Danyal Akarca, Co-founder at Callosum: BREAKING: Callosum raised $100m Seed

0:00Danyal Akarca:Who do you think are going to be the biggest winners in the AI stack over the next three years? Easy question. At Kilosum, we're building the software to enable AI to work faster, better, stronger, and cheaper, utilizing combinations of different chips. We've just raised$100 million. Being the first company backed by some AI, you kind of have, I call them the Neo chips. So we had the old school chips that were pre-AI, and now you've kind of got the peri-AI ones that came around the renaissance of AI, and then you've got ones that are being built. since. Yasha and I actually dreamed about a future in which the future of how you build really truly autonomously intelligent, even perhaps conscious intelligent systems, just like we are as humans, heterogeneity is one of the core primitives of how you do that.

0:49Hello, welcome back to Skating Europe show. I'm joined by Dan. Welcome back. Second appearance in five months. Thanks, Seb. I don't think I was expecting it to be so soon. But it's It's amazing. But it's amazing. Before we get into the news and what you've announced, can you give a brief introduction to what you've built and what that says about the world that you think is coming?

1:08Danyal Akarca:So at Kelosum, we're building the software to enable AI to work faster, better, stronger and cheaper, utilizing combinations of different chips. So what that means is basically that I think most people have this idea that AI is generally driven by kind of single models, which is true, and that those models are very large. And that's also true. Models have got larger over time. And that's been amazing. You know, so many feats that have come from this. And largely, they've been running, if you think about how they actually work, on kind of large data centers of chips. And those chips have largely been what I call homogenous or the same.

1:54Danyal Akarca:And that's been great to train these models and has obviously been in the news because there's been huge amounts of data center build outs for these chips. What we've kind of discovered over the last kind of couple of years is that when you actually want to run these models, you know, you want them to be economically useful to build new products or you want them to help you with scientific discovery and so on and so forth. Actually, there's many, many models that exist now. So actually people are discovering loads of different models of different types, of different sizes, of different specializations.

2:28Danyal Akarca:And people have realized, given that this case of multi-model future that we're kind of moving towards, that you can also run those on many different types of chips as well. So you don't necessarily need to rely on one single type. And me and my co-founder Yasha kind of clued onto this and made a bet with building Colossum that actually this is an inevitable trend of which we're going to move towards, where if you want to actually kind of run AI in an economically sustainable way, you're going to need to move to a direction of heterogeneous computing, which basically just means a world in which there'll be many, many different types of chips.

3:04Danyal Akarca:However, the problem that you get is the software to be able to do that is pretty basic. It actually doesn't even exist or it didn't exist before we started Colosum. And so Colosum is on that mission to enable that system software to enable them to work. So I can tell you a little bit more about that. but at a high level it's making AI better using combinations of different chips. Amazing. Can we talk about the current generation is like predominantly you use a large language model Frontier Lab and that's leveraging a homogeneous type of chip, predominantly Nvidia. The better you're making we're starting to see in the market already is that we're going to have a new generation of chip companies that are specialized in different things.

3:40Yes. Can you talk about what maybe what that looks like today and maybe some example chip companies and how they're doing things differently to the sort of the homogenous world that we've been living in so far.

3:48Danyal Akarca:Totally. So you have GPUs and that's the NVIDIA case and there's actually many different types of GPUs. The other famous company is AMD, which also does GPUs amongst other types of chips. And what we're finding now is that because the... Just to say those chips can do a lot of things. They don't just necessarily do AI. They weren't even designed originally for AI famously. but what has now become clear is that because the market has now grown so significantly for inference which means just running models right the market is so large it justifies a huge amount of investment for new types of chips as you just alluded to and so these chips can take what we call kind of new architectures they have different kind of profiles and different ways of designing those chips in new funky ways they have things like different what's called memory hierarchies for example, or how they store information, how they compute information is just fundamentally different.

4:45Danyal Akarca:A little bit like when Ford came up with a car, right? There was different types of vehicles that were being made. And so different flavors that people were trying out, you know, cars with three wheels rather than four, you know, all these kinds of cool things. And so there's effectively been this explosion of new chip architectures, some of which are closer to what we're used to. So, you know, they're kind of more canonically like typical and some of them are dramatically different including from photonic chips that we're all the way to kind of thermodynamic chips and so on so the companies that you can include here are ones such as grok which got acquired by nvidia and they're particularly good at running chips faster running ai faster and cerebrus which just went public which is a wafer a wafer a wafer chip and they can run AI incredibly fast.

5:35Danyal Akarca:And then there's many other companies that we're aware of and that we're partnering, including ones like Accelera in the Netherlands and Rebellions in South Korea, amongst several other new companies such as Tendrils in the UK, who all kind of have different conviction bets on what is the right chip architecture. Yeah. And so could you give us an example, because you mentioned it's all about the economics of AI. So for a real use case, how would a company use Colosum in combination with different chips to achieve an outcome either better in some way than they would today? Great. So what Colosum does for end users is something that we like to call tailored inference.

6:17Danyal Akarca:So tailored inference is more to say that inference is not a kind of a blanket, well, Well, inference is a blanket term, but actually it's relatively uncertain to what extent this is economically viable at scale. It's still being worked out exactly how to run AI, kind of, such that it can be economically sustainable. And this is for the basic reason in my mind that we have a pretty much like a single hammer approach to everything. We generally speaking just kind of try out the new model on like a generic group of tasks and they're running generically on like quite homogenous computing infrastructure, as we were kind of describing before.

6:53Danyal Akarca:what our approach and our view is is that to actually make ai economically sustainable so that we can have g really positive gdp impact of ai technologies not only in this era of llms which i'll talk about in a moment but all of the new types of models and new types of intelligent systems we can build you're going to need to do the following which is what we do which is to understand a little bit more about what tasks people are doing so that's the level one like What are the tasks that people are actually using AI for, like concretely? Is it for document search? Is it for coding agents, which is obviously one large one?

7:26Danyal Akarca:Is it for being able to process cybersecurity risks and these things, such as some of the companies that we work with? Then you have to then match the correct combinations of different models to those tasks. As I said, not just a one size fits all, but actually many different types. And then what is the chip and computing technology that is optimally placed for this? So it's a three level thing in our mind. And that is in that process of correctly mapping and orchestrating that process, you can make AI far cheaper and far better. How do people access this? Well, there's two ways. For people that want AI, we just provide APIs.

8:01Danyal Akarca:And so we can tailor our APIs directly to users to provide them with optimized models and combinations of models to solve tasks that are optimized on the chips that are ideal for them. So that's what I call tailored inference. There's lots we can talk about in that. And the other is that we work specifically with people who want to actually build data centers and who want to be able to do that in a new way where they aren't just relying on one chip to build their data centers. And that we can actually enable people to have their chips and ultimately to generate out of that data center those APIs.

8:35Danyal Akarca:Those APIs that actually can deliver the intelligence in a much more cost effective or optimized way, depending on what they believe their customers need. And you came out of stuff five months ago. Yeah. Right. And at the time you announced coming out of the pre-seed round today, what are you announcing? So we've just raised$100 million and we're incredibly excited, led by Atomico and with Plural, who are backing us again after our previous round, including DCVC in the US. And in addition to this is the UK Sovereign AI, who we were the first investment of. We're so excited about that. In addition to a host of world class investors all over the world.

9:19Danyal Akarca:And so it's very and so we're super, super excited about that. Our team has scaled significantly since we met five months ago, Seb. And we think basically that this is one of the most exciting times to be innovating on the intersection, not only of the computing landscape, which is it's driving the world economy right now, but also in how we actually form a new era of how we make AI economic at scale. And so we're scaling up to meet the challenge. So super excited. And this is one of the largest seed rounds ever in UK history. Only five months after you came out of stealth. What has changed in the market that has given investors more conviction?

9:59And I don't want to ask what's changed internally that has, again, given investors conviction.

10:05Danyal Akarca:Let's say the market, what's happened in the market. Yeah. So maybe just before we started Colossum, people thought that me and my co-founder Yasha were crazy. So genuinely, and this is, it's funny to think about it even now, only even maybe a year and a half to two years ago when we were socializing these ideas to people. It wasn't the norm to think that even there would be new chips that would be able to supersede what is the dominant force. and what has happened even since our raise our first sorry our first announcement was that we came out with some really stunning results so we showed actually that for the problems that people are actually solving in the real world not um not kind of just training uh models generically to do kind of quite cool things but actually to solve real world productive tasks that people actually want to use AI for, we showed that it's becoming increasingly heterogeneous.

11:02Danyal Akarca:And what I mean by that is that if you think about the tasks that you try and solve on a day-to-day basis in your job, they're highly varied. Some things you want to happen kind of quite quickly. Actually, there's many tasks that you're doing simultaneously often or interleaved with each other. And what this means is most likely the intelligence that we, those are the tasks. And so as a result, it's become clearer, I think, to the market that those tasks are decomposable and that they can therefore to be done efficiently likely need what i'm talking about which is more specialized infrastructure to be able to to do that the problem with specialized infrastructure um is is that it's specialized and so as a result you kind of have to have lots of them and uh and then you have a new problem which is this kind of how do you wed that together yeah and so that is the new bottleneck and that's what obviously colosum is built for so that is the market uh problem that has really accelerated um that colosum is addressing in addition to the fact that many people are building chips and so in the interim period we've had amazing companies in the UK, in the US, in Europe all showing really amazing promising results on their new chips and so this problem compounds.

12:09Danyal Akarca:In terms of the company and what we've been doing well our team has been maturing amazingly we've hired some unbelievable people across the whole tech stack. And we've also really matured in our relationships with all of the new chip providers, which has been super exciting. As with our launch, as I mentioned, loads of different new chip companies that we're super excited to be working with. And how did you get in the front door with those people? These are some of the most exciting chip companies in the world. How did those partnerships come about? So in many cases, I guess it depends slightly on their technical and commercial maturity.

12:44Danyal Akarca:So the way I put this is that you kind of have, I call them the neo chips, right? So we had the old school chips that were pre-AI and now you've kind of got the peri-AI ones that kind of came around the renaissance of AI. And then you've got ones that have been built since. The ones that are more technically mature, they really, really, it's very clear our value proposition to them because ultimately, like we are a mechanism to actually demonstrate even better performance gains on those chips. So I'll give an example is that we're announcing our partnership with Cerebrus. And with Cerebrus, we find a great synergistic relationship because Cerebrus on the market have some of the fastest chips in the world right now.

13:24Danyal Akarca:So if you want to run AI fast, which actually is a massive bottleneck for many, many applications across many different domains. Actually, what we at Colosum do is enable the easy integration of those chips and to be able to use them to make them usable. And not only to make them usable, but to actually make their token economics when integrated into your systems better. So we have really great results that we're announcing to make them cheaper, to utilize these newer chips and actually to benefit from the speed. So for these for these kind of that epoch of amazing chips very well, because we can actually just demonstrate performance gains and so on.

13:58Danyal Akarca:And and with with many of the others, it's very exciting because ultimately we're all going to a shared mission of shaping the kind of compute landscape. Right. This is really a major change in how we think the market should ingest these new chips. And so we like to think ourselves a very friendly partner to many of these players. Yeah, it's really interesting because I guess they're helping introduce you to customers while you are actually helping to build the future market of the world where people use multiple chips, including theirs. Exactly. So it's very clear kind of how we think about this.

14:34Danyal Akarca:We've basically built our AI infrastructure under current assumptions of the current availability of the computer infrastructure. And you can think of effectively the new computing technologies that are entering the market who have some degree of specialization afforded to them, have a bit of a market challenge on their hand, which is kind of determining the various niches that they need to be able to serve for the data centers who want to purchase those chips. And also how they can market themselves in context of like the growing, changing, quick changing demands of AI. And so Closin really helps with that.

15:05Danyal Akarca:You can think of us as effectively a neutral platform that basically helps expose what are the unique benefits of those chips and to be able to help bring them to market. How do you work, if at all, with data center companies? I don't know, is that an education piece? How do you work with them? No, so data centers are very, very aware of this, depending on what kind of scale we're talking about. So many of the hyperscalers are building their own chips. So they're realizing the value of the chips at the moment, which is where at the moment all of the value accrues is within the chips. So all of the hyperscalers are building to some extent.

15:39Danyal Akarca:Google famously with their TPUs and we're ahead of everyone else. But we are also working super closely with partners in AWS, for example, who've built their tritonium chips. And effectively, what we are finding is that they are very, very, very useful at certain things. And so they are very aware that there are kind of new chips to be built. and in many cases they are also purchasing these new chips to try and integrate them within their data centers and so this is a whole new kind of era of heterogeneous compute how is it that you can utilize the different benefits of these computers ultimately to solve these kind of really ever changing ai systems it feels like you're the only player in the market of building the software that enables this why do you think that is well i'd say that there are other people in the space that i can i can outline they uh generally speaking um either focus more on the data centers themselves which is which is great so basically how can you help a new chip enter into the data center or there are people who have for a long time and still continue to do build tools that allow you to port the code from one chip to another um so we kind of call this kind of making approaches that can make your um compilers um hardware agnostic some people have different approaches to this So some people have ways in which they can just port it.

16:59Danyal Akarca:Some people have built new programming languages from scratch to do this. And so all of these companies do amazing work. This is absolutely needed. And this market is going to grow significantly. The approach that we have at Colossum is slightly different. We actually started off not as inherently a heterogeneous computing company. Yashar and I actually dreamed about a future in which the future of AGI or the future of how you build really, truly autonomously intelligent, even perhaps conscious intelligence systems, just like we are as humans. Heterogeneity is one of the core primitives of how you do that.

17:41Danyal Akarca:It's more about how you build intelligent systems that you need this rather than as building the tools to allow porting of chips and so on. We come from a vision of like, how do intelligence, how does intelligence arise? And then you kind of get to this conclusion over many years. And this, I guess, like calls back to your PhD, right? You studied brains. Yeah, exactly. Is that where this idea started? Yeah, totally. So the kind of core thing that Yasha and I were doing for our PhDs was really thinking, what are the principles of efficiency and intelligence that you observe in the brain? What are the principles in artificial intelligence and in computing?

18:23Danyal Akarca:What are shared? What are not shared? What's interesting? What's not interesting? And what we basically came to realize is that many of the lessons that we were learning about why we think the brain is such a fascinating organ, obviously, is its heterogeneity. Heterogeneity is itself something that has evolved in order to be able to run the brain at 20 watts, right? The fact is that we are a multi-scale system, so it's operating at the macro-meso-micro scale, meaning at multiple levels of abstraction. There's many different specializations across those scales. and if you just look at a brain it's not the same everywhere right so so the core insight is is that probably if you want to be efficient and you see this not only in the melian brains this is true across the animal kingdom i see it as actually a cambrian explosion of intelligent systems all with their different niches all have evolved this heterogeneity um and uh you know you take those ideas a few steps forward and you realize probably the way that we evolve the algorithms of intelligence and how we evolve the hardware that enables the intelligence to run should probably be not thought of as two separate things.

19:36Danyal Akarca:You should probably kind of combine them. And so you need to have this kind of co-evolution of the intelligence and the chips together. We just got a lot of deeper philosophical ideas, but it's kind of what led us to think that this is such an important problem to work on. And this trend of the computing industry becoming more heterogeneous is a reflection of that deeper point. Yeah, yeah. And so our thesis is that this is likely going to continue for a while. You know, you've been given$100 million now and that like a lot of capital for a software company. What do you still need to build or prove in the next 6, 12, 18 months?

20:14And how are you going to do that with the capital that you have to further bring this vision to life?

20:18Danyal Akarca:Great. So there's two core things that we're doing right now. One is that we are aggressively scaling our team. um so we um have just uh we are based in london predominantly but we're opening up new jurisdictions um uh and we're ready for that so that's that's that's one kind of core part um we have our hiring uh pages online so please go there if uh you're listening to this uh and um and we are looking to concentrate the world's best talent in close um across the stack we have loads of very very varied roles that we're looking for in the company so that's number one is hiring aggressively the absolute best, most ambitious, high agency people, I want to speak with you.

20:59Danyal Akarca:That's number one. The second part is in securing appropriate compute. So we need to be able to have enough compute, not only to service our customers, which is the main priority, but also to do this so that we can be sustainable over the coming years. And so we effectively see ourselves in a very, very interesting period of rapid change. And so having that capital allows to secure us to ensure that we can sustain ultimately being a key player in this compute industry, as there's a huge amount of capital involved in this industry, obviously, that we can make the moves that we need to, to be able to really capitalize on the fragmentation of the industry.

21:40Danyal Akarca:Ultimately, that there are many, many new chips coming, as I've been outlining, and we need to be able to secure appropriate capacity with the best partners that we work with and so this sets us up for that um interesting and uh and so it's a kind of key key key aspects in terms of what we're now proving over the next uh next next kind of epoch of the business um we are now really scaling up um uh our commercial cadence of the business so um you can think of us in many ways a neolab where we're really pushing the frontier of what's possible on these new chips and so we're posting out some absolutely amazing results um and we're super excited to really be pushing the frontier from a cost speed and capability performance.

22:18Danyal Akarca:And now we're really scaling up, kind of working with the AI consumers. So people who want better AI systems that are tailored to them, be that super, super, super cheap, much cheaper that's available on the market or faster as well. But also the people that want to build the next generation of compute infrastructure. And so that's really what we're laser focused on over the next period of time. And what does that go-to-market look like specifically? I know you've got some partnerships and chip companies. Is it also a bit of going to companies who need more value from AI and saying, hey, you could be using different and varied chips?

22:52Have you thought about that? And we can enable you to do that.

22:55Danyal Akarca:Exactly. So one example is a really great company called Helmgard. So Helmgard is a company that we are providing tailored inference for. And so what this is, is Helmgard is a cybersecurity company. They want the best AI system so that they can service their customers to ensure that they can appropriately assess the risk profile of customers that they work with. And to be able to do that, they ideally want to do it really, really fast. And so we are providing them with optimized systems to be able to do that, but also to be able to provide to them and to other companies cheaper ones, cheaper systems that can do background tasks.

23:33Danyal Akarca:So this is where you can actually ideally not only just have AIs that are really, really fast and can do things that are really rapid because that's needed for something like cybersecurity, but also that you can do background analysis, for example, and you have 24-7 AIs that are always on. And if that's the case, right, if you really want this abundant background intelligence that's going on, you ideally want that to be cheap, right? Because ultimately, the upside of that is like there's infinite amount of AI that you could be running on that, right? And so you want it to be cheap, but also you want it to be really, really highly performant.

24:09Danyal Akarca:And so what we do in that case is that we can optimize those models to be run on different chips in different locations. For example, on the different specialized chips that occupy different clouds. And we can actually find really optimized ways of running that and drive the cost down. And we can do that in a way that I call tailored cheap inference. And this is another way. And so that is kind of, as you said, exactly the chip companies, but actually going to people who really can benefit from this. And this is something that's really kind of very new, very, very new. And over the next kind of coming six months to a year to 10 years is going to be an explosion of new ways to build intelligence that utilizes these different chips.

24:49And price is such a hot topic in the market at the moment, this gap between the amount of money people are spending versus the ROI. I imagine you've got a bit of that tailwind behind you. Is that the biggest problem that people are facing when they're talking to you? Is it about actually want to go faster, or is it really how do we bring the cost down to prove the ROI?