What happens when you use AI to optimize AI and make AI models run fast anywhere?

Tired of choosing between performance and freedom? We sit down with Stefan Crossin, CEO and co‑founder of YASP, to unpack how a hardware‑aware AI compiler can speed up training, simplify deployment, and finally make model portability real. The story starts with a distributed team in Freiburg and Montreal and moves straight into the heart of the problem: most AI groups burn time on infrastructure and juggle separate stacks for training and inference, all while staying tethered to one dominant vendor’s software ecosystem.

Stefan lays out a different path. YASP converts models into a clean intermediate representation, plugs into the tools teams already use, and applies a closed‑loop optimization system that learns the target hardware. Instead of forcing a new language or workflow, a few lines of integration unlock dynamic kernel generation, graph‑level tuning, and one‑click deployment to different chips, clouds, or edge devices. The result is a practical bridge between “write once” ideals and real‑world performance, where being hardware‑aware—not hardware‑bound—delivers speed without lock‑in.

We also dive into the market dynamics behind portability. Incumbents protect moats; challengers need bridges. Cloud providers fear shorter runtimes but win when customers get more value per dollar and per watt. With credible benchmarks showing meaningful gains in training and inference, YASP is courting chip makers, CSPs, and end users through a focused beta, a clear roadmap to launch, and a business model that combines free access with subscription tiers. If you’ve been waiting for proof that AI can be both faster and freer across architectures, this conversation makes the case with clarity and detail.

Enjoy the episode? Follow the show, share it with a colleague, and leave a quick review—what platform or accelerator would you target first with true portability?

Transcript

Origins of YASP and Global Team

SPEAKER_00 0:01

Okay, here we are, HAI partner, and we're with Stefan Crossin from EASP, CEO and co-founder of EASP. Stefan, welcome.

SPEAKER_02 0:11

Yeah, thank you, Pete. Nice to meet you.

SPEAKER_00 0:13

Yeah. And uh so I we had talked about this just before we started recording, but you're currently in Germany, is at Freiburg?

SPEAKER_02 0:21

Yeah, exactly. It's in the south of Germany. I'm I'm privately based in Freiburg, but we have an office in in Munich as well. And also the other part of the company is um based in in Montreal in Canada, where we have also an office as well.

SPEAKER_00 0:37

How did you pick Montreal? Just curious.

SPEAKER_02 0:40

Yeah, it came from from the origins of Yesp. So um I founded YASP um beginning of 2023. It's not a classical um startup, it's more kind of spin-off as well of some activities in the past, where we then directly hired an the core development team. And um the core development team, the most of them um of those members um have been already in Montreal, around Montreal and Canada, and then also one or two of those um members in in Germany, and then we um decided directly to um have that kind of organization in Canada and in Germany. Makes a lot of sense also covering the time zones as well. So minus six, so we almost work uh and can cover 24, huh? So 24. That's right. And so that's that really helps a lot.

SPEAKER_00 1:39

Interesting, yeah. I know there's McGill University there, some really good universities. So going back to kind of the restart, now we're talking about kind of where Yasp came from for folks that don't know. So, what's the origin story behind Yasp? You said you found it in 2023, so it's not too long ago.

SPEAKER_02 1:54

Was it was it a spin-off from some research, or was it just uh yeah, it was it it was it was a it was a spin-off of a of a cooperation of my um of former companies, one of my co-founders company and and um myself. And um I have technical background, I studied electrical engineering, worked as an um embedded engineer, led also a company as well, um, grew that. And um the my my partner or the the other co-founder, he had an um AI startup. Um, then um that AI startup uh got acquired, but in between we were kind of um yeah collaborating already together. Um were discussing um different kinds of projects, and um within the other startup, um there was then a kind of core core team um where we had really the chance then to directly um take that over, hire that team, and continue directly then also with our activities um around our um agentic AI compiler.

SPEAKER_00 2:57

Yeah, so let's talk about that. So compiler companies are always interesting. So uh so actually before we get into the tech a little bit, so you're a uh uh uh you have a seed round, I think I read that. So you have a little bit of funding, right?

Why Build an AI Compiler

Breaking Vendor Lock‑In

SPEAKER_02 3:14

We had, we had. So yeah, um let me let me start at the at the beginning. So why where did we um um started that that activity around the um AI compiler? Yeah, um initially, as we have also um figured a lot of um yeah issues or challenges in the market, so we had dependencies towards a very big um yeah chip vendor um for our AI workload um to train our own models as well, but also um then um to deploy that. Um that was a big hurdle also and very costly as well, quite expensive. And um, what we usually then um have also um seen is um you have this kind of dependencies towards this this big player, the green player here, and um at the same time you have um a team um running the training for the models, but on the other side it's a completely separate team um for for the inference to make that more efficient to run those models either in the cloud or also on the on the edge devices. And um we we took a decision or we we have seen um then also during our development where we said okay, um more than 80% of our time is spent on AI infrastructure. Um we were not really also um being able to to focus on our core development itself, and um, but also from from the market um feedback as well. We have seen this, where we then said, okay, um let's try really to to crack that nut that knot um and and and solve that that problem. Why not using AI to optimize AI? So that means why not building an um compute stack in between the standard training frameworks and the vendor-specific um architectures and compiler stacks. We do not want to compete with the with the big players, and um they have a massive, they have an army of compiler engineers at all, but um, there is a gap in between. How to run the models very efficient, um also custom models, very efficient to the specific targets, to the specific architectures as well, and also get rid of the dependency. So that means being more hardware agnostic on the training side, but also how to get that model seamlessly deployed for the inference. That was the challenge, right?

unknown 5:39

Right.

SPEAKER_02 5:40

Yeah, exactly. And then we started with our with our team working on that. Um had a lot of great ideas. First, we have a really excellent um team. The good thing here is we have a really experienced team. Um, they're not just um graduates from university. So the team at all has more than yeah, or each team member has more than 10 years of experience in the industry. Yeah, so and um then we said, okay, let's let's decide for um focusing really on the infrastructure, creating that unified compute stack to optimize the development of AI in general, yeah, of any kind of AI workload. So that means at the end, any model, any target, any hardware, any compute. Yeah, that's exactly our claim long-term. So meaning um at the end, the customer doesn't have now to change its code base, it's just a few lines of code integration of our compiler. And um, with parameters, you can decide I want to run on vendor A, vendor B, vendor C, and um with a one-click solution, you can then very seamlessly deploy the later on the um the model to the to the target.

Unified Compute Stack Vision

SPEAKER_00 6:57

Do you assume that it's uh so there's um do you assume that on the source end, on the training end, there's sort of a single target? You mentioned the big green partner, like that's typically where a lot of people train their models, but um, and then there's like lots of different endpoints, or are you assuming that either side is variable?

SPEAKER_02 7:16

It's it's it's variable because um I would say the the good thing is that the the the the green company um is um is claiming themselves as a software company, and that's exactly the right way. Yeah, so it's all about the software stack and the software support, and then you can really get the the benefit, the the last juice out of the out of the architecture, out of the chip. Yeah, that's that's why very important. And the others, they started, yeah, we are a chip vendor, we build chips, yeah, we build the hardware, but are lacking on the software support. So some of them go um open source, get community support and and all those to try to to speed up the process, but they're still lacking, and at the end it's a user experience. So the user decides, yeah, the end customer decides um for a solution where the ease of use is given, yeah. And but they can still get more juice out of the target, also for the green customer. And this is what we proved. So a lot of benchmarks are very promising within our um development. And um if everything goes goes the right way, we will launch our product by the end of the year.

SPEAKER_00 8:26

Excellent. And so um is it are you using some sort of intermediate? Uh we should probably get you on for a big long technical talk, and you can actually walk through all the code and stuff. But are you using like Onyx or some sort of intermediate format that you're going from as you're using as a portability framework, or you just sort of like you know, brute forcing it?

Training to Inference, Seamlessly

SPEAKER_02 8:48

No, no, no, no. So um our our idea is not to build up something um proprietary. So our idea is really to have an intermediate representation of the models, convert that um intermediate representation, and then have really the standard interface and the standard models. Our goal is not to um force the customer to learn any new languages or any new toolings, um, they have their existing um development framework for training and inference. And as I said, with just a few lines of code, they can very seamlessly then integrate our compiler, um, do some configurations as optimized for um speed, optimized for latency and throughput. And at the end, they really get impressive um speed up of the training itself of the model, and then it makes automatically ready for a seamless deployment for the for the edge inference or for the inference at all.

SPEAKER_00 9:52

Got it. And uh is there an opportunity for there's so many, as you know, AI acceleration semiconductor companies out there doing interesting things. I mean, all the way out to the neuromorphic edge as well as kind of more the MPU class. Is there some kind of, or are you expecting some kind of SDK or DDK where someone could you know plug their funky chip into your system?

Hardware‑Aware, Not Hardware‑Bound

SPEAKER_02 10:17

Yeah, that's that's a really good question. And that's also one of the reasons why I'm so excited really to join that um foundation here, the edge AIO Foundation, as um really getting the the network and the connection also to those kind of um chip makers, but also maybe some some kind of um champions or rising stars um in the background who have specific um accelerator uh accelerator chips. Yeah, so because our compiler works like this. As I said, we are we are using um some AI as well, yeah, at some part, but um our approach is to have a kind of um closed loop system, we have to um be aware of the of the hardware as it's optimizing itself in a closed loop all the time and generating not only the dynamic the dynamic kernels for the specific target, but also optimizing the entire module, uh the entire um model then yeah for that. So that means if an accelerator um champion, yeah, at the end, an edge accelerator champion comes up and approaches us, it's not a very big deal also to start um supporting those architectures then as well, yeah, because it's a fully automated process at all. Yeah, we we always claim we are hardware agnostic, of course, we have to be about aware about the target, so that means a better wording would be maybe hardware aware, yeah, hardware uh aware um compiler.

SPEAKER_00 11:44

Sure. Yeah, no, makes sense. Well, you know, I mean, uh the as you know, the the the history of computer science and technology is littered with write once run anywhere type of uh efforts. But I think you know it sounds like we're first of all, the need for model portability is huge, as you know, and that's kind of so you're definitely filling the gap there, and then using kind of AI, agentic AI, and this kind of closed loop validation optimization techniques um makes a lot of sense. Uh so so yeah, no, it sounds like it's uh exciting thing. So you said you're gonna be looking to um ship toward the end of this year, which is coming up. And and are you selling into like um is this a direct developer kind of business model, or what's your model here? Is it like a subscription thing into tool chains, or who's your target?

Product Benchmarks and Launch Timing

SPEAKER_02 12:33

Yeah, so we are we're um at the moment we are in a beta phase. Um we're working together with um pilot customers. So um pilot customers can be on the on the chip level, can be on the CSP level, cloud service providers, but also can be then um the end customers as well. And that's really the good thing. Yeah, um when we start working with a chip maker, with the vendor itself, or when we work together with a cloud provider, we can then directly then get um access to the end customers and scale over those. So these are kind of um channels for us as well to get the access. And um at the end, of course, we we plan also um partly um we plan also to to go open source as well, some points. So there are a lot of discussions as well, but um at the end it will end up in a subscription um based model as well for the customers where we'll have uh free package as well, and then specific um um packages and for for the individual customers as well.

SPEAKER_00 13:39

Got it, got it. Yeah, that makes sense, and um, so is this your first startup?

SPEAKER_02 13:45

It's my first startup, actually, yes, yes. So um worked um over 15 years now, so more than 20 years now in the industry. Um and during those times um I worked in our um family-owned business, um, grew up there, um, worked as an um embedded engineer, then um started leading the company, and then finally really said, okay, that's that's a really great opportunity. We have a great solution for a really huge problem at the moment, and um the right timing. I want to go that way. I want to go that journey. Um yeah, so took the decision to do that.

SPEAKER_00 14:29

Awesome. Yeah, no, it's exciting. I mean, um, you know, I mean the NGIA foundation itself is a small business as well, so that's uh, you know, we have to we have to run it as a startup as well. Uh, and there's lots of different decisions and uh things you can bet on and experiments you can run, but you have to run the business. I was gonna ask you, so uh maybe you can give us a little background on the word YASP. Yet another what? Or is that is it is a reference to that acronym?

Standards, IR, and Ease of Use

SPEAKER_02 14:58

No, no, no. Uh yeah YASP is just a kind of it's it's just a random name, honestly. Uh oh. I I feel it's it's it's positive. So before we had another name, um in the in when we have been in stealth mode, um, then we we closed the seed round um beginning of the year in in April, April, May, May, um this year, successfully. And then we have realized, oh my god, okay, the previous name was already trademarked um in euro. Um, but it was not that that bad because we have been in stealth mode. The name itself was not that famous or well known in the market. And we said, okay, now let's take the chance and find a new name. And um, yes, sounds positive. It's a yes with a P. So um internally within the team, it's always uh it's not a no, it's always a yesp. So yes, yes, exactly.

SPEAKER_00 15:56

The uh so I used to be a BIOS engineer, and we used to have a tool called YACE that was stood for yet another CMOS editor. Uh and I've also you know I also work with uh on a project called Acri, which is another Kubernetes resource interface, AKRI.

SPEAKER_01 16:12

So that's why I was wondering if YASP is like yet another software something, but it's just maybe yet another software provider, but internally within the team, we're already joking about that. So um there are a lot of ideas.

unknown 16:24

Okay, good.

SPEAKER_00 16:25

Well, no, that's a good positive name, and you know, and it is pretty unique. I mean, when you go and Google it and stuff, there's not a lot of YASP stuff out there, so I think you've got a good edge on it. I was gonna ask you, um you know, coming into the foundation, we have a number of startups, uh pre-seed and kind of series A startups. What what have you seen, especially in the super noisy AI environment? Like what are you what are your concerns about kind of cutting through the noise and getting the attention that you need? Like what how are you addressing that?

Supporting New Accelerators

SPEAKER_02 16:59

So I I I'm always uh um a fan of um having really something in in hand to prove um that means um we can we can really convince with our product, with our numbers. So as soon as we have really impressive numbers for um the um making the training better, you know, speeding up the training times, but also um improving the the inference. And if we say and talk about X um X um one point, I don't know, like 1.1, um this is not impressive. But we have already now benchmarks um where we see um really high numbers um already um for inference but also for for training. And as soon as we have those numbers and we can really do the communication, I am always convinced saying, okay, the product I cannot say the product sells itself, but we make enough, we make enough noise. Yeah, we yeah, so technically we always have to convince technically uh with the product itself, and then you can start really getting visibility, and then um yeah, you you you get also accepted in the market, right?

SPEAKER_00 18:12

Right. Well, yeah, no, it makes sense. I mean, yeah, at the end of the day, if the product's good, then uh that's the core, that's the core thing. And like you said, with your stuff, it's kind of the proofs and the pudding in terms of the numbers.

SPEAKER_02 18:24

So but maybe, maybe, maybe just adding to that, um, of course, it's important to to build up a network. That's that's also quite very important. And um, as we see now in the the foundation here, the HAI Foundation is really growing. Uh what I see as well, also when I talk to um the um the colleagues from the HAI Foundation, really great guys, very supportive as well. And um the um got also a lot of introductions already, yeah. Oh good events, so makes a lot of sense.

SPEAKER_00 18:55

Makes a lot of sense also. I appreciate that. Yeah, there is a stack, I call it layers and players. So there's there's layers and then there's players, and then it has to sort of all work together. Actually, before we started recording this, I noticed in the news today that Qualcomm acquired Arduino. So speaking of uh partners making friends, uh that seems to be NGF Foundation seems to be the place to make friends. I think that's our new slogan. Um because uh it just seems to be a lot of that stuff going on, but it's a dynamic space. Fantastic. So you're so you're in Germany, you're in Montreal. Are you doing any business in Asia right now, or you're you're kind of focused on European and uh not not in Asia, but also in in the US market as well.

SPEAKER_02 19:40

So we have a lot of um yeah, um POCs, projects, pilots going on.

SPEAKER_00 19:48

No Bay Area folks yet?

SPEAKER_02 19:50

You haven't uh cracked the not not yet, but we are talking to a lot there as well, and um, so that means our core markets are um of course Euro, Canada, and then also US. So we will open very soon also an entity, and um we get visibility also in the US very soon. So yeah, it's not growing now.

Go‑To‑Market and Beta Pilots

SPEAKER_00 20:09

Yeah, well it's yeah, one step at a time, one success at a time, right? It's interesting when you talk about model portability, so you know, and silicon partners, it's an interesting dynamic. You know, there's I I believe in the tech world everyone wants to commoditize everybody else. That's kind of the dynamic. So, you know, and to a certain extent, silicon partners don't want model portability because they don't want people moving from their chip to someone else. That being said, if you're in second place, you do want portability so that you can get uh the models from someone else's platform onto yours. So it's an interesting dynamic. I'm sure you're seeing that when you're sort of working with partners, like which partners are really motivated for portability versus others that are sort of grudgingly um, you know, engaging in in and portable efforts. But um yeah.

SPEAKER_02 21:00

I think I think you you always have to be open, uh you always have to be open also for the others. It doesn't make sense to close the doors um towards others. Um also the the the word self-competitors is um also not really sure if that is always um the the right wording, also for the big players as well. Yeah, there can be maybe uh a new startup, um building a new um AI accelerator um chip for specific applications. Of course, you can call that as a as a green big company, it's a competitor, but you could also maybe somehow benefit out of that, yeah. So um I'm always more open for um collaborations as well. Um the market is big enough um at all, um, like this. And if you really say, hey, um I don't want to have this portability of of modules because this customer has just to be um um has just to um be focused on on my side. Yeah, but what are you doing? Uh you're um forcing the customer creating code specifically for only for that for their architecture, they will not be happy in the future, right? They want to have the the independency um of uh want to be independent of um taking a decision at the end for going for other defenders. The same on the cloud side, yeah. The cloud, same on the cloud side, all the CSPs. It doesn't make sense to say, yeah, that's that's micro customer as well, yeah, because we had also discussions about okay, we are speeding up the training times. Okay, directly we can say okay, the cloud provider can say no, um yes um reduces the the revenue on our side. No, that bring added value to the customer to get more benefit to bring more service to the customer.

SPEAKER_00 22:44

Exactly.

unknown 22:45

Exactly.

SPEAKER_00 22:45

Yeah, no, that's good. Awesome. Well, it sounds like you're on a good journey. Um, and like I said, we should schedule a deeper tech talk so we can kind of walk through the tools and the code and uh the results and all that stuff. But yeah, it's kind of uh it sounds like you're in a spot to solve a really serious problem that's gonna open up a lot of opportunities. So appreciate you being part of the foundation and uh and making friends and um look forward to meeting you in person at some point in the future.

SPEAKER_02 23:12

Yeah, of course. Thanks for having us. Thanks for having us and maybe meeting us um in the uh in the Taipei event.

SPEAKER_00 23:20

Taipei, yes.

SPEAKER_02 23:22

Okay, perfect. Then let's meet there.

SPEAKER_00 23:24

Okay, all right, take care, Stefan. Thanks.

SPEAKER_02 23:27

You too. Thanks, Pete. Bye better.