and they're cheap and they're quick, but the complexity of trying to provide base load power with them complicates matters. When all you want is on demand power if it's completely reliable, which kind of brings you back to gas, and obviously at a time when the US is swimming in gas. But even cggps have a pipeline of five plus years before you can get
on the list. So they're now reverting to OCGT. And so I think in a stargate there's something like ten to fifteen of these operating thirty five megawats each in XAI facility again talk of thirty to thirty five of these are reoperating. So you've got like hundreds of megawatts these three or five hundred megawat's worth of open cycle gas turbines operating that I guess we'd normally expect to only see running at like a five to ten percent max load factor.
That's being on the spaceload.
Yeah, you're one hundred percent right. All those guys they made like all their climate pledge, and we're going to be in net Zio and the first thing they do is they put and I can even tell you the brand. It's all the same. It's the Titan three fifty by Caterpillar. And you know what what is great with the your administration is, you know andamantal law no onenviromuntalo anymore. You want to pollute, please pollute. It is amazing. For the
sake of speed, the laws don't apply anymore. The only thing people are complaining about is when you start linking to the grid that it's kind of mixed with the great payer and I know there's going to be a backlash. I empathize with the necessary speed of development to reach Agi. There's a thickness in the way infras being built which you almost need to reverse engine. What is the grid
able to deliver to me? And then start thinking not about nuclear because nuclear is going to come in ten years and it's gonna cost you three hundred dollars and make what I wur now if you're willing to pay, that's good enough. It's not gonna be now, and it's not gonna be cheap. It's not going to be cheap.
If you look at the basic numbers. Why do company like Oaklow or Fermi who literally have zero project, a bunch of slide, not even a product, and there are value twenty billion dollar that is for me part of the frenzy. And there's going to be some tough awakenings because I have no doubt the Microsoft and Amazon they'll continue. But there's certain number of things in the supply chain
which are for real and stuff which are ballooney. The question now is, knowing that the limiting factor is the energy, is, are they not going to rethink the way they do data centers or compute our software all this based on this limiting factor.
I don't think so. Clearly.
There's efficiencies being made all the time. I know there's been a lot of efficiencies made on water usage by having close cycle kind of water loops, for example. But I think it's a matter of, like everything, it won't be a case of, well, we made the efficiencies, we need less power. It's just like we've made the efficiencies, now we can get more out of that power. They will continue to build more and more because we have
this race that's a three to five year race. I guess the logic would be, well, look, it's some short term pain. We would like to do this in a climate friendly way, but the reality is we can't. We can't move at the pace we need to, and it's a national security concern, et cetera, et cetera. But you know what if we can get to that point in three to five years, then the machines will fix this for us. We will find out solutions at that point.
The tech is for the longer term, and this five year period will just be a short plit up until the point at which we then have a level of intelligence that means we can we can solve a lot of these problems. And I'm not saying I believe that, but I think that would be a justification that would get used.
Excellent No, no, sorry, no sorry, I mean Jiard you say excellent, but me I don't say excellent because you know, a few take billionaires said you know what, I take the cash now, and it's a short term paign. But not worry. The future is going to be better. Leny used to speak like that, but it was the Communist Party, it was not the tech bros. But at the end of the day, it's a minority of people who are capturing the common good, which is our oxygen, and say,
you know, the future is going to be great. Sorry, in the meantime, I need to pull youute like crazy, but you know it's for the greater good. I'm sorry you think it's great.
I don't you can look like that, or you can look at it as an opportunity to use the fact that what we've got is power the man going up for a start, and secondly, somebody who's willing to put large amounts of capital to work in energy infrastructure, and that gives you the chance to really really innovate. And I'm also clear in my head who wins. The winner is those who get electricity quickly to market, and so
that pushes innovation and drive solutions. And for me, if I look at it, yes, there's no doubt it's small gas engines, but you're put the gas engines besides solar, wind batteries, et cetera, et cetera. And you're also going to flexibilize the way the data centers work. And I think Lron, we've talked about this before, is they're already doing this anyway, because at the end of the day, when we in Europe and where and during our working day, guess what we're using servers in the US? Why because
it's capacity over there? And guess what we got to bed they're using our capacity over here. So we're doing this or there is a flexibility here already. And I think then we actually actually embrace these technologies to enable us to electrify in a better way. That's my positive science.
Yeah, but child, I agree with you. If the government's there, which of course they're not, put some technical and enjunmental constraint, they would start being intelligent. The moment is where okay, you want to put fifty ocgts. Be my guest they'll do it, because that's the easy solution. If they were more constrained ehether they would move elsewhere or not known in the north of Scotland, or they would put the
data center where there is load. And now they say we can manage our own fleet, so that's number one. Number two they would say, we need to put more effort in our softwares or our chips or the design of our model, so it's much more efficient. And if the fact that you can get all the energy you want and you know, as polluting as possible, now I think that's intellectual laziness. Sorry, we can debate, prove me wrong.
I kind of see both sides of this. If you believe in the way that tech companies do in this suite of five year timescale, you can absolutely see the need for me quickly, and the only option they have
is basically gas gen sets in the short term. It feels like though the investment potential and the longer term investment potential from these guys, with the amount of capital they have and the amount of will they have, do incredible things to help with the energy transition, which is kind of stalling in some ways, I think in the West. So the question is like, how do you combine those two things together, and how do you put conditions in place that force that medium to long term investment in
cleaner solutions in new technologies alongside shorter term measures. Seems to me it's like we're kind of in a singularly bad political situation right now to kind of force that
to happen. Maybe five ten years ago, it would have been easier to try and make this trade off work and perhaps put conditions around the use of these OCGT engines that you have to also correspondingly invest in technologies and have a un the comparent some kind of roadmap for how you're going to develop your energy production, which I haven't.
Seen any of.
Having said that, there have been investments clearly in nuclear, there's investments into SMRs, there's investments interfusion even that these companies are making, how far like those are realistic and you see a realistic kind of roadmap of one going into the other, or some way of understanding the relationship.
That's really unclear. It's like really unclear.
And there's clearly regulatory input you would want into the way these investments are being made and into the way that these things are being developed that we're not getting. And I just don't know how the permitting's being done and how the things that EPA in the states, like what position they're taking on this or how they're operating. I know a lot of stuff is done at state
level as well. I don't know if you guys have any insight on that, but it seems like these things have been put up incredibly quickly and probably without any oversight on the environmental concerns of it.
Well, and it just maybe to wrap up, could you talk a little bit about what you actually do.
We try and help companies and organizations adopt AI in the right way. I mentioned at the start kind of adoption gap between technology progress and an actual usage. The whole bunch of things that drive that gap, and it's not that the fact that the technology can't do a useful things. It's because it's really hard to develop user centered products and tools and solutions within organizations, and that
involves a whole hybrid of skill set. By what we do is a business, we bring those skill sets of data science, machine learning, engineering, product management, and technical architecture, commercial delivery, business case understanding all of these sorts of things, and we operate both in the generative aispace. So maybe half of our workers a businesses how you deploy large language models in useful ways within businesses, how you create
solutions around them. But also half of our work would estimate is more like traditional machine learning type solutions, so self built machine learning tools with very specific purpose. And actually a lot of our work and energy is more
in that space. So while we've talked about the generative world a lot here is actually that a lot of the value of AI within energy sits within the way you apply more traditional machine learning techniques to specific types of solutions, things like how do you use flexibility effectively on the network by forecasting your demand in generation better? How do you schedule dispatch of generation in the system, how do you manage the you've got like a thousand
plus balancing mechanism units rather than tens of them. Now you've got this complexity to the where you plan and
run networks, the way you operate them. How do you prioritize connection cues when you've got just masses of generation and new demand queuing up to connect and trying to understand how you prioritize that and design interventions that respond to how it impacts your career, a non redistection on networks and predictive maintenance in generation, and how do you manage customers better as an energy supplier all sorts of things across generation, networks and supply where we can make
things not just more efficient and more productive, but optimized, and so a lot of the work is run how you make better pocisis when I look at the energy transition, and I've been in the space of fifteen odd years. Now we're really at that point now where we have
generation of scale. A lot of the challenges how do you integrate the stuff at scale in a way that optimizes the way the network operates, where the system operates and prevents us from rebuilding and duplicating infrastructure because we're able to use what we have efficiently, and I think that's what machine learning has just a huge part to play.
Well, honey, thank you very very much for this world one tour of AI being very enjoyable.
Yeah, sorry for sometimes the level of the debate. But I am convinced AI is going to change everything. But I see stuff which are not normal, and I don't want to talk about the valuation of anthropic or whatever. I have no clue, but it's more inside the supply chain along the way. I'm always wary that a lot of investors are going to lose money because they look for the moon but they don't concentrate on the finger. Yeah.
The fact that is that we're over around in a few years time and we can see who was right and wrong. So all of these I thought the things are gonna get tested exactly.
Thank you so much for coming one.
Thanks for having me. It's a lot of fun.
So what's your takeaway?
My takeaway is when you discuss with AI, guy, they have two or three magic words to make sure that you don't really understand what's going on, and I have them. The number one is inference. They say, yeah, yeah, I like your solution, but what about inference? Of course, I have no idea what inference is, so you can of okay, fine. And then the second one is Argentic model. First I thought it was like Argentic, like but from Argentina or no, no, no, no,
it's Argentic. And then so of course I had to go on AI to kind of figure out it was. And it's a totally different universe and they're just rushing, rushing, rushing. They have really no idea where they're going, but they don't want to be the last, so they throw everything at it. And that's what makes the old conversation fascinating.
What I take out of it, I suppose, is it's not a race for AI, because we have a AR right in lots of different forms. It's really a race for artificial human intelligence. That's really what it's about. And there's a mixture of fear, oh my god, what if we don't get it and the Chinese get it? And on the other side there's also greed, which is if we get this, we're going to rule the world, or we're going to be the top tech company in the world.
That's what's going on, and it's driving crazy evaluations. That's what I take.
Out of it.
Right when I'm looking at I looked at one of the utilities the other day and there's the US independent power producer and trading at seventy times earnings. I've never seen a utility and independent power producer trading it's seventy times earnings. You know a bull market. Prices are up and you know economy is booming, and maybe at twenty times what's seventy times earnings? Mind bo So there's crazy stuff going on in terms of evaluations as well.
Right, okay, job, the big question is is this all ai RA is going to reshape the power grid? Or will the inertia and complexity of two day's greed will deliver some revolution Behind the matter, we'll see how we're going to build in America the equivalent of two hundred nuclear plans within five years to fuel all that potential demand. Very difficult to say, but it's definitely fascinating universe.
Yeah, Aliston, I want to thank and Andy Perry for coming on the show. It was really great to have his view on it. And that's very different. No, London, let's say maybe the view that we'll have in the energy space right totally.
Well, finally, we'd like to thank the BMW Foundation A Berkman for supporting the show and job. I'll talk to you next week looking forward.
To thank you for listening to Redefining Energy.
Don't forget to rate the show and subscribe on Apple Podcast, Spotify, or the platform of your choice.