Casper Sønderby (twig) | AI & The Future of Power Trading

Hello and welcome to episode 48 of the Nordsearch podcast where we speak to leaders in the power trading industry. So today I'm delighted to welcome back one of my personal favourite guests back on the Nordsearch podcast. It's been two years since we last spoke. So Casper, Sonneby who is the CEO, co-founder of Twig, welcome back to the podcast. Thank you. It's my pleasure. Yeah so as I mentioned it's been two years Casper and obviously we caught up last week and a lot has happened in those two years. It's really impressive what you've managed to achieve. Well I guess for those who didn't watch the first episode it would be great just to have like a short reintroduction to yourself. And that's where and then we can talk about all the great things that have been going on in the last couple of years. Sure. Yeah so yeah my name is Casper. I'm CEO and co-founder of Twig. My background is in engineering so like did a like machine learning and computer science degree and then eventually a PhD in machine learning like catching the early days of deep learning and then I moved to Google and worked there for few years in the research department and eventually ended up working on weather forecasting there. So we imagined all these AI forecasting model that are now arriving in power sector. We basically did a lot of the early work on this and basically replaced the physical model with a machine learning based model. Then after Google I moved back to Denmark and a little bit by coincident ended in the power trading business. Even I actually had accepted a job of for moving to London to work on self-driving cars by a great company in London but then I mean for family reasons I decided not to pursue that and then we moved to Copenhagen and Denmark has a strong position in energy and that's really how we got started. Yeah 2000 would be proper as well wasn't it? Yeah so my brother he was actually in in in Cupertino working for Apple. So he came back more for I mean he was like it was during Covid so there was like I mean they were just coming back more for like to avoid being locked up in an apartment in in San Francisco and then I mean we started talking about power trading and you know visiting power traders and I mean we saw a pipe in the sense that we care about computers so we thought we could automate a lot of the operations and and maybe that was a little bit the you know optimistic of how easy that would be but we had this like little napkin clan that we could automate power trading and then we could start doing asset management and eventually actually take part in deploying assets. So that was kind of a napkin clan and taking a little longer right than you think but I think we have state relative to that. Yeah I know five years in I think you're proven you're executing that plan pretty well. No it's funny so you were an AI researcher before it was cool and trendy and mainstream. It was almost being cool. Yeah yeah but I mean I was like wondering I mean yeah the language model was not a thing right so we also worked on like actually some of the the early versions of how chat TVC was developed in the lab I was in Google it was not by me but by someone else. Also some of the early works on diffusion models is really interesting to see how some of that scales right because I remember the way that you train a large language model today it's kind of a very it's not clear why that approach would work and I just remember we have this discussion in the lab about like that someone was proposing that you could train on text only and and just like do that and and people are very skeptical that I would work and and now we're like five years later right and it just take you know the world. Yeah no for sure no because I're Google with the with the front runners on AI at work when they and I read something that yeah you have Google had the capabilities to launch a chat GPT type brother I mean it just didn't just didn't ship it yeah and then chat GPT ship to it and then here we are today with model like large language models and text only being what everyone uses every every every day yeah yeah so so yes that's what that's why I always find it really interesting about your background because it is so different to the large majority of backgrounds of of our trade-in founders so I think because you've got that different background you you see things in a different lens and you try to do things in a different way and it's really impressive now to see five years in the results that you've managed so far and obviously you released your annual report last week and how is extremely impressed with the numbers so could you give us a bit of an overview of you know where you are today I think that would be interesting for people to yeah so maybe I can I can just like go back and then say that we started like just pre like we were probably live in in the power markets like just pre Russia's in relation of Ukraine so there is a very abnormal time to start and we basically had the zero background in power markets when we started so we didn't really know what to expect so first year was a little bit like a wild ride but since then I think we have basically been able to prove that we can by improving our machine learning models increased the amount of data we have increased the compute right and then have a like strong conviction in that that will eventually work out then I think we are basically seeing the results of that now five years in that that we can we can just scale everything and it just gets better but where we are right now I mean on a kind of more like commercial side I think like revenue a bit I know that are growing pretty fast so like almost doubling the last few years every year which of course we are very happy with but I think also maybe on more like on concrete milestones so we entered all of Europe a last player on the entry day market in Europe we are in pretty much all of the US and then we recently that's actually almost two years ago in Japan so we are in those market as a trader and then on the asset management side we are active in northern Europe right now and Baltic states and operate about 100 assets in those markets various kinds like solar wind hybrid assets power to X all kinds of other variable load assets also like last flea the EV charges so we have a lot of different things there and I think we are getting quite good at the asset management as well it actually turns out we started pretty early on the asset management so after I think about one year we started building that and that is taking actually pretty hard I would say to build that business and also keep it basically maintain good unit and I think it's because like the amount of work is a little bit harder than if you are like trading shop so that's also another great thing I think and then on the market access I'm like been extremely impressed by the number of markets we have been able to enter and that goes to our market access team in Japan was hard for many reasons like culture reasons, it applies to reason, Poland also extremely challenging right for other reasons so there's all these kind of things that I think are you just need to require great people to pull them off at a relatively fast pace that pretty much my milestones. Yeah actually one last thing actually very important things so one another thing we have a pipeline now about 500 million hours this project where we are a partner on the projects together with a with a German fund so created to them for doing a lot of the work but we are like a capital partner and also a trading partner on those projects and I think that's the comeback to the third point I mentioned right that it's actually very satisfying in a way to have a real world impact and actually see that you can go to the site and you can actually see that we have been a part of enabling assets that will eventually make our grid a little bit closer to carbon neutral. That's very nice and then the last thing I also we're starting to invest heavily in compute and I think that's also a big main milestone for us. Yeah a lot a lot to begin to that just to summarize like that is such an impressive list of accomplishments in five years you know even just on the trade we just think about the trade inside just to be in it as many markets as you are in in five years is rare to be covering all of Europe to be in the US and to be in Japan like there's that is like the road map for a lot of firms but not many are already trading all those countries inside of five years so I think that is super impressive and like you said you must have an amazing market access team because I know this it's not easy and it doesn't happen overnight to get into places like Japan. Was there any big takeaways for you there like with Japan I think you mentioned like it's it is very different when you enter the Japanese market compared to maybe the Europe or the US. I mean just from one thing is just that everything is in Japanese so it's meeting right where everything is in Japanese and that means there's a challenge right in it's yeah because I don't speak Japanese I understand a few words but I mean very very right so you were super reliant on having a very good partners and local people also. We have been very fortunate to hire a local Japanese guy, a ex-Mitsubishi guy, who has helped us with a lot of understanding the business culture and also low-level stuff like acting as a translator and meetings. But also, Japan is very relationship based so open a lot of doors to businesses that basically you can tap into relationship of these big trade houses that exist like Mitsubishi and Sumitomo and these ones. And there's this very different from Europe, strong kind of a relationship with people who have been in these companies. So if you can somehow tap into that, that's a great help for foreign companies. Yeah, so it's a show and you're going to be there for a lot of the long term and having that those local talent like on the ground in like Tokyo to make sure that you're able to actually build the personal relationships that maybe are more required than cancer in a year or a year. Also, just getting the meetings with the right people at the right level, we require relationships. That's just new in a way. You can actually link them, for example, it doesn't really exist in Japan. So that's the thing that Japanese, they're much more loyal to their employer. So having a linked info for them is a little bit like a sign of disrespect to your current company. So that, I mean, just for practical reason, that may extremely hard actually to find people that you want to engage. I have to do it in a different way. I think it's a different way to learn how a different way of working, right, that has all the strength, right, and all the challenges, but it's just very interesting to see. Yeah, for sure. So do you have a job at office in Tokyo? Yeah. Also then are you working with customers also in Tokyo? Yeah. Do you have to have assets to train? Yes, that's something that depends, I think, answer. The day I had market, is that the structure is very similar to the UK market. There's more zones, but otherwise it's actually the same as a half hour-based market and have their head and something akin to it today. What is attracting you guys to Japan? Why did you want to make on moves so early on? I mean, scaling as fast as we can Europe, and basically we can just see that we will be constrained on market size and the same goes for the US. And then we're looking for other big markets and Japan is a two times the size of Germany. So that's, I mean, and then it's liberalized relatively recently. Obviously China is also a very market, but that's not really accessible for foreign companies. So Japan is basically the second or third biggest market in the world. Yeah, that's a great progress there. So that's the trade inside, I can say, on it in itself is really impressive the progress you've made. But then when we talk about, you've also got a customer business as well to have that alongside the trading business. Like how different is it getting within a proprietary trading business to an asset business? It was a big, you've not done either before. Like there were both completely new to you. It would be interesting to know what were the things that were challenging and the key differences is one harder. Yeah, another big question. I mean, I think our physical asset team, that's what we call it here, have done a great job of handlaces. Very complex, I would say, to operate assets because, I mean, we install like a computer on its assets we control, at least most of them. And then we integrate that with our like server infrastructure level then commands. But for various reasons, actually the actual control happened locally. And that's down to like a hundred millisecond level pretty much that you need to control the battery. So just kind of getting that up and running reliably and making sure that you, and this is real hardware, right? If you do something really unfortunate, you can black out or like part of the grid. You need to be a little bit like certain about what you do and not make mistakes. That's of course, it's one thing. And I think the other thing is, like just engaging with customers, right, is a different type of skill than engineering. So we have to learn that. I had to learn that myself. I'm very much an engineer, but I have learned to basically how you communicate with people who are not engineers and what they would like to know and how you can kind of information in a way that's more suitable for people who are maybe not coding themselves. I think that's also very interesting to learn and we have been on the good journey to learn that. But I think operationally, it's just a different beast to sum it up compared to trading. It's much harder from an operational point of view. People and you have customers who really care, right, because they maybe put like tens of millions of euros into some methods. So they really care about that you don't let it not run for a few hours for whatever reason, right? And that's of course very different to trading. They can actually just decide to shut down if you want to fix something. That's not really an option in here. Yeah, yeah, absolutely. Yeah, I get so, did you treat trading as purely an engineering problem that you could solve with your engineering skills apps? Yeah, I would say like, I mean, we're an engineering company. So we try to build like very strong engineering culture. And I think what we have good is the problems where we can like turn them into engineering problems, where we can like basically get them framed in a way. So we can optimize them, we can automate them, and then we can build like great software that will solve the problem for us. And then I think what we have convinced ourselves now and have proven is that we can want to be able to do that, we can build like better machine learning models, collect more data and like build more compute, and then we will actually make like the models and a better system, both for the customers, obviously also for the trading business. Yeah, and did what when you started dealing with customers, and you were talking, you know, your background is different to many customers, to my users around there. So would there are any things that you needed to convince potential customers on it to begin in? Yeah, I mean, obviously the track record is a hard one for the first US, it's hard to get around and you basically need to build the relationship with someone to like convince them to take a chance with you. So that of course is hard. I think the other thing that we have learned is, so it's like some feedback we did get really on was that some customers, like we had a meeting with them rather than they basically our engineers would tell them ask them what they want, what do we want us to build, we can build whatever we want. And then they're like kind of in a way paralyzed a little bit by because maybe they don't actually know exactly what they want or they get a little bit a, it's actually a hard question to answer right? So we can learn that we should actually never frame anything like that, we should frame it as like we have two options maybe. This one and this one and it's up to the customers to choose right but we cannot send back them too much in information and that doesn't work, that's one thing I think is in engineering fallacy in a little bit. Yeah, I think Steve Jobs famous is that customers should tell you what they want, you should be able to provide, track them by the forum, word to word, in fact. So obviously then, why don't you have the track record with the early assets, I guess that's allowed you to grow the portfolio because then it's based on results, is that's how, is that how things are grown from there? Yeah, yeah, I think that's that's accurate. So obviously we have a large customer base and then some of them will do very well and then we grow with them but then we also of course as we grow we can kind of target more like enterprise customers. So we recently signed like several projects with a large IPP, like hybrid projects, wind projects and best projects and also some power to ex facilities. And of course that I mean they will not sign with someone who doesn't have any track record. So that of course, I think that's very nice for us. We have a scale I would say and we're better at handling larger assets and that's just basically like it's a cost for us is pretty comparable, it's not the same of course but there's some, I mean there's a certain amount of work that goes into an asset and that doesn't really scale so much actually with the not make a watch size of the PV park for example. Yeah, so you have a bit of broad spectrum of assets, you're not just focused on that across the renewable start. So maybe our business model is pretty much, I would say we follow I think the existing balancing responsible part is they actually invented a great business model I think. I mean you balance assets and you have a trading business on the side because they are complementary in many ways. But I think what we're doing differently is just we have a change the way that it operates internally. We offer many of the same products of course we believe that our products, we can do better products but the kind of fundamental business model is very similar to like actual or currencies or these types of companies but we have just don't have a like a trading company internally, we had to company internally. So the way they would have like teams of people physically trading around the assets to optimize on any given day or hour you just don't have that. It's fully automated. Yeah, I mean you can maybe look at their org chart and then you can stop with the IT department and then everything that's kind of upstream on that doesn't exist here. Quant like in the traditional sense traders all of those we don't have those we have. Is she not in people and I mean they do some of the same things as quants do but it's much more like systematic I would say it's very much focused on how can we get it into code and it has to run automatically. Yeah that's pretty revolutionary compared to what exists in the industry so that that means you must be so much leaner as well. Yeah I think you're probably I I mean we're about 35 people right now. Yeah. And those are that's that's your trading that's your personal business and asset business. Yeah that's that's all of it. I mean it's super heavy on tech. We're building a commercial organization of right but we're very very very on the tech side right now and I think that's a little bit like a you know a fallout of of my own and and a surrounds background like the engineers right so we like to build the technology build a great product and then you know think a little bit about how we commercial that. Well think maybe commercialization is following after we have built the product and and I can just see I think it's also very interesting to like look at other companies whether you do it in the all the way around it maybe build the brand first and actually a lot of customers and manage to do that and and are very good at the storytelling and and and then they need to build the product afterwards and it's just it's intriguing to me how you can do that because that would be that's completely like the all the way around the I'm thinking I don't know what the way I just think it's interesting to look at. Yeah no for sure and yeah I think people probably listening to this from the industry thinking how how is it possible to do it with with some few people. It's a question that I have you said like traditional quants you don't really have them. I'd love to dig into that like what's the difference in what a traditional quant may look like compared to the types of people you have in your business like the skill sets what are the roles and responsibilities that they have. Yeah that's a good question obviously I don't know what a traditional quant do because I don't know exactly what all the people do but my understanding is that a quant in many companies is that you have the traders and then you have the quants to do a lot of analysis and maybe contents like some of the like ideas from like coming from the trading floor and then they do some analysis and eventually have some ideas and and maybe automate some of the like signal generation and so it's very much like kind of a data exploratory and a lot of analysis of data. With us I think the kind of comparable role would be a machine learning engineer. They obviously also do a lot of the exploratory work but our force is never to support a trading floor in the same way like we need we need people who can put stuff in production themselves and put it at least close to production I would say. Like some of it requires especially like like specialised soft engineers who know something about like building fast system and writing in for various kinds of accelerators and whatnot. Right that's not something we are but we need people who can put it relatively close to production right and really understand how the actual system works because I'm kind of my personal opinion is that ideas are relatively cheap but actually executing them is much harder. We have just seen that we have seen a few times here where we have hired people who are not good enough at coding and then they end up I mean they actually generate a lot of ideas but they in a way they just generate work for other people but they actually don't really execute on it and that's the hard path of I think that's the execution path like actually really bang home completely and making sure that it runs every day doesn't break all that stuff right that actually comes after the analysis path. I think that that to me is maybe 80% of the work and then you can iterative improvement on your system right once it's actually running. I think that's maybe the probably the biggest difference obviously and then maybe all one is that like machine learning software engineering is first first class citizen here that's where we get all of you they in a way don't serve it. I want to say pretty much like the entire organization service the IT department or like the kind of instead of the all the way around right and I think that's probably different from how other people have set it up. It's not just that I think the other approach is also actually good approach just that we don't have the right people or the right skills to do that. Yeah no super interesting so does that mean that you guys don't necessarily need people with power trading backgrounds if they are hardcore engineers or yeah and is that what you value more like the hardcore engineer which the deep learning machine learning expert over the I've been in the power intraday power market for five years as a quam and I know the rpm mark inside now yeah and I've got a track got in that so you go with the hardcore engineer. Yeah any day I would do like we actually don't test that all for any any knowledge on power markets when we hire people for these like rules obviously if we also have some power market experts who are more like in the asset trade side that needs to know a lot about the salary service markets and how the regulation works and and really understand how actual asset operates but for the power trading and the suffers side we simply don't care right it's not as negative to know something about power trading but I don't think we put a lot of emphasis on that we believe that we can teach people what they need to know. Also I think actually people our people are more interested in kind of engineering in a way that in power markets actually I don't actually think they know what the power prices I mean they don't know what power prices but I don't actually think they know what the price is right now which I think is probably very different from if you are a power trader I mean you probably know have a got feeling for what the power prices tomorrow. Yeah for sure that's really good. Yeah so it's harder to find the engineering knowledge than I guess you can teach someone who is world class engineering what they need to know on the power market rather than the other way around. Yeah I mean I just think like we I don't know if it's harder or I just think we need to need to have skills more and then we value that and I think every artist probably like well maybe to put it in this way that the amount of power and the power that people need to know is relatively easy to know they're out of it not in the general sense right but the engineering skills they need is not something you can never learn in three months. Yeah I know for sure and I look I know this I know this as well you guys have a very high talent bar to even get into the engineering to use it to make I guess you have to be like the top the top 5% of wishing. I think you're taking higher but I know this is incredibly tough to find that back to a person and is that something that you've done just through you know your network looking at PhD recent students and recent master students and then developing nurture in your own people. Yeah I mean in a way we don't really care about people's background I would say but obviously on the other hand right if you have a PhD from a good school you are I mean you have caught like basically being able to pass a high pass at some point right so I would say I was successful with PhD students from who have done a good PhD thesis is higher than people who have another degree but we are also higher higher a lot of all of them. Yeah I don't care if people have a PhD or say but I think PhDs have been trained at least like good ones have been trained in solving problems right and are aware that you will fail 10 times before you succeed. Yeah that's a very valuable thing for us but how we hire I mean initially right we hired a lot from our own network I mean you risked the quickly to plead that out there and then we of course you need to think how you track great people and I mean for us it has been very much like outbound search for us so we kind of you know find the people we interest it in and we basically write them or call them or something or and a little bit pushy on getting in touch right it's still like relatively low success rate I would say like a lot of people then they're not interested they cannot move with family or something like that and so we are also in office I should say so we actually would like people to move here which of course was a little bit challenging sometimes in Copenhagen. Yeah coming I mean actually surprisingly for many people Texas is actually quite nice for our own skilled workers in Denmark but obviously like just the practicalities of moving is something that will prevent some people from actually pursuing a job with us. I don't think there's a lot of time on hiring I think it actually works to either founder and actually reach out to people and show that you actually care and actually want that particular person. Yeah so you're actually like to die some of your personally still involved in us to founder trying to find the best people you sell. I don't know like I think my greatest achievement has been you know a great hire. I think that's really a really interesting point because you know I don't know how many people who are as you know founders or directs are still actively involved they tend to you know outsource that or build a machine underneath that will help them but I guess what you guys are focused on is that special talent engineering talent which is so hard to find it could be so valuable when you when you hire great people. I think also right we of course I mean I can judge I was able to be reasonable accuracy if someone has a strong background in in machine learning right and that just comes from working in that fields myself for a long time I think but actually it's actually extremely hard to judge the skills of someone in a field where where you're not an expert. So, for example, I think some of the greatest hires I had. We have done, I mean, obviously all the tech people, there's really great hires we've done there. But I have been actually been very, very certain about most of them that they're really good because I mean, I feel like I have a good sense of them. But actually some of them all, like, you know, on the operation side, on the sales side, on the maybe on power market, expert side. Some of those hires, I in a way, I wanna say more, but I think they have been actually also very instrumental in a way, right? I've been a little bit high risk because I have less knowledge about me personally and also the entire team. What is a great operations person, or what is a great sales person? How do they look? That's actually not easy. I think if you have not, have not never really done it yourself. - Yeah, no, for sure. So how have you managed that then? - I mean, I think that's the same, you know, you talk to people, I, you talk to them right, and then I try to get a good sense of it. If I think people are kind of in general smart and know what they're talking about. So I try to, you know, ask, like, keep on asking questions until I order. And so they can explain something, is that I don't know anything about, so I actually understand it. And I think if they kind of fail to do that, I think that maybe show that, either I'm not kind of skilled enough to understand whatever they're talking about, right? I mean, one explanation, or the other one is that they actually don't have a kind of good enough understanding themselves, so they can actually tell the points in a way where a non-expert can understand why this is a good idea. - No, no, for sure. So something that I wanna talk about as well that you mentioned is that you're infrastructure and compute for, in the world of obviously AI that we're in today, and you guys obviously be in a, an AI first business. I would like to talk about that, 'cause I can see, you know, some GPUs in the background. Are you building in-house your own capability, compute capabilities? - Yeah, I mean, we actually pretty early on, I'm talking a decision that we wanted to post most of our development compute, I would say, that stuff that's not actually running production models. That we will host that locally or in local, like somewhere in data centers, but actually not in the cloud. And we did that for multiple reasons. I mean, one is that we, I think like we really try to maintain a strong engineering culture, and I think one part of having a strong engineering culture is that you actually know how you are kind of a primary machine work, and for us, our primary machine is computers, right? So we built them ourselves, we know how they work, we know what the limits are, how fast they are supposed to be, and that goes also to the network stack, right, of how they are connected, and how we can serve data to them, and all that stuff. And I think that's a little abstracted away if you go to the cloud, because that's just, you know, you log into S3 or whatever, right? And you, I mean, you don't actually know the hardware, how that's, like, what are the actual hardware limitations of what I'm working on? And we want to squeeze out every single percentage of performance of our hardware, right? And then we just need to know how it works. So it's kind of one thing. And the other one is just simply that we can actually not buy compute at, like, the right compute that we want. And then maybe that actually something that happened later, it's not an original argument, but then it had also been extremely turned out to be extremely expensive to be in the cloud compared to building on-premise. Interesting. I think right now, what last kind of analysis we did is, like, payback time of our hardware is less than a year. And I think that's a great trade-off for us. I don't see this trend stopping, honestly. And of course, it's also nice to have-- I mean, we know that our compute is available, right? Because we own it, so we don't need to. It's not like we cannot rent, because it's full capacity at the data center somewhere. Yeah. So I guess what are the other than the cost? What are the advantages of you guys having built in your own data center in house? I mean, I think that the main one is just like speed. It's much faster, honestly. You can do the right trade-offs for the right use case if you know what you're doing. So we have extremely GPU intensive right, and we can put a lot more GPUs in our servers. And we also know that, for example, one bottleneck we have is to serve data fast enough to the GPUs, then we can work on how can we serve data fast enough. And we can build the actual physical infrastructure to do that. It's a little hard to do right if you're in the cloud right, because then you get a shelf product right, that's just different. And anything actually just the physical kind of act of, actually putting stuff together, that actually makes you understand a lot of our house stuff works. You know, what type of cable is it like? What does a fiber look like? And how does it actually connect to the switch and all that stuff? I think that actually makes you understand and appreciate a little bit what you can do with certain things. Yeah. So you mentioned speed. So is that speed for trade in? No, that's like training. That's just like a brute force throughput on training models. Right. OK. But does it have advantages for your trade in activities as well? Having GPUs and-- Yeah, I mean, a lot of the stuff we do cannot run on CPU, also not in production environments. But I mean, obviously, you can see some of the GPUs here, and I think there's around 50 hosted out there. I mean, this is obviously not something we will use in production, because we need more like high reliability. This is still fairly reliable, but we don't have like a 99.9999 uptime. That's not a priority. So the actual kind of operation is hosted in high quality data centers, where we know that you know, and it works, power, all that stuff doesn't return. Right. Yeah. But I guess what has been the learnings and what have been like the challenges for building your own data centers? I mean, I think we actually started like a few years ago on this, maybe two or three years ago, I'm unsure. But the biggest problem is it's actually very interesting. I think because it turns out that, I mean, obviously, you need to buy the hardware and know what you need to buy and all that stuff. That doesn't take so long. I mean, that's the kind of information gathering and you need some people to buy. But actually, then it turns into more like low-take problems. How do you cool your hardware? How do you keep it cool? Then you need to build a cooling system, right? So we have built a cooling system in there, and you know, have a bunch of radiators on the roof now, and that extracts the heat. But then also, like some of the things I was getting enough power, like, obviously, I mean, pretty quickly, right? We ran into that, you know, there's only that many blocks and the circuit boards in a normal kind of office building, and you very quickly, depending that actually. So at some intermediate phase, we have these kind of cables running around in the entire office right, because we have to kind of connect to all those different circuit boards in a pretty big room to kind of power our servers right? So we didn't trip the circuit breakers. But then now we have a, given this kind of a cable pulled in from outside, that didn't do much more power. I think it's just interesting, this kind of, almost like a low-take, right? It's the same technology as you used like a hundred years ago. Just like, yeah. Think a big fat cable, right? It can't deliver a lot of power. No, it's funny because, like, obviously, if you listen to any big tech founder now, and they talk about, whatever they talk about, AI, they don't talk about resources, they don't talk about people, they talk about capital to much being the constraint, the constraint is actually physically getting enough power to run the data centers to do what they want to do. And I totally agree. We kind of, obviously, we cannot build like 10 megawatt data in our office building. We could build something relatively last-right, but we're basically looking at a version 2 now, right, where we need to host it somewhere else, where there's a bigger recognition, essentially. That's the kind of limiting factor. Yeah, I was going to ask that, like, what's the solution for a business like Twig? There's a start up. Then there aren't a mega capital, and that's a business like a big tech company that are spending billions and billions and billions, can't be physically spend enough on this. So what's the solution for a business like Twig? I think we are in a phase right now where we can also spend, like, a significant amount of money ourselves. So, I mean, that's one thing. I mean, we're fortunate enough to be in that situation. But, I mean, obviously, we also don't need the same kind of a compute as Google does, and we need a very particular type of compute. So we don't need to spend a billion dollars, but we still probably need to spend in the millions of euros right. But I don't think there's a good way around that. I think, before us, that's the kind of. I think it's actually quite interesting. We started out being capital-wide right, and very much like kind of a software business, and you know, had a few GPUs. But then, actually, we actually start to buy a lot of hardware. And one problem we have is getting rid of all the cardboard that we can get into the office right. But. And it's like, interesting. They're almost the hardest thing, right? But, yeah. But then, I mean, but actually, it costs money right. I can just see that we will evolve into more like a kind of a capital-intensive business, both on the compute. But also, on hardware, I think. I think there's some interesting things around, like, taking part in helping projects actually being. Like, then I'm talking about grid batteries, or solar, whatever, actually being pulled across the financing line. And sometimes you just need a little bit of extra capital, and I think that's where we could have a role in that. But I think it's just a completely new thing for us, right? That we, like, being kind of ambitious, but also, like, discipline right, with how you spend your money. Yeah, for sure. And I think that's what we're seeing now with, like, cutting edge, quant trading firms. They've been coming, like, hardware businesses as well. Like, like, I saw. So, Jane Street, they've got a huge data center. There was an interview them showing someone around recently, and. Have you not fight there yet, right? But, I mean. I've read it. I mean, obviously, they are an inspiration, right? Those type of people. I also think, actually, companies like to go. And other. There's a compute company, right? Like, Jane Street and HTX is also, I think, actually. They actually invested in Anthropics. So, I mean, that was probably a home run from an investment fund of you. But also, like, companies like Glyncore, Trafigura, or these ones that come out of the. I mean, they come out of an oil trade trading business. I mean, it's not just that we don't want to trade oil. But I think they have. All of them have kind of moved upstream of the actual trading business to be. easily secure, like supply or like should be a little bit more entrenched in their position right. So they own right factories or like mines like Glen Cornhouse, one of the biggest mining companies right in the world. But they actually started out as an oil trade. I think this is a very interesting route that you kind of move more and more towards actually having a real kind of physical presence. Yeah, no, it's a great point. And I guess is that what you're trying to do on the renewables side by making investments yourself? So like you actually have skin in the game with the assets, some of the assets, some of the projects that you've put your own capital into. Yeah, I think that's one way to look at it right. I mean, that we have another one is just that it's a kind of partnership right where we bring some money right. Obviously, we're not an investment fund in a traditional sense. But I mean, yeah, we have skin in the game right and then we bring a lot of knowledge on how to like sizing and all that stuff. But we also do that to all our other customers, I should say. But I think actually it's more the kind of a sometimes you just need a little bit of additional like capital that has a high risk tolerance to basically enable all our types of finance. And we can sometimes I think have an outsized impact by providing that. We're seeing the same right. There's a bunch of the all our trading companies that are thinking kind of going down the same route. I think like second foundation, for example, have done it to a pretty last degree. Yeah, very impressive. What they have done, I think, but there's also all the companies right that that somehow trying to do the same thing. And I think I mean, another thing is just like it's actually harder to do right than being a trading company. Just being a trading company is also I think that the hard part is to make sure that you stay ahead of the game and be, you know, make sure that you don't get complacent. But you can also build like stuff that takes years to build right under like as a management management side and on the actual project like enablement or whatever you want to call that. But that takes a lot longer time to scale that business. And I think for the same reasons, there's more in a way, defensible now because it's fewer people want to do something where you need to spend five years on actually like getting to a place where you can even have a seat at the table. Yeah, no, for sure. And I like you say, I think that is there's a couple of companies that are doing that extremely well. And yes, it is extremely impressive to see what second foundation have done in a relatively short space of time with their asset business. I regard them as a similar type of business to yours actually in same engineer in my set for sure. But given your AI background and a lot has happened in the last two years in the world of AI since we spoke. So I'd like to just go into your thoughts on how AI is being used from a powered trading perspective. I just didn't just get your thoughts on where we are right now. I mean, for us right, we are, I would say, I mean, I call us an machine learning company, right? I think that's the same as what most people call like AI right. So I mean, we build our own machine learning model right. And in there, we got, I was a real super heavy on our like high conviction in that is the future. I was a little like this interesting right development on language models, which is a little bit like a separate kind of topic, right? That you have all these chatbots, accounting agents and all that, which in a way, I don't think they're very useful for actual power trading because they, but I think they are useful tools for like actually producing certain types of systems that you can use to support your operations. I don't think you're anywhere near where you can kind of tell a claw to build your system. That's a complete fantasy, I think. But I think you can actually make your software engineers more productive. And it can help on certain systems like we have had, I think, good success with front-end engineering on using agents. And I mean, one reason is that it's a little bit lower risk in a way because if there's a glitch right, it's a visual glitch. That's a core. Not great, but it's not like, you know, trading system and blowing up. So in those kind of areas where I think like kind of stuff that doesn't, well, there's not a lot of stuff depending on it. We have moved a little bit faster on AI adoption. And we use it for like a tool of to make our engineers more productive. But it's on a voluntary basis, I would say. So if people think they can do more, then that's great. I think that's good. But it's not like we have a kind of a mandatory unit to spend seven million tokens per month or whatever, because that's the KPI that we, I don't really believe in that. Yeah. So I think it's in a way like an enabler right. I think it makes software engineers and many other people more productive. But I think actually we'll solve it interestingly that we need more software engineers because we can produce more software, like more efficiently. And that would actually be more attractive to produce even more right. So that's the, I think we need. Yeah. Well, that's actually a nice thing to hear, because obviously your ear is like AI Doomsday things that are like, oh, it's going to take all the jobs, but you think it'll actually for engineers will have the opposite effect. Yeah, I think I mean, there's this guy called the Javons paradox is actually British guy, but he basically observed that if you make a steam engine's more efficient, so the use less coal, then you see a net increase in coal, coal usage because people will run their steam engine more. And I think it's a little bit similar here, right? We know we know we see an increase in efficiency, right? And they actually want to do it even more because it's a way like marginal cost a low. I think the other thing that you don't need software engineers or anything like that. I think as a complete nonsense, I think people who do that they will just missively fail, I think, but I mean, it's a free world, of course, so you should try right, but I don't believe that. What is the reason that you feel so strongly about that? Is it? I mean, one like for the model capability or? No, no, I think the model likes in general good, but I think one thing is like, I mean, they will build whatever you tell them to build, right? So I mean, and you need to be very specific what you want. And if you don't know what you're talking about, then you cannot be specific. Then you can say, build me a trading system. Yeah. What is going to do? It's going to build you probably going to build your trading system, but whatever, I mean, probably not going to work right. So one thing, the other thing is, you know, when things doesn't work at 11 p.m. on Saturday night, right? You just want people who know how everything is wired up. What can kind of cost this type of error? How do I fix it? Is it serious? Like, what's the kind of impact? All that stuff right? Is anywhere near trusting AI to kind of give me an evidence on that? It can maybe help sign all things, but like the final call I want that to stay with people. And I want, yeah, I mean, I see that a little bit like it's very calculated made, you know, our accountants more efficient, right? Then spreadsheets made them even more efficient, right? But it's a little bit the same, right? It's not like I don't have any accountants anymore. Yeah, no, for sure. It's a tool. It's a very, very useful tool for an expert that knows what how to use the tool. Yeah, I think of course, then there's certain things that will be more like exposed. But I actually don't for me, it's not really programming, I think I think actually more like white color work, traditional kind of like, white colors a bit worth, but more like it's extremely good at like generating text that has the kind of right form, like reports, stuff like that. I think that can be completely automated a lot of that. You know, where the maybe the kind of way to do this kind of, you know, you get all these kind of the kinsies or whatever, they're probably already generated by AI. And I think we will just see more of that. And I think that's actually very interesting. I think because people actually thought that a lot of these things, I think actually AI in a way reveals what hard and was not hard, or like hard for my kind of my sound a little bit offensive, sorry, but you know, it's very, very strong on like the first thing a language model could do with between rhyme and write a rap song, all that stuff. And that's just because it's a very kind of particular format, right? Actually, but actually the content doesn't really matter so much, but actually if you get the style right, and there's a little bit the same right with some of these, you know, reports, right? It's a little bit more like a format or like a particular style than actually kind of novel content. That's more like a gathering of existing stuff. Yeah, so when there's like a formula? Yeah, like a kind of, you know, you just like collect all the facts and present them in a particular way. That one is unbeatable at that, I think. Yeah, but for you, we're in the context of power trading. What do you think is extremely good at? And where is it perhaps not still at the level that would be needed? I mean, like if you're talking about like using language models for power trading and then I mean, we don't do that at all. We build our own machine learning models. I mean, one thing is they're way too slow. I mean, another problem is I think it's just duchastic. So I mean, you can ask it seven times about the same thing right? And we'll give you maybe four times the same answer, then three times something else. That's just not great for reliability. That's another problem, I think. And then I mean, that's the general, I don't even know how you do this from a compliance part of you. That's another problem. I think one one issue for us actually with language models is that of course, we a little bit kind of vary about a lot of this stuff is hosted like by big companies or like what do we actually, and essentially like you send everything to them, right? The leaders in your bank and answer. Of course, you should be kind of cautious about what you're doing. And also a lot of the very that is implemented is that actually the kind of a coding agent, for example, would have the same rights as the person that are coding. And that is just also not something I see as the way forward in general, because I mean, that means that in principle, it could wipe my co-group position. If because it has the same rights as I would have. And I think that's not the way forward. So we are trying to do like stuff around like building systems right where we really strongly can control the rights of the agents and we host them then internally. I don't think we just maybe next generation the open source models are I mean the new GLM 5.2 I think was called is pretty good. But I think then we can control right we know that like this, we give it a task like fix the spark ride, you know, it has right to the only the stuff it needs to know. We know it runs on our GPUs, all of that. I think that that is really a very important thing. And I also see kind of security breaches in going like somehow coming through these agents and just very let's something we have cautious about. That's really interesting. So the proprietary models. So when it comes to trade in coming up with actual trading strategies. to answer the market you would not use an AI model for them. I mean, I'm like, you know, the usual answer, like a guiding principle, I'm well, as your answer is no. But I mean, maybe my reasoning is more that, like, if you can just write like it's incredibly easy to use these tools, right? And there's almost like addictive right in a way. But you can just, if you can just write, like, write me a profitable trading strategy. It's just not going to work, I believe. Like, that's super easy. It's low effort, like, everything, right? Like, the chance that will work is extremely low. And even if it would work the first time, then it would be arbitration the way immediately. Right? So, I mean, my, like, our kind of, we have this kind of a little guiding principle that if something is easy, then we should be a little bit cautious of why it's so easy. Then probably because we are not thinking about it the right way. So we are much more confident when we solve a problem that we think is hard and we get good results from, from some, because like, then there will be fewer people who are, like, putting in the effort or have the capabilities to actually do it. I think you can ask, you can use, like, these coding tools to actually maybe move faster on some of the implementation side. And not so much the kind of excellent production system, but on some of your experimentation, you can, I think they allow you to move faster. But the actual idea generation, and actually kind of, I don't think we have seen anything that are producing stuff that we could use for. I will also be like extremely surprised if people believe that. I mean, in a way, I think, why, I mean, I would ask the opposite question of why should I work? It would be very surprising. But I guess I get what you're coming from that if someone can create a really profitable trading strategy on Claude with heavily using Claude's capability, then I guess anyone else can do the same. That would be my reasoning. It's not to say that it will not be possible. I just have a, I would be, you're extremely kind of also valuable to other people doing the same thing. So you use it for productivity, but not for IP creation? Yeah, that's probably a good way to put it. Yeah, cool. And I guess that's where we're seeing a lot of these open source versus closed source models. Yeah, basically the Chinese ones are more open source, but like you have the, the Quinn, more than things called, right? And I think it's called TLM. They are actually open source and open weight. So we can host them in our, on our GPUs if we want. We already do for some of them. And that I think is something we will kind of push on to use those. Like Claude and Codix are still superior. I would say especially from a kind of, they can do more, but then they have this issue around like, you know, rights and you don't really control right. In a way, given away the keys, right? So yeah, I would at least like to have an alternative, especially for like stuff where we actually build a system around it. I also feel like there's something around pricing, right? That I mean, you are very, very exposed to, like, I mean, in a way, it's kind of the opposite of the software based model, right? Because the software is like, you know, actually put in a huge amount of cost upfront. But then actually running the software is cheap or like free. But it would be like if you actually rely heavily on these services for actually operating your system, you have flipped it like actually turns into very expensive stuff to run because these tokens are, I mean, they actually subsidize right now, right? So they're actually too cheap. But that might actually mean that software is actually very expensive to say, high-orbit right, but actually low cost of entering the market. I think that's a terrible position to be in and do as a business. Yeah, no, that is really interesting. There's something I'm seeing across like the tech podcast that I listened to about like the Chinese models versus the US models. Something that obviously I want to talk to you about, given that you came from a weather forecasting background before entering a power trade in. You know, I see so much in marketing now about like the advancements in weather forecasting models because of AI. Be interested to just get your views on just weather forecasting now, the capabilities it has and what you use at Twig and how to use it. No, that's a good question. I mean, these weather forecasting models are pretty much all the same in a way. Are they at least take the same approach, right? I mean, they essentially try to predict like replace the like a physics-based model with a learning and model. Learning that dynamics are the atmosphere, but don't do it from a kind of first principle point of view with physics, but do it from a kind of just a new network and a bunch of matrix multiplies. So that's the kind of a backdrop, but they're all like in a way similar. They do this. It's the same training data. They are all trained on the same data. Then some add like a little bit of proprietary data or not, right? But it's essentially the same. It's trained on this ECMWF data, this era of time. That's it. But I mean, we benchmark a lot of these external providers and try to like test them in our system. In general, we have not seen so much gain from it. I would say maybe it's our style. I don't know, but we have not really been able to see great improvements. So we run a lot of it in-house. It's not just that our models are better, but I think one reason for running in-houses is that you there's this kind of problem, right? If you are a weather provider, like a forecasting provider, in general, actually, but also in particular, also for weather. If you provide like forecast, then you provide them to financial markets. Then actually the more customers you have the less valuable your product is because like more people will act on your forecast so they will kind of price out the advantage or like actually reduce the value of your product. That's a kind of negative network if they don't know. Which is of I think one reason why it's hard to have a really great weather forecasting because I mean, what is something successful people will pick it up and they will be less useful. So we run our own and I don't think our models are actually better from a kind of a root-meant square era of the spectrum, but we hope that there's officially, they make different errors than what most other people do. And that from a monetary point of view that makes our errors cheaper because we are, well, we are hopefully a little bit less wrong at the same time as everyone else. And that just makes it cheaper. That's just the way the power market works. So that's the kind of fundamental one reason for us. The other one is we can integrate the model a lot more tightly with our other setup right. So we don't have to query like a lot of these providers. They have two types of interfaces. The one is like you give it a longitude latitude, you give it a forecasting time, then they give you back like solar radiation, wind, temperature, whatever. And that's basically API-based. We can tap into our own models, most more deeply, like integrate them more tightly with our other machine learning systems. But think that's one one advantage. The other one is speed. It's faster to run on premise. Yeah, I think that those ones are probably the main source. Maybe I can add one one thing. I can see the other interface that I'm a little bit skeptical about is that people like, you know, there's this language model in the way right now that people put up like then essentially they have a weather forecasting model right. Then they want an interface that a human can use and humans are not really good at, you know, looking at the atmospheric maps of something. So then you put a language model on top so you can kind of query it like, oh, will the forecast be too high as it loads tomorrow or in London or something. And that is just not the way we operate. That's kind of a super opposed to the way we have set up. And we would really like to cut out that step of needing to support a essentially a human decision maker in the loop. So your model, you've got your own proprietary model you built in house for weather forecasting. You're able to do that at real speed. And then there's no evaluation by any human. The weather signal is then interpreted and used by your machine learning. Obviously it's not just the right and backtested and we have a bunch of car wheel right on everything but I mean in live setting we just did go straight through right. There's no human intervention. Yeah. And how big an edge do you think that is for a trade-in firm? I mean there's what work for us. I mean I have not, basically, I mean we have failed to deliver a lot of value from like a provider forecast. So not a provider but like third party forecast. So that's the, I don't know about exactly how valuable it is. But do you regard that as a strength of tweaks that you are? Yeah I think that's one thing we're good at right. I think we in general good at machine learning I think. So we try to build machine learning systems and then once we have a machine learning system how can we scale it? How can we throw up computer learning? And I think once we are in that kind of operational mode then I think we have a large advantage compared to many other companies. Not of course everyone what I mean 95% of other companies are not set up the same way to do that. Yeah. Something I'd like to close with is just looking into the future. So regardless you guys is like probably the most forward looking business in power trading. That's my view. So what do you think power trading looks like in the next five years? Yeah, how do I mean one thing we see right now is there's a large increase in the number of like orders in the market or we literally we see pretty last increase in the in the kind of order flow. But I think maybe from what I highly respected I think it will move more and more towards like something more akin to a creature trading. Not that I mean power has its own kind of a characteristics right but more from a kind of like the depth and liquidity of the market. The transaction costs I think we'll see that see that that's my clear belief. And I think we already seeing that right the orders and the amount of trades just keeps going up every year. The volume is like not increasing as far as the number of trades or number of orders but it's still increasing and I think we'll just see that continue. I can obviously why that should stop. And I think it's actually great because it actually allows like if we can get the transaction costs down right we allow more people to participate in probably in adjusting their load or the production to whatever the current prices are. And I think just countries that have a liquid market also I think it's just fundamentally better off and you can see that in every thing of the metric under power might right they have lower cost on imbalance they have lower cost on anteliric services and just because they can actually handle a lot of it on on a liquid entry market. So that's my belief like just increasingly like liquidity will keep on going up. Yeah awesome and so that means I guess the firms that are more focused on automation because if you look at equity trading now. It's pretty much fully automated. Yeah, I think that's true. Yeah. And also, I mean, there's still like people who do like mid-range stuff, but they're actually also doing it by a computer. So I think like if you look at the kind of traditional power trading companies, they have been a little bit had a rough ride the last few years. I mean, obviously coming from a very high peak in 22, right? But they have been going down. It's like 90% per year, something really fast. And I think we are just seeing other companies that are more tech-driven, have an advantage and a bit of cost-based from an operational point of view. That's one thing. Yeah, that's probably the most important thing. I think that's a big takeaway for me. It's like you guys have been producing 90% year-on-year increases in a market where you look at a lot of the competition. They've seen numbers decrease by 50% or more. So I think that kind of tells the story as well. And I also think like when we started, there was a little bit like people saying that you cannot automate power trading and all the usual stuff. People say, right? And we heard that, oh, I know, again, also in my previous machine running, you could not automate translation. You cannot automate what they always say. I have this thing right. If you can make a decision in less than a minute, right? I don't see very little reason for not believing that you can somehow automate that eventually. And I think we just see that now. Like now, power trading execution happens to be a very fast-paced thing where you make tons of decision very fast. And I mean, humans have a limited or compute capacity. There's no good reason for not believing that why you should not be able to do that with a computer. And I think that we'll just then, I mean, then we were doing like short-term stuff in the beginning right. Then we're just, of course, making our way out the curve. And I think we will see the same right that now we are looking at the intro. Like initially was short-term holding, holding for a few minutes. And now we are more also in the head markets and stuff like that right. But I think then we are moving into more like in future markets and all of these longer-term markets. And you see exactly the same right in equities that you have these high-frequency shops that are extremely short. But they are actually not necessarily the ones that are doing the best that might be the mid-range, what do you call them like mid-range companies, right? That's not a very good execution right, but actually they hold a little bit longer. And some of them even go now holding like, have targeting more like month or four-need-fans, or even season. And that's, I mean, very comparable to what you see in future trading, right? In power. Yeah, no, that's the really interesting trend. I think, you know, obviously the case for fully automated trading in the short-term is, you know, no one has to argue that anymore. But I think the next is future power trading in futures. Like it's still predominantly done fundamentally driven discretionary. Whereas I think the dinner that I hosted, which you attended, I think it was a really fascinating conversation going on with someone from one of the biggest commodity trading houses. And then someone from my second foundation saying, why can't we come into your market and do what we're doing right now? And that person was saying, well, you can't do it because it's too difficult. Whereas I think there are now for sure businesses like yours that are focused thinking about, well, it works in short-term. So why can't we just use our approaches to futures? Do you think it can be done? Yeah, I'm sure it came done. I'm a high conviction on that. I mean, it might take a more effort right. And also the stuff around latency that are a little bit more competitive on futures. But I don't see any reason for why that cannot be done in a similar way. I mean, also actually on the future side, it's interesting that people who were on just maybe as a last mark on that. But as consumers actually make a ton of mistakes right on futures as well, right? It's not that like the homostrate crisis, right? Like I think that probably blew out a lot of fundamental straighters because they were just randomly caught on the wrong side, right? They might have the wrong, you know, maybe it's true that the market should be structurally short, but no, so long in the sense that the prices will go down, but like I also applied, I guess, is the right words, sorry, but then you know, randomly someone through the dice somewhere and then close the straight. And then there's lots of ton of money, right? So it's just more than that. I mean, humans make those mistakes as well. And it's like it's not, I think actually computers in a way would actually in a way have a if you do it right, you could actually account for some of the tail risk. And then you would know that this is a risky position to be in. It happens every five years that someone macro politically do something and you're if you find trouble right. Yeah, no, no, for sure. I know, and it totally, the systematic desks that I'm aware of in the energy futures market have done really well in the crisis because, yeah, they're not looking, they're not tracking fundamentals and they actually do well in these these circumstances. So I think that is a sign for me that it obviously can be done very successfully. And I think it's interesting that the companies like yours who have done well in short term, who've now got the capital to go into the more capital intensive markets. I think it's going to be a really interesting couple of years seeing who enters that space and the success that can be generated there. Yeah, Lizzy. Yeah, but Casper Locke fascinating talking to you. I could talk to you for hours, I really could. But I think, yeah, look, it's hugely impressive to see the last two years, the grow from all different angles. Like I said, on the trade inside, on the customer side, exciting seeing you invest in your own projects now. So I'm sure if anyone's listening to this who has assets, they should for sure reach out to you guys. Thanks, yeah. And a pleasure to be on it. I think it's really, I think building a podcast community is kind of a, you know, a non obvious way of actually, you know, differentiating yourself when you are in the hiring business. I think it's very very nicely executed. No, it certainly is become a big passion of mine. And there's real appetite in the power trading community to listen to different perspectives. Because there are so many different ways of being in this industry. It's not, it's not just one way of doing things. And it's such an important industry. And I think, you know, that's why alongside the podcast, the events that we're now hosting, the dinner that you came to, we're going to be doing loads more of that because, because there is an appetite for it in the industry. So, yeah, it's not good. I think it's very, I mean, like if you deliver something valuable, right, then people like it, right, I think that I meant just great and great service to everyone. Yeah, no, thanks. I've no doubt that people will find this one very interesting, because it's still the first episode with one of our most I talked about. So, yeah, thanks again, Kasper. I really appreciate your time. Thanks.