I run Wells Fargo's semiconductor and systems investment banking business, and I'm joined today by Dinakar Munagala, who is the CEO of Blaize. It's great to have you, Dinakar.
Thank you, Preeti.
How are you?
Thanks for having me here. Pleasure being here.
Great. Is his mic working okay? Can you guys hear us?
Mic's good.
Mic's good? Okay, great. Great. Look, I thought we'd start just maybe having you talk about Blaize. What was the strategy and vision when you started the company? Obviously, that's morphed a little bit as this AI wave has played out. Maybe just talk about that and how you're thinking about the go-forward strategy for the business.
Sure. Blaize, prior to starting Blaize, my co-founding team and I, we were building graphics processors at Intel. Chances are if you used a MacBook or a Centrino laptop, those GPUs are something that we worked on and built. At Blaize, what we have is a true platform innovation for edge AI, edge and inference focused on that. The core of the innovation is a novel processor architecture, which is more efficient and lower power and all the specific demands of the edge. We do deliver it in the form of a complete platform, hardware plus software, because we realize that AI is a full-stack problem. A lot of the use cases, especially on the edge, are quite diverse, and they want one thing, really. The edge, they want a return on investment. We focus on delivering practical AI solutions.
Our key markets are smart city, defense, industrial automation, and how we actually create value for the customers is key. That is why we're getting all the business that we are.
Yeah. Let's double-click on that concept of value. When I think about value at the edge, it's performance, it's cost, it's power, and it's also programmability. Maybe just talk about your solution and kind of how you've thought about those parameters in these markets, that smart cities, auto, et cetera.
You're absolutely right. It's usually that, hey, the chips on board a drone and making real-time decisions are in a car or a robotic machine. Real-time is very key. Lower power, often they're battery-operated devices. It can be a camera on a lamppost doing traffic monitoring, et cetera. The third aspect is diverse applications. The edge has everything from a certain class of computer vision problems for industrial automation to education, safety, et cetera, et cetera. Being able to do all of that in a common platform is important. One of the most important things is ease of deployment. Because, of course, in the data center, you have all the experts, and they're all there. Outside in the field, often these are city planners, some kind of a defense analyst, but they're trying to get an AI solution built that they can deploy.
Having a platform, a software layer that can actually make all of these come to life with ease of deployment is key. Of course, the single most important thing is they're not as much in the field of experimentation of AI. They want real deployed solutions. They want to deploy. That's their motivation. AI is more about how it helps their top line or bottom line. This is the single most important growth factor that we're witnessing. Packaging all of this and giving them a platform view is where we focused on.
Got it. When you--I mean, it's obviously gotten to be a fairly crowded field of startups and inference at the edge. How are you positioning Blaize, and how are you guys competing in that market? Who are you competing with, and how are you winning some of these areas?
You're absolutely right. There are about 30, 40 AI chip companies, and almost all of them, they're more dedicated fixed functiony. They focus on, hey, I can do a certain type of YOLO or a resonant very well, and that's their thing. If you look at just AI, even in the last three, four years, the rate of change of AI algorithms is so rapid. Silicon design takes a fixed like three, four years. You can't build something faster. The only way to go about is making sure that your solution is future-proof. One of the foundational principles at Blaize is how do you actually build a future-proof programmable processor that is efficient? If you compare us to the others in the field, and if you say, hey, lesser programmability, more fixed function, and more programmability will be on the right side. GPUs typically.
Being the most programmable.
Most programmable, exactly. Most programmable. GPUs are typically on the same scale of programmability, except that we also are more efficient. We are almost as efficient as fixed function. We give the customer that benefit.
Got it. Okay. That's very helpful. Let's maybe shift gears and talk about the market strategy and your go-to-market. Maybe just walk us through how you guys thought about the biggest pockets of demand from an edge perspective and why you chose to focus on some of the markets that i maybe just start with where you're focusing today and why you're going after those markets specifically.
Clearly for us, the focus is there are about a billion cameras installed out there. These can be in cities, in airports, stadiums. It can be across perimeters of countries. They all want to make them AI smart. They have a choice: rip out all the infrastructure, rebuild, very expensive, or pipe all of this back to a cloud and do it all on GPU, which is also expensive. What they're picking is this whole mode of how do I achieve my business outcome in a hybrid fashion. Blaize, a tiny box potentially at the edge, right behind where cameras are, doing some real-time traffic analytics. Some of it piped back to a server, which is sitting in an on-prem or a cloud with a Blaize chip, again, the common software infrastructure, being able to do all these analytics and so on.
Smart traffic, defense, these are some of the use cases. There's also a hybridization that is happening here. The way the customers look at it is they experiment on GPUs. When they want to deploy, typically their CAPEX budgets, OPEX budgets. How do they actually make a project successful is fitting in these metrics. This is where Blaize plus GPU has a better TCO than GPU alone. This is the big thing that we're finding demand for.
Because of the offload of the.
Because of offload, exactly, into specific areas. It is about how do you actually do this in a seamless fashion. We have also invested in orchestration software. How do you offload the workload easily? How do you coexist with GPUs? We do not look at GPUs as a competitor, but rather we look at it as, hey, Blaize plus GPU is better than GPUs alone. The customer is pretty much taking to that.
Is the traction you're finding mostly with kind of the smart cities, maybe some of the industrial constructs? Where do you see the biggest uptake today?
The biggest uptake is definitely smart infrastructure in that I'm lumping cities, airports, stadiums, et cetera. Defense is another key sector for us. The third one is industrial automation that you mentioned. Also, retail is picking up right now for us. It's kind of in an early POC kind of stage, but that's looking promising as well.
Defense is defense surveillance.
Surveillance, yes.
You're working with the DOD and the like on that?
Exactly. Defense agencies for securing perimeters and so on, those kind of use cases.
Got it. Got it. Very interesting. Let's maybe double-click on one aspect you started your comments on, which is the software stack, the importance of the software stack from a go-to-market perspective. Just maybe talk about what investments you're making there and how are you sort of driving that part of the product stack to help your go-to-market strategy.
So once.
Different folks in the ecosystem have taken different approaches. Some of them have actually spent time building out their stacks, building out their compilers. Others have not or maybe less so. Where are you in that spectrum and kind of how are you thinking about that?
Right. No, no, very good question. In fact, I would say software is equally, if not more, important than the hardware itself. We've taken this approach. The moment you have a programmable processor architecture, the way that you actually expose that programmability out is your software layer. We've consciously invested years in compilers, drivers, fitting into open frameworks like ONNX, coexisting with the likes of CUDA. A lot of thought has gone into software. Now it's actually yielding results for us when it comes to business generation. Software and ecosystem go hand in hand, being able to do this. Within this, the way we're going about is we're creating a platform layer where you can actually easily onboard customers, and customers can deploy with ease. The platform can comprise Blaize software. It can also be third-party software. There's a ISV partner for retail.
There's an ISV partner for facial recognition for something else, et cetera, et cetera. Having multiple.
It's compatible with.
It's compatible, exactly. Making sure that you have complete applications. Because especially these edge customers, they want a solution, and they want to take something and deploy it. They don't want to too much.
Have to tweak with it.
Yeah, tweak, exactly. Focused on solutions and focused on creating market ready, take this and go with it kind of thing. That is where the software layer comes in. There is another key thing. If you look at the whole big picture AI, OpenAI is of the world, et cetera, building these massive models, AI models, there is a trend right now happening that all these models are shrinking. For every model that they create, they create a $600 billion parameter and then $80 billion and $7 billion and so on, $1.5 billion. The trend is that the sweet spot from one of the model companies that we have heard, hey, $12 billion is the trend, $12 or $13, I forget. The point is that the $12-$13 billion parameter model is almost as accurate as the $600 billion parameter model.
Increasingly, the smaller models are being more downloaded on the gates of the world, and they're being deployed more. The reason of bringing up all this is having a software infrastructure that can take care of this and ease of deploying smaller models and being able to run at the edge is key.
Yeah. Yeah. No, that makes sense. When you think about the roadmap, and this is more from a product perspective, where are you looking to take the hardware and software roadmap, thinking kind of maybe your couple of years ahead?
We now have customers in our chosen verticals, and obviously they want, hey, tell us what your roadmap is, and we want these capabilities. It is a very customer-driven roadmap in a sense. Also, to answer your question, we've initially focused on camera, video, image, satellites, those class of workloads. There is a commonality to it; a lot of these are computer vision-based problems. We are now, given the processor is completely flexible, the next generation, we do want to grow the TAM into other use cases. It could be these small language models and so on. The idea is how do you actually serve 70-80% of the world's AI problems in a very cost-efficient, low-power and real-time experience. That is what is driving the roadmap. When I say roadmap, to your point, we do look at it as a comprehensive.
What is the chip roadmap? What is the platform? What are the hardware form factors? What are the use cases? Who are the partners? Everything. All of that is, of course, happening real-time.
There is both going to be a focus on performance and modulating for some of these low-cost use cases. Potentially multiple generations of multiple variations, sorry, product for those use cases.
Yes. Of course, we'll customize at the software level. I mean, the moment we, for example, the current generation, it was one silicon that we built, but we've created many platforms, many use cases, et cetera. Similar strategy, but with, of course, more compute because more compute is being demanded, but also being able to deliver it efficiently and the way I described to all these use cases.
Yeah. Yeah. No, that makes a lot of sense. Maybe we'll switch gears and we'll open it up shortly for questions. Let's maybe switch gears and talk about what is your revenue model approach. Is it product-centric, IP, some combination of both. How does that evolve. How do you think about that.
Definitely product-centric. We sell servers and boxes. That's the hardware layer. Software, we have this unique opportunity given we're a platform that we have built is there's a software recurring revenue model possible for us. The beauty is because of this whole practical AI, that's what we're doing. Customers do come to us with saying, hey, I want to build a certain industrial problem. Can you help me with it? Once we engage with that customer, that gets built into our platform, and we can offer it to other customers. There's a platformization approach that we're taking so that there's repeat recurring revenue, et cetera. Hardware and software, but these are the components.
It is a product first.
It's a product first.
Product first.
Absolutely. Product first and clear business outcome focused. I mean, if the customer makes money, we're good. We get paid. In fact, there have been instances where customers asking us, in fact, of course, we've not yet done it, but customers like it when we go in with, okay, hey, we'll deploy this. If you make money, give us a percentage of your upside. These things are possible. Of course, it's too early for us for those. Those discussions do happen.
Got it. A royalty type of customer?
It's not so much a royalty. It's almost like a rev share. Imagine you've, I'll give you an example. We're not doing this yet, but it could be pretty lucrative, is traffic fines. One model is, of course, you sell the hardware, sell the subscription. Yet another model could be that, okay, hey, if you as a city or a country or a nation are able to do revenue generation from traffic fines and we are able to help you detect that, pay us a cut. They're open to it. That could be actually more lucrative than selling the product.
Right. Right. Makes sense. You have talked about the big pipeline you have. I think it is, what, $175 million or so in the first 18 months. If you break that down, how do you feel about the components of that pipeline? Where is the bulk of that coming from, from a market perspective, but also is most of the uptake here in the U.S.? Is it in emerging markets? How do you kind of split up geographically in terms of that opportunity?
I'll talk about the pipeline and the specific contracts. The pipeline is large. We've announced, I think, a quarter ago about a $700 million-$800 million pipeline besides the contracts we have. This spans U.S., Asia, Middle East, and Europe. Now, the specific contracts that we announced are primarily from Asia and India and Middle East. We have a customer that's a $120 million contract that we announced in Asia for various smart use cases. It can be smart agriculture, smart city, et cetera, mainly camera analytics. The second one is with a data center in India called Yotta. The use cases are around toll management, toll booth. A very interesting use case here is that the number plates in India, as you know, are not standardized. It's not just numericals or alphabets. It can be cursive handwriting.
It can be different languages and so on. It is actually a very complex AI model that you need to use to detect that accurately and Blaize passed the test. That is the second one. We have also been working in the Middle East. We have announced some collaborations with TCC, which is owned by PIF and so on. We are growing in the region. In the U.S., we have announced some, but the pipeline is growing.
What is the from announced to revenue, what does that duration look like for some of these contracts?
It varies. Typically, it's anywhere between six months to a year. Yeah, but it varies.
They're not ripping out infrastructure?
They're not.
They're not. This is just an example.
This is just, yeah. Exactly. That is what they prefer because they already have the cameras and everything. That is the easiest. Of course, there are new smart cities, et cetera, happening as well. Those projects are in discussions. Typically, when there is a new city built out, the scale is good, but the time to revenue is longer as well because the city needs to be built and then the infrastructure.
Yeah. The broader pipeline, which we should talk about a little bit, of $725 million or so that you guys have talked about, is that still a pretty smart city and industrial IoT-centric pipeline? Do you see a broadening set of use cases there or talk about that a little bit?
Definitely, it's still around smart city, defense, drones, those class of use cases and industrial automation and retail. The commonality does come down to camera, video, analytics in this pipeline. Now, we do, of course, get customer demands for additional things like, hey, can you solve a combination of multimodal image and NLP? There are some of these in the pipeline as well. It's a very interesting space that the advent of language models, what they want to do is have language models integrated with computer vision to understand what is the traffic pattern on Highway 101. I'm talking about an Asian country, but equivalent of that. Automatically get a report. There's fog here. There's traffic there. Consider diverting to the city planners. Consider diverting elsewhere. Things like those.
The mix of these two is actually a very powerful and useful tool for city governments that we're seeing.
These are multi-year sort of deployment contracts, especially on the city side?
City side, usually it is a so we work with the system integrator and they, of course, with the city. Yeah, they typically have multi-year contracts.
Multi-year contracts. Okay. Very helpful. Maybe just talking about the cost side of the business. You've talked about paths to margin migration, margin improvement over time. Maybe just spend a little bit of time talking about that. What gets you there? What are the accelerators? What are the risks to getting to that?
Interestingly, the chip that we have right now in the market is on a 14 nanometer node. Despite being on a 14 nanometer node, we're able to demonstrate that when you take this at rack scale, we're about two and a half times better. Pairing Blaize plus GPU is better than GPU alone by two and a half times better, be it CapEx or power efficiency, OpEx, et cetera. Even not going to the latest process generation. The next generation chip that we would build would be on a more leading edge node, naturally, so that it's that much more future proof. Margins, of course, as we will improve as we engage more and more on the software platform layer, that's an additional high margin business for us. Also the next generation silicon products that come up, these are two things.
Coming to risks, of course, we've crossed the architectural risk and adoption risk in the business, but that's been adopted now. Execution risks, as with any other company, do exist. We're, of course, working on it daily to make sure. Supply chain, et cetera, we've figured all of that out through the first generation products.
Sure. Sure. I assume along with all that, the ASPs for your products are going to see some improvement, particularly as you migrate node and performance. Are you going to go 7 nano or what is your?
We're not publicly disclosed, but it'll be on a leading edge. You're right. As we get into leading edge, the margins are much, much higher, naturally. Also, the software, not to neglect there, it's very good software. Especially as we get adopted more and more, there's like the learning that you have on a factory floor can be applied to a city, can be applied to an agriculture problem. We've always had this platform view. That's all coming together now. That has higher margins.
How do you feel about your ability to fund the expansion, the investments over the next some years from a current standpoint?
I mean, fortunately for us, the rationale for us to go public earlier this year was to actually be able to fund these initiatives. That is working out well. We feel we are well capitalized to execute on our current contracts and the next generation chip. Of course, as business grows, if there are more demands, we will look at it. For now, I think we are in execution mode. We feel good. The contracts that we have in hand greatly diminish the risk for revenue for this year and next year. Of course, the pipeline is growing. We continuously keep executing and as a function of success.
It sounds like you have pretty good visibility into the next whatever year, certainly 2026, in terms of the contracts and the pipeline. That gets you to, I guess that gets you to a break-even point, right?
We're moving closer. We've not disclosed the break-even quarter, but yeah.
Yeah. Okay. Great. Maybe we'll pause here and see if any questions before we continue. Sounds like not. Maybe one last thing and then we can wrap it up. You think about kind of milestones of the business in terms of things that you guys are focused on. What does 2026, 2027 look like from that perspective?
Revenue, of course, is a key milestone. And we've given public guidance will be $130 million plus. That's one. The other one is the platform is actually becoming a thing for us. It's real. Customers like it. So marketing this and actually getting tangible customer wins onto the platform. The third one, of course, is more like an internal thing, making progress on the next generation technology. I think these are three key things that we're looking at as. And we feel good about how we've started off as a young company and gotten it to the stage earlier this year as a public company, I mean, and build on top of that quarterly success and cadence, become more predictable, et cetera. That's where the focus is.
Is there sort of a land and expand type of approach to your customers in terms of how you would grow the business?
Yeah, absolutely. In fact, good that you asked. It is actually most of the customers that we're working with, the moment we win their trust on a certain problem, they're beginning to open up saying, "Hey, oh, can you also do this? Can you also do that?" There are actually two things. There is a land and expand within the large customer. Plus there is also land and expand across geographies. If it is something, for example, in the Middle East, something good for a country number one, country number has got to have it. We get involved from there or we go there. The other piece is as we engage, typically the win happens with a combination of us and a system integrator and a software provider. They begin taking us to, "Hey, oh, okay, you're good. It's a well-oiled solution.
Let me get you to my other partner, et cetera. We are witnessing all of those things. That is why we feel good about it.
Okay. That's great. Maybe we'll take one more check on any questions. If not, we can wrap up. Okay. Great. Thanks for joining me, Dinakar and Sprank.
Thank you very much.
It's great to hear about Blaize and the story.
Thank you very much.
Appreciate your coming.